# The Information Content of Realized Losses

The Information Content of Realized Losses Abstract Examining the trades of company insiders, I find that a sale of stock at a loss is a much more negative signal about future returns than is a sale of stock at a gain. I consider a range of explanations for my results and find that the evidence is most consistent with the idea that investors derive direct disutility from selling a stock at a loss. Since selling a stock at a loss is painful, an investor who sells at a loss must have particularly negative information. This result offers a novel measurement of the strength of the disposition effect. Received December 7, 2015; editorial decision December 18, 2017 by Editor Robin Greenwood. Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online. The disposition effect is the robust empirical fact that many groups of investors—even sophisticated ones—have a greater propensity to sell assets at a gain, relative to the reference price, rather than at a loss.1 One way of summarizing the disposition effect is to say that investors are averse to realizing losses. This aversion has several possible sources, both economic and psychological. For example, it may result from rebalancing motives: a desire to rebalance their positions means that investors will have a greater propensity to sell assets at a gain rather than at a loss. Alternatively, it may be due to “realization utility,” in other words, to a direct preference for realizing gains and for not realizing losses. In this paper, I investigate how the disposition effect interacts with informative signals. If an informed investor does not like realizing losses, then observing her realize a loss suggests that she must have especially negative information about the future returns of the stock. Indeed, the information must be so negative that it leads her to override her inclinations to not realize losses. As such, I predict that a realized loss by an insider—an informed one—is a more negative signal for future returns than is a realized gain by an insider.2 To test this prediction, I must define the terms “realized gain” and “realized loss.” Presumably, an investor views a sale of stock as a realized gain if the price when sold exceeds some reference price she had in mind. In my analysis, I consider three reference prices. I focus on the price at which the stock was purchased, based on FIFO, or first in, first out, accounting. I also consider a reference price based on a time-series average of recent month-end prices and a reference price based on the most recent purchase price. With this definition of realized gain/loss in hand, I test my prediction using data from the U.S. Securities and Exchange Commission (SEC) on the trades of company insiders. I find that when an insider sells company stock at a loss relative to the FIFO reference price in the lagged month, the subsequent 6-month return for the company is 188 basis points (bps) lower than all other firm-months. On the other hand, when an insider sells shares at a gain relative to the FIFO reference price in the lagged month, the subsequent 6-month return is 9 bps higher than all other firm months. This suggests that insiders require a strong negative signal to realize a loss. The predictive power of realized losses allows for the construction of portfolio strategies that, at least before transaction costs, generate excess returns. An equal-weighted portfolio that buys stocks that have been recently sold at a gain by insiders and sells stocks that have been recently sold at a loss by insiders (using the FIFO reference price as the reference price) earns a four-factor alpha of 67 bps per month. In short, stocks perform more poorly after a sale by an insider at a loss than after a sale by an insider at a gain. Having confirmed my basic prediction, I then try to understand, at a deeper level, what is driving the result. My maintained hypothesis is the one described earlier: given the aversion to realizing losses that is inherent in the disposition effect, a sale of stock by an insider at a loss should be particularly informative. The investor must have a particularly negative signal about future returns if she is willing to sell despite a typical aversion to doing so. I first offer evidence against some alternative explanations for my result. For example, it is plausible that insiders who have a stronger propensity to sell at a loss execute more informative trades. Then, we would observe more negative return predictability from realized losses than from realized gains, but this would not be driven by an individual aversion to realizing losses. Rather, it is driven by the fact that informed traders tend to realize more losses. By comparing realized losses and realized gains within insider, I show that this is an unlikely explanation if the insider’s propensity to sell at a loss is static. It is, however, possible that this aversion is not static. For example, an aversion to realizing losses could change with the insider’s position in the firm. Furthermore, access to information likely changes as an insider’s position in the firm changes. Then we might observe more negative return predictability from realized losses than from realized gains, but this would not be driven by an individual aversion to realizing losses. By comparing the propensity to realize gains with the propensity to realize losses for each insider position type and comparing the return predictability of realized gains with the return predictability of realized losses for each insider position type, I show that this is an unlikely explanation of my result. The evidence suggests that my result is driven by an aversion to realizing losses. I examine this suggestion more closely: what exactly is driving insiders’ aversion to realizing losses – an aversion that makes their sales at a loss particularly informative? The evidence I offer is consistent with an explanation rooted in realization utility. Under this view, investors derive utility directly from realizing gains and losses: they experience a positive burst of utility when they realize a gain, a burst whose size depends on the size of the gain realized, and a negative burst of utility when they realize a loss.3 A sale of stock at a loss is then particularly informative about future returns: the investor must have negative information that overrides the pain felt when selling at a loss. I am able to go one step further by using my data on company insiders to comment on the source of realization utility. One view, the “heuristic” view, is that insiders feel pain when they sell an asset at a loss because they have in mind a rule of thumb, namely that selling assets at a loss is a bad idea. Another view, the “cognitive dissonance” view, is that investors feel pain when they sell an asset at a loss because doing so forces them to admit that their earlier purchase decision was a mistake. The company insider data allows me to examine these alternative hypotheses. When insiders acquire a position in a stock, they can do so in one of two ways: they may actively purchase shares or they may be endowed with shares. Under the dissonance view, an investor will not find it painful to sell shares she has been endowed with: since there was no active purchase decision, she does not have to blame herself for a trade gone bad. As a result, a sale of endowed shares at a loss should be no more informative about future returns than a sale of endowed shares at a gain. Under the heuristic view, however, a sale of endowed shares at a loss will be more informative than a sale of endowed shares at a gain: the investor finds it painful to sell even endowed shares at a loss. I test this prediction in the data and find evidence in favor of the dissonance view. My basic finding – that a sale of company stock by an insider at a loss has more negative predictive power for future returns than a sale at a gain – is most consistent with realization utility, in other words, with the view that people experience pain when they close out a position at a loss. This, in turn, means that when they do close out a position at a loss, they must have particularly negative information, information which manifests itself in low stock returns over the next few months. 1. Related Literature A large empirical literature investigates whether the trading activity of corporate insiders predicts returns in the cross-section (e.g., Lorie and Niederhoffer 1968; Rozeff and Zaman 1988; Lin and Howe 1990; Bettis et al. 1997). Seyhun (1998) reviews the evidence and concludes that insider trades contain predictive power for future returns. He notes that the information content of insider trades is higher when insiders purchase. Similarly, Lakonishok and Lee (2001) argue that the informativeness of insider trades comes from purchases, not sales. Finally, Jeng et al. (2003) highlight that whereas insider purchases earn abnormal returns, insider sales do not. In general, the literature attributes the weak predictive power of sales to the fact that sales are often driven by liquidity and diversification concerns, not private information. There are some papers that relate insider sales to private information. Cheng et al. (2007) show that sales filed under Form 5, which allow for delayed disclosure, predict negative returns. Marin and Olivier (2008) show that insider sales peak months before a large drop in the stock price. Relatedly, Gao et al. (2016) find evidence that insider silence predicts negative returns. In contrast, Huddart and Lang (2003) document that returns are higher when option exercises are low. Ali and Hirshleifer (2017) show that the sales of insiders, who make profitable trades prior to quarterly earnings announcements, have information content. Jagolinzer (2009) looks at the trades of insiders who pre-specify sales using 10b5-1 plans and finds that sale plans are terminated before good performance. Cohen et al. (2012) develop a filter based on trading patterns that decodes whether a sale is likely to be informative (opportunistic) or not (routine). The authors show that opportunistic sales have predictive ability for future returns. I extend our understanding of the information signals that insider trades reveal. The conditioning variable I suggest is motivated by the large literature on the aversion to realizing losses. Numerous researchers have documented this investor tendency. For example, Odean (1998) documents an aversion to realizing losses in a data set of 10,000 trading accounts from a large discount brokerage. Genesove and Mayer (2001) find an aversion to realizing losses in the downtown Boston housing market. Grinblatt and Keloharju (2001) obtain data on the trading of people and institutions in the Finnish stock market and find that investors are reluctant to sell at a loss. Frazzini (2006) documents this behavior in mutual fund managers. Hartzmark and Solomon (2012) look at a set of NFL betting contracts at Tradesports.com and uncover evidence consistent with the disposition effect. Heimer (2016) shows that social interaction can exacerbate the disposition effect. Kallunki et al. (2009) document the disposition effect in the trades of Swedish insiders. Heath et al. (1999) offer evidence that corporate insiders consider reference price heuristics when timing their option exercises. In particular, since option exercises do not have an associated purchase price, the authors show that corporate insiders consider dynamic reference prices. Specifically, the authors show that employee exercise activity approximately doubles when the stock price exceeds the maximum price attained during the previous year. My research offers a new way of measuring the disposition effect. Namely, by looking at returns following a realized loss, I can proxy for how strong of a negative signal insiders require in order to sell at a loss. In this paper, I argue that, as a consequence of this aversion to realizing losses, insider sales at a loss will have more negative predictive power for returns than sales at a gain. After confirming this prediction, I examine the data to understand the source of the aversion to realizing losses. There are a number of theoretical explanations for this aversion. Odean (1998) notes that rebalancing is a possible explanation, but provides evidence against it. Another possible explanation is a belief in mean reversion. However, Weber and Camerer (1998) and Hartzmark and Solomon (2012) provide evidence against this view. Barberis and Xiong (2012) show that realization utility—the notion that investors directly derive utility from realizing gains and losses—and time discounting can explain the disposition effect and a number of other puzzling facts. Finally, using brokerage and experimental data, Chang et al. (2016) provide evidence for the view that investors are averse to realizing losses because it is painful to admit that a prior purchase decision was a mistake. I find evidence consistent with the theory that it is painful to admit that a prior purchase decision was a mistake. 2. Data and Reference Price Construction 2.1 Data set construction To conduct the analysis, I collect data from several sources. I obtain return information, monthly closing prices, split-adjustments, and daily low/high/closing prices from CRSP. Other firm-level information comes from the CRSP/Compustat Merged database. I consider ordinary common shares listed on the AMEX, NASDAQ, or NYSE. I do not consider REITs, closed-end funds, ETFs, or Americus Trust Components. Following Shumway (1997), I replace missing delisting returns with a return equal to $$-$$0.3 for performance-related delistings. I construct a book-to-market control equal to the log value of common equity divided by market capitalization, where market capitalization is equal to the quarterly closing price times the number of common shares outstanding. I only consider observations that have a pre-log book-to-market ratio greater than 0 and less than or equal to 100. Size is calculated as the log value of market capitalization. I construct a control for momentum by calculating the previous year’s return, excluding the most recent month (i.e., I use the return from $$(t-12)$$ to $$(t-1)$$). To construct a control for the capital gains overhang, I need a weekly return series. I use CRSP’s daily stock return data file to construct a weekly return series. To mitigate any microstructure effects, I exclude observations where the month-end share price is below one dollar from my return-predictability analysis. I download factor data from Kenneth French’s website. I obtain insider data from Thomson-Reuters. The SEC requires corporate insiders, or “a company’s officers and directors, and any beneficial owners of more than 10% of a class of the company’s equity securities $${\ldots}$$”, to file their trades. I exclude observations with a cleanse indicator equal to A or S, as these indicate a failed cleansing attempt. I consider insider trades with a transaction code equal to P, S, or A from 1986 to 2015. That is, I consider open market purchases, open market sales, and grants or award transactions (hereafter, endowed shares), respectively.4 From here on, when I refer to a transaction by an insider, I am referring to one of these three types of transactions. I aggregate trades by personid, firm, transaction date, and transaction code. I use a share-weighted average, split-adjusted transaction price to compute the daily transaction price. If the transaction/endowed price is unreported, lower than the daily low price, or higher than the daily high price, I use the split-adjusted closing price for the corresponding day. I present summary statistics for the Number of trades in Table 2. As it may be difficult for “active” traders to keep track of their reference price, I drop “active” insiders from the sample. Specifically, I drop any insider that has more than 10 sale days, more than 10 buy days, or more than 10 acquisition days. I present similar results from the full sample in the Online Appendix. Table 2 Number of transactions by insider Number of acquisition days Number of purchase days Number of sale days 5th percentile 1 1 1 10th percentile 1 1 1 25th percentile 1 1 1 50th percentile 2 2 2 75th percentile 5 4 6 90th percentile 10 8 13 95th percentile 15 15 22 99th percentile 41 45 58 Max 430 1,057 1,551 Number of acquisition days Number of purchase days Number of sale days 5th percentile 1 1 1 10th percentile 1 1 1 25th percentile 1 1 1 50th percentile 2 2 2 75th percentile 5 4 6 90th percentile 10 8 13 95th percentile 15 15 22 99th percentile 41 45 58 Max 430 1,057 1,551 This table looks at the number of trading days for each insider and firm combination. I present the percentiles for number of acquisitions, purchases, and sale days by insider conditional on having at least one acquisition, purchase, or sale, respectively. I drop insider-firm combinations from the sample that have more than 10 purchases, 10 sales, or 10 acquisitions upon entering the month. Table 2 Number of transactions by insider Number of acquisition days Number of purchase days Number of sale days 5th percentile 1 1 1 10th percentile 1 1 1 25th percentile 1 1 1 50th percentile 2 2 2 75th percentile 5 4 6 90th percentile 10 8 13 95th percentile 15 15 22 99th percentile 41 45 58 Max 430 1,057 1,551 Number of acquisition days Number of purchase days Number of sale days 5th percentile 1 1 1 10th percentile 1 1 1 25th percentile 1 1 1 50th percentile 2 2 2 75th percentile 5 4 6 90th percentile 10 8 13 95th percentile 15 15 22 99th percentile 41 45 58 Max 430 1,057 1,551 This table looks at the number of trading days for each insider and firm combination. I present the percentiles for number of acquisitions, purchases, and sale days by insider conditional on having at least one acquisition, purchase, or sale, respectively. I drop insider-firm combinations from the sample that have more than 10 purchases, 10 sales, or 10 acquisitions upon entering the month. I construct earnings controls using financial data from Compustat Fundamentals and IBES. Considering Compustat Fundamentals quarterly data, I first drop duplicates that are a byproduct of fiscal year-end changes and then I drop duplicates that are a byproduct of multiple issues by the firm. The data from IBES is constructed from IBES actuals and IBES’ summary of analyst forecasts. I collect the unadjusted median estimate, the unadjusted actual earnings, the standard deviation of earnings estimates and the number of earnings estimates. I drop observations where the stated quarter end is one day before, on, or after the earnings announcement. I remove duplicates by retaining the most recent median forecast. I consider firms that use U.S. dollars as the currency. To merge Compustat Fundamentals data and IBES data I use the iclink linking table. I drop duplicate observations after merging with the iclink table by only keeping observations with the highest matching scores. Following DellaVigna and Pollet (2009), I define the earnings announcement date as the earlier of the Compustat and IBES dates. I calculate the earnings surprise as equal to the actual estimate minus the median estimate times 100. I winsorize the earnings surprise, the lagged earnings surprise, the standard deviation of earnings estimates, and the number of earnings estimates at the 1% level. 2.2 Reference prices My prediction is that a sale of stock by an informed investor that represents a realized loss is a more negative signal for the stock’s future return than a sale of stock by an informed investor that represents a realized gain. To test this hypothesis, I need to define “realized gain” and “realized loss.” Presumably, an insider thinks of a sale as a realized gain if the sale price is higher than some reference price she has in mind, and as a realized loss if the price is below the reference price. Therefore, to define “realized gain” and “realized loss,” I need to specify what this reference price is. In this paper, I consider three plausible reference prices. I list them below, along with details of their construction, and then discuss them. Recall that an insider can acquire a position in her firm’s stock in one of two ways: she can actively purchase shares, or she can be endowed with them. When I use the term “purchase price,” I am referring to the weighted-average split-adjusted purchase price from the day of purchase; when I use the term “endowed price,” I am referring to the weighted-average split-adjusted price recorded on Form 4 for grants or awards on the day of endowment; and, finally, when I use the term “acquisition price,” I mean either a purchase price or an endowed price. The reference prices I use are described here. FIFO reference price. I focus on this reference price throughout the paper. I determine this reference price by calculating a weighted average of past acquisition prices. Each acquisition price is weighted by the number of shares acquired that day (from the associated method) that have not been sold. I determine which shares have been sold based on FIFO accounting. That is, I will consider the oldest possible shares acquired to be the ones sold in a given period. Most recent purchase price. Moving-average price. I take the average of the previous six split-adjusted month-end prices. I focus on the FIFO reference price because it is likely the reference price used for inventory considerations.5 This reference price depends on previous purchase and previous endowed prices. I also consider a reference price based on the most recent purchase price. This reference price allows for a more reasonable discussion when I consider the short swing rule. Additionally, the most recent purchase price is plausibly the most salient purchase price due to recency effects. Several papers have suggested that purchase prices are natural reference prices. For example, Shefrin and Statman (1985) suggest that, when an investor buys a stock, a mental account is opened, one that closes only when she eventually sells the stock, at which point she evaluates the transaction by comparing the sale price to the purchase price. Similarly, Barberis and Xiong (2012) suggest that people think of their investing history as a series of investing episodes, each characterized by the name of the asset, the purchase price, and the sale price (“I bought IBM for 80 dollars and sold it for 120 dollars”). The last reference price I consider is different: it is a time-series average of recent monthly stock prices. I use this reference price for a number of reasons. First, and most importantly, it strikes me as a plausible proxy for when an insider might consider herself to be trading at a gain or loss. Suppose that an insider bought stock of Company X 10 years ago and enjoyed a meteoric rise in Company X’s stock price over the first 9.5 years. However, over the past 6 months, Company X’s stock is down 50%. I suspect that this insider might consider the stock to be trading at a loss. Specifically, if there are recent higher prices, the individual may regret not selling at the recent higher prices, and may consider herself to be trading at a loss relative to those prices. As attentive people, insiders may have more than one reference price; “I’m doing worse than I was doing last month, but I’m doing better than I was two months ago.” The moving average reference price tries to average a few of these recent reference prices. Of course, this reference price is most relevant for those who have held the stock for a long time. In Table 3, I show that the average holding periods of insiders in my sample are quite long. Specifically, the average time from the most recent purchase to a sale is 1,165 days and the median period is 782 days. The average time from the most recent acquisition to sale is 592 days and the median period is 294 days. Second, the moving average reference price does not require a history of purchases or acquisitions – as such, it is easy to consider people who have acquired shares through derivatives/who have acquired shares in an unclear way. Relatedly, it does not interfere with Cohen et al.’s (2012) selection criteria for opportunistic/routine insiders, which I will make use of later in the paper. Finally, the moving average reference price is a firm-wide reference price. This allows me to easily proxy for the “ability” to sell at a loss on a firm wide basis. Table 3 Average holding periods Average holding period Median holding period From most recent purchase to sale 1,165 days 782 days From most recent acquisition to sale 592 days 294 days Average holding period Median holding period From most recent purchase to sale 1,165 days 782 days From most recent acquisition to sale 592 days 294 days This table presents insiders’ average and median holding periods in my sample. The most recent acquisition refers to either the most recent purchase or the most recent grant or award of stock. Table 3 Average holding periods Average holding period Median holding period From most recent purchase to sale 1,165 days 782 days From most recent acquisition to sale 592 days 294 days Average holding period Median holding period From most recent purchase to sale 1,165 days 782 days From most recent acquisition to sale 592 days 294 days This table presents insiders’ average and median holding periods in my sample. The most recent acquisition refers to either the most recent purchase or the most recent grant or award of stock. 3. Company Insiders In this section, I examine the trades of company insiders. I first test my prediction that a sale at a loss predicts more negative returns than a sale at a gain. Having confirmed my prediction, I examine alternatives to my hypothesis that my result is driven by an aversion to realizing losses. After finding evidence that the result is driven by an aversion to realizing losses, I examine potential drivers of this aversion. 3.1 Disposition effect Central to my prediction is the premise that insiders are averse to selling at a loss. If this is the case, the proportion of losses realized by insiders should be less than the proportion of gains realized by insiders. I consider a panel of insider-firm combinations. I include monthly observations for each insider from the month of her first transaction at the firm to the last month she appears in my data set. I consider it to be possible for an insider to sell at a loss if the previous month-end price is below the insider’s reference price. I calculate the proportion of losses realized, or PLR, by dividing the number of months when an insider sold at a loss by the number of months when the insider could have sold at a loss. In Table 1, I present the PLR and the PGR, or the proportion of gains realized, for the three reference prices that I focus on throughout my analysis - the reference price using FIFO accounting, the most recent purchase price, and the previous 6-month moving-average price. I find significant evidence of the disposition effect. For example, I find that, relative to the FIFO reference price, the PLR is 1.35% and the PGR is 2.28%. The difference is statistically significant at the 1% level. This result adds to a recent literature which shows that even “sophisticated” investors exhibit a disposition effect. Frazzini (2006) documents this behavior in mutual fund managers. Massa and Von Beschwitz (2015) show that short sellers exhibit the disposition effect. Kallunki et al. (2009) document the disposition effect in the trades of Swedish insiders. Table 1 Disposition effect Gain Loss Gain Loss No sale 4,494,639 3,642,119 No sale 4,442,653 2,065,232 Sale 210,164 101,302 Sale 103,529 38,539 PGR=0.0447 PLR=0.0271 PGR=0.0228 PLR=0.0135 Gain Loss Gain Loss No sale 4,494,639 3,642,119 No sale 4,442,653 2,065,232 Sale 210,164 101,302 Sale 103,529 38,539 PGR=0.0447 PLR=0.0271 PGR=0.0228 PLR=0.0135 Gain Loss No sale 2,909,502 1,577,226 Sale 68,671 25,049 PGR=0.0231 PLR=0.0156 Gain Loss No sale 2,909,502 1,577,226 Sale 68,671 25,049 PGR=0.0231 PLR=0.0156 The first row of the left table shows the disposition effect for the moving average. The first row of the right table shows the disposition effect for the FIFO reference price, and the second row of the table shows the disposition effect for the most recent purchase price. PGR, the proportion of gains realized; PLR, the proportion of losses realized. Table 1 Disposition effect Gain Loss Gain Loss No sale 4,494,639 3,642,119 No sale 4,442,653 2,065,232 Sale 210,164 101,302 Sale 103,529 38,539 PGR=0.0447 PLR=0.0271 PGR=0.0228 PLR=0.0135 Gain Loss Gain Loss No sale 4,494,639 3,642,119 No sale 4,442,653 2,065,232 Sale 210,164 101,302 Sale 103,529 38,539 PGR=0.0447 PLR=0.0271 PGR=0.0228 PLR=0.0135 Gain Loss No sale 2,909,502 1,577,226 Sale 68,671 25,049 PGR=0.0231 PLR=0.0156 Gain Loss No sale 2,909,502 1,577,226 Sale 68,671 25,049 PGR=0.0231 PLR=0.0156 The first row of the left table shows the disposition effect for the moving average. The first row of the right table shows the disposition effect for the FIFO reference price, and the second row of the table shows the disposition effect for the most recent purchase price. PGR, the proportion of gains realized; PLR, the proportion of losses realized. 3.2 Firm-level regressions The central prediction of this paper is that insiders’ realized losses have more information content than their realized gains. To test this, I examine the predictive power of realized losses relative to realized gains for the future one month return. I run pooled ordinary least squares (OLS) regressions with month fixed effects. Specifically, I estimate the following equation for firm $$i$$ in month $$t$$: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1}&= \beta_{0}+\beta_{1}\:\textit{Short-term}\ \textit{reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}\\ &\quad +\beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}\\ & \quad +\beta_{6}\:\textit{Sale}\ \textit{below}\ \textit{reference}\ \textit{price}_{i,t-1} \\ & \quad+\beta_{7}\:\textit{Sale}\ \textit{above}\ \textit{reference}\ \textit{price}_{i,t-1}\\ & \quad+\beta_{8}\:\textit{Capital}\ \textit{gains}\ \textit{overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t}.\label{eq:1} \end{split} \end{align} (1) I estimate the equation using a sample of all firm-month observations since 1986. I include all firm-months as Ali and Hirshleifer (2017) argue that limiting the sample to firms with insider trades may bias the results in favor of finding return predictability in sales. The dependent variable is the return on stock $$i$$ from the end of month $$t$$ to the end of month $$t+1$$. The most important independent variables are Sale below reference price – a dummy that takes a value of one if an insider in firm $$i$$ sold stock at a loss relative to the reference price in month $$t-1$$ – and the analogous Sale above reference price. I lag the variables of interest as I do not want to pick up a contemporaneous correlation between firm returns and insiders’ trade signals. My prediction is that $$\beta_{6}$$ is less than $$\beta_{7}$$. In other words, I predict that a sale at a loss has more negative predictive power for subsequent returns than a sale at a gain. Before turning to the results, I first explain the other independent variables in the regression. I control for the previous month return, the previous year return, excluding the most recent month, the book-to-market ratio, and the size of the firm to ensure that my results are not driven by the short-term reversal phenomenon, the momentum anomaly, the value anomaly, or the size anomaly, respectively. I also control for the capital gains overhang (CGO), which, for a given firm, is a measure of the average capital gain embedded in investors’ holdings of the firm’s stock. Grinblatt and Han (2005) and Frazzini (2006) argue that, as a consequence of the disposition effect, CGO will predict subsequent stock returns; they confirm this prediction in the data. Since CGO is likely correlated with my Sale below reference price and Sale above reference price variables, it is important to control for it in my tests. Following Grinblatt and Han (2005), I construct a capital gains overhang control using weekly stock return data. To calculate the value of this control, I first calculate an aggregate reference price equal to: $$R_{t-1}=\frac{1}{k}\sum_{n=1}^{156}\left(V_{t-1-n}\prod_{j=1}^{n-1}(1-V_{t-1-n+j})\right)P_{t-1-n},\label{eq:2}$$ (2) where $$V_t$$ is the week $$t$$ volume turnover, which is calculated as the weekly volume divided by the number of shares outstanding.6$$P_t$$ is the week $$t$$ closing price, or the last daily closing price in the week. The coefficient on $$P_t$$ equals the probability that a share was purchased in period $$t$$ and has not been sold since then. The scaling factor $$k$$ ensures the sum of coefficients on $$P_t$$ equal one. As such, this reference price attempts to proxy for the cost basis of the typical investor in the market. Following Grinblatt and Han (2005), I then define capital gains overhang for firm $$i$$ in week $$t$$ as:7 $$Capital\;Gains\;Overhang_{i,t}=\frac{P_{i,t-2}-R_{t-1}}{P_{t-2}}.\label{eq:3}$$ (3) Finally, I also include month fixed effects as there could be common shocks within a period (e.g., changes in investor discount rates or investor sentiment) as well as within-month correlation across firms in my dummy variables. Some months may have a higher number of realized losses and those months may have negative 1-month returns as a result of aggregate market shocks. For example, tax considerations likely motivate insiders to realize losses in December. Additionally, there is likely correlation across firms as to when bonuses are paid/stock compensation vests, which would influence selling behavior, and also likely correlation as to when firms’ insiders are trading at a loss. In general, I am more interested in the firm-specific information content of insider sales. Finally, to address correlation within an industry/investment category in a particular month, I cluster standard errors by month. Having defined the control variables, I turn to the results. I focus on the FIFO reference price, which is defined in Section 3.2, and Table 4 presents the results. I look at a number of different return horizons, and find robust evidence that sales at a loss have more negative predictive power than sales at a gain. At the one month horizon, I find that a sale below the reference price predicts a return 61 $$(t=-5.32)$$ bps lower than all other firm-months in my sample. On the other hand, a sale above the reference price predicts a 1-month return only 21 bps lower $$(t=-2.58)$$. I reject the null hypothesis that the two are equal at the 1% level $$(F-\mbox{statistic}=7.44)$$. Table 4 Predictive power of sales at a loss relative to sales at a gain 1-month return 6-month return 1-year return 1-month return 1-month return Short-term reversal –2.38*** –1.16 1.84 –2.38*** –2.38*** (–2.71) (–0.86) (0.78) (–2.71) (–2.70) Momentum 0.311 0.611 –0.853 0.311 0.309 (1.63) (1.49) (–1.45) (1.63) (1.62) Book-to-market 0.321*** 1.68*** 3.35*** 0.321*** 0.321*** (3.58) (5.15) (6.97) (3.58) (3.59) Size –0.018 –0.154 –0.422** –0.018 –0.022 (–0.42) (–1.23) (–2.39) (–0.42) (–0.51) Buy (t-1) 0.647*** 1.84*** 2.82*** 0.646*** 0.635*** (7.81) (8.31) (6.61) (7.81) (7.70) Sale below FIFO (t-1) –0.609*** –1.88*** –2.31*** –0.510*** (–5.32) (–5.25) (–3.96) (–4.31) Sale above FIFO (t-1) –0.206*** 0.0898 0.592 –0.204*** (–2.58) (0.42) (1.60) (–2.58) Capital gains overhang 0.326 0.790 –2.01 0.326 0.326 (1.14) (1.09) (–1.55) (1.14) (1.14) More than one person selling below FIFO (t-1) –0.760*** (–2.74) More than one person selling above FIFO (t-1) 0.002 (0.01) Complete liquidation below FIFO (t-1) –4.83*** (–4.29) Complete liquidation above FIFO (t-1) –1.21** (–2.16) R-squared 0.1101 0.1215 0.0965 0.1101 0.1101 F-statistic 7.44*** 20.98*** 16.82*** 11.22*** 9.81*** Number of observations 1,352,175 1,297,278 1,229,697 1,352,175 1,352,175 1-month return 6-month return 1-year return 1-month return 1-month return Short-term reversal –2.38*** –1.16 1.84 –2.38*** –2.38*** (–2.71) (–0.86) (0.78) (–2.71) (–2.70) Momentum 0.311 0.611 –0.853 0.311 0.309 (1.63) (1.49) (–1.45) (1.63) (1.62) Book-to-market 0.321*** 1.68*** 3.35*** 0.321*** 0.321*** (3.58) (5.15) (6.97) (3.58) (3.59) Size –0.018 –0.154 –0.422** –0.018 –0.022 (–0.42) (–1.23) (–2.39) (–0.42) (–0.51) Buy (t-1) 0.647*** 1.84*** 2.82*** 0.646*** 0.635*** (7.81) (8.31) (6.61) (7.81) (7.70) Sale below FIFO (t-1) –0.609*** –1.88*** –2.31*** –0.510*** (–5.32) (–5.25) (–3.96) (–4.31) Sale above FIFO (t-1) –0.206*** 0.0898 0.592 –0.204*** (–2.58) (0.42) (1.60) (–2.58) Capital gains overhang 0.326 0.790 –2.01 0.326 0.326 (1.14) (1.09) (–1.55) (1.14) (1.14) More than one person selling below FIFO (t-1) –0.760*** (–2.74) More than one person selling above FIFO (t-1) 0.002 (0.01) Complete liquidation below FIFO (t-1) –4.83*** (–4.29) Complete liquidation above FIFO (t-1) –1.21** (–2.16) R-squared 0.1101 0.1215 0.0965 0.1101 0.1101 F-statistic 7.44*** 20.98*** 16.82*** 11.22*** 9.81*** Number of observations 1,352,175 1,297,278 1,229,697 1,352,175 1,352,175 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade(s) occurred in the associated firm-month. The reference price FIFO is based on FIFO accounting. Short-term reversal equals the return from $$t-1$$ to $$t$$, and Momentum equals the return from $$t-12$$ to $$t-1$$. Book-to-market is the log value of common equity divided by market capitalization. Size is the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I drop all observations in which the month-end price is less than one. I multiply all coefficient estimates by 100. In the first three columns, the F-statistic tests whether the coefficient on sale below FIFO $$(t-1)$$ equals the coefficient on sale above FIFO $$(t-1)$$. In the fourth column, the F-statistic tests whether multiple sales at a loss predict the same return as multiple sales at a gain. In the fifth column, the F-statistic tests whether the coefficient on complete liquidation below FIFO $$(t-1)$$ equals the coefficient on complete liquidation above FIFO $$(t-1)$$. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than 1 month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 4 Predictive power of sales at a loss relative to sales at a gain 1-month return 6-month return 1-year return 1-month return 1-month return Short-term reversal –2.38*** –1.16 1.84 –2.38*** –2.38*** (–2.71) (–0.86) (0.78) (–2.71) (–2.70) Momentum 0.311 0.611 –0.853 0.311 0.309 (1.63) (1.49) (–1.45) (1.63) (1.62) Book-to-market 0.321*** 1.68*** 3.35*** 0.321*** 0.321*** (3.58) (5.15) (6.97) (3.58) (3.59) Size –0.018 –0.154 –0.422** –0.018 –0.022 (–0.42) (–1.23) (–2.39) (–0.42) (–0.51) Buy (t-1) 0.647*** 1.84*** 2.82*** 0.646*** 0.635*** (7.81) (8.31) (6.61) (7.81) (7.70) Sale below FIFO (t-1) –0.609*** –1.88*** –2.31*** –0.510*** (–5.32) (–5.25) (–3.96) (–4.31) Sale above FIFO (t-1) –0.206*** 0.0898 0.592 –0.204*** (–2.58) (0.42) (1.60) (–2.58) Capital gains overhang 0.326 0.790 –2.01 0.326 0.326 (1.14) (1.09) (–1.55) (1.14) (1.14) More than one person selling below FIFO (t-1) –0.760*** (–2.74) More than one person selling above FIFO (t-1) 0.002 (0.01) Complete liquidation below FIFO (t-1) –4.83*** (–4.29) Complete liquidation above FIFO (t-1) –1.21** (–2.16) R-squared 0.1101 0.1215 0.0965 0.1101 0.1101 F-statistic 7.44*** 20.98*** 16.82*** 11.22*** 9.81*** Number of observations 1,352,175 1,297,278 1,229,697 1,352,175 1,352,175 1-month return 6-month return 1-year return 1-month return 1-month return Short-term reversal –2.38*** –1.16 1.84 –2.38*** –2.38*** (–2.71) (–0.86) (0.78) (–2.71) (–2.70) Momentum 0.311 0.611 –0.853 0.311 0.309 (1.63) (1.49) (–1.45) (1.63) (1.62) Book-to-market 0.321*** 1.68*** 3.35*** 0.321*** 0.321*** (3.58) (5.15) (6.97) (3.58) (3.59) Size –0.018 –0.154 –0.422** –0.018 –0.022 (–0.42) (–1.23) (–2.39) (–0.42) (–0.51) Buy (t-1) 0.647*** 1.84*** 2.82*** 0.646*** 0.635*** (7.81) (8.31) (6.61) (7.81) (7.70) Sale below FIFO (t-1) –0.609*** –1.88*** –2.31*** –0.510*** (–5.32) (–5.25) (–3.96) (–4.31) Sale above FIFO (t-1) –0.206*** 0.0898 0.592 –0.204*** (–2.58) (0.42) (1.60) (–2.58) Capital gains overhang 0.326 0.790 –2.01 0.326 0.326 (1.14) (1.09) (–1.55) (1.14) (1.14) More than one person selling below FIFO (t-1) –0.760*** (–2.74) More than one person selling above FIFO (t-1) 0.002 (0.01) Complete liquidation below FIFO (t-1) –4.83*** (–4.29) Complete liquidation above FIFO (t-1) –1.21** (–2.16) R-squared 0.1101 0.1215 0.0965 0.1101 0.1101 F-statistic 7.44*** 20.98*** 16.82*** 11.22*** 9.81*** Number of observations 1,352,175 1,297,278 1,229,697 1,352,175 1,352,175 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade(s) occurred in the associated firm-month. The reference price FIFO is based on FIFO accounting. Short-term reversal equals the return from $$t-1$$ to $$t$$, and Momentum equals the return from $$t-12$$ to $$t-1$$. Book-to-market is the log value of common equity divided by market capitalization. Size is the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I drop all observations in which the month-end price is less than one. I multiply all coefficient estimates by 100. In the first three columns, the F-statistic tests whether the coefficient on sale below FIFO $$(t-1)$$ equals the coefficient on sale above FIFO $$(t-1)$$. In the fourth column, the F-statistic tests whether multiple sales at a loss predict the same return as multiple sales at a gain. In the fifth column, the F-statistic tests whether the coefficient on complete liquidation below FIFO $$(t-1)$$ equals the coefficient on complete liquidation above FIFO $$(t-1)$$. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than 1 month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. I then consider the same regression model as Equation (1), except that I consider returns at longer horizons. Since many observations are overlapping, and there is more autocorrelation in returns at longer time horizons, I cluster standard errors in two dimensions. Namely, I cluster standard errors by firm and by month (Petersen 2008). At longer horizons, the difference between the return predictability of realized losses and realized gains is even more striking. A sale below the reference price predicts a 6-month return 188 bps lower than all other firm-months, whereas a sale above the reference price predicts a return 9 bps higher. The difference is statistically significant at the 1% level $$(F-\mbox{statistic}=20.98)$$. The difference is even bigger at the 1-year horizon. A sale below the reference price predicts a 1-year return 231 bps lower than all other firm-months, whereas a sale above the reference price predicts a return 59 bps higher. The difference is also statistically significant at the 1% level $$(F-\mbox{statistic}=16.82)$$.8 I present coefficient estimates from various return horizons in Figure 2. Figure 1 View largeDownload slide Magnitude of the sale at a loss or gain This graph charts the return predictability of sales based on distance from the FIFO reference price. The x-axis is a measure of the 1-month return predictability of the associated sale. The y-axis indicates the distance from the reference price. Figure 1 View largeDownload slide Magnitude of the sale at a loss or gain This graph charts the return predictability of sales based on distance from the FIFO reference price. The x-axis is a measure of the 1-month return predictability of the associated sale. The y-axis indicates the distance from the reference price. Figure 2 View largeDownload slide Predictive power of sales based on relation to the FIFO reference price This graph plots the coefficients on the dummy variables Sale below purchase price and Sale above purchase price from estimating: \begin{align*} \begin{split} \textit{Return}_{i,t\rightarrow t+j} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{Reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}\\ &\quad +\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}+\beta_{6}\:\textit{Sale}\;\textit{below}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{7}\:\textit{Sale}\;\textit{above}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{8}\:\textit{Capital}\;\textit{Gains}\;\textit{Overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t.} \end{split} \end{align*} over different return horizons $$j$$. The x-axis displays the time horizon, in months, over which returns are predicted. The y-axis displays the expected difference in returns when the dummy variable equals one compared to when the dummy variable equals zero. Figure 2 View largeDownload slide Predictive power of sales based on relation to the FIFO reference price This graph plots the coefficients on the dummy variables Sale below purchase price and Sale above purchase price from estimating: \begin{align*} \begin{split} \textit{Return}_{i,t\rightarrow t+j} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{Reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}\\ &\quad +\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}+\beta_{6}\:\textit{Sale}\;\textit{below}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{7}\:\textit{Sale}\;\textit{above}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{8}\:\textit{Capital}\;\textit{Gains}\;\textit{Overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t.} \end{split} \end{align*} over different return horizons $$j$$. The x-axis displays the time horizon, in months, over which returns are predicted. The y-axis displays the expected difference in returns when the dummy variable equals one compared to when the dummy variable equals zero. Under my hypothesis, it is plausible that firm-months which have more than one insider sell at a loss will predict more negative returns than firm-months which have exactly one insider sell at a loss. To test this, I add dummy variables to Equation (1). Specifically, I add dummy variables that equal one if there was more than one insider who sold at a loss (gain) relative to the reference price in the associated firm-month. I estimate the following equation for firm $$i$$ in month $$t$$: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}\\ &\quad +\beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}\\ &\quad +\beta_{6}\:\textit{Sale}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ & \quad+\beta_{7}\:\textit{Sale}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{8}\:\textit{Multiple}\;\textit{insiders}\;\textit{selling}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{9}\:\textit{Multiple}\;\textit{insiders}\;\textit{selling}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{10}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t}.\label{eq:4} \end{split} \end{align} (4) My prediction is that firm-months which have more than one insider realizing a loss will have more negative predictive power than firm-months with only one insider realizing a loss. That is, I predict that $$\beta_8<0$$. The results are consistent with this prediction. I estimate $$\beta_{8}$$ equal to $$-$$0.76 $$(t=-2.74)$$. That is, a firm-month which has more than one insider sell below the reference price is associated with a 1-month return 76 bps lower than a firm-month that has exactly one insider sell below the reference price. I also find that if more than one insider sells above the reference price, this predicts a 1-month return of the same magnitude as if only one insider sold at a gain. The predictive power of realized losses allows for the construction of portfolio strategies that, at least before transaction costs, earn alpha. At the end of each month $$t$$, I construct an equal-weighted portfolio that goes long a firm if one of its insiders sold shares at a gain relative to the FIFO reference price in month $$(t-1)$$ and go short a firm if one of its insiders sold at a loss relative to the FIFO reference price in month $$(t-1)$$. As shown in Table 5, I find that an equal-weighted portfolio earns a four-factor alpha of 67 bps per month with a t-statistic equal to 3.85, significant at the 1% level. I list other equal-weighted portfolio alphas in Table 5. Value-weighted portfolios do not earn alpha. This suggests that the greater negative predictive power of sales at a loss relative to sales at a gain is concentrated within small firms. Table 5 Portfolio returns 1-month return 1-month return 1-month return Intercept 0.60*** 0.67*** 0.67*** (3.63) (4.05) (3.85) Excess market return –0.09** –0.09** (–2.35) (–2.29) SMB –0.07 –0.07 (–1.35) (–1.35) HML –0.06 –0.06 (–1.05) (–1.03) Momentum –0.00 (–0.06) R-squared 0.000 0.0240 0.0240 N 351 351 351 1-month return 1-month return 1-month return Intercept 0.60*** 0.67*** 0.67*** (3.63) (4.05) (3.85) Excess market return –0.09** –0.09** (–2.35) (–2.29) SMB –0.07 –0.07 (–1.35) (–1.35) HML –0.06 –0.06 (–1.05) (–1.03) Momentum –0.00 (–0.06) R-squared 0.000 0.0240 0.0240 N 351 351 351 This table presents regressions of portfolio returns on return factors from Kenneth French’s Web site. The portfolio returns are monthly—end of month t to the end of month $$(t+1)$$—returns earned by an equal-weighted portfolio that goes long firms that had an insider sell at a gain relative to the FIFO reference price in month $$(t-1)$$ and goes short firms that had an insider sell at a loss relative to the FIFO reference price in month $$(t-1)$$. I require at least 5 stocks in both the long and short sides of the portfolio. The sample period runs from 1986 to 2015. I drop all observations in which the month-end price is less than one dollar. I put t-statistics in parentheses and multiply all coefficients by 100. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 5 Portfolio returns 1-month return 1-month return 1-month return Intercept 0.60*** 0.67*** 0.67*** (3.63) (4.05) (3.85) Excess market return –0.09** –0.09** (–2.35) (–2.29) SMB –0.07 –0.07 (–1.35) (–1.35) HML –0.06 –0.06 (–1.05) (–1.03) Momentum –0.00 (–0.06) R-squared 0.000 0.0240 0.0240 N 351 351 351 1-month return 1-month return 1-month return Intercept 0.60*** 0.67*** 0.67*** (3.63) (4.05) (3.85) Excess market return –0.09** –0.09** (–2.35) (–2.29) SMB –0.07 –0.07 (–1.35) (–1.35) HML –0.06 –0.06 (–1.05) (–1.03) Momentum –0.00 (–0.06) R-squared 0.000 0.0240 0.0240 N 351 351 351 This table presents regressions of portfolio returns on return factors from Kenneth French’s Web site. The portfolio returns are monthly—end of month t to the end of month $$(t+1)$$—returns earned by an equal-weighted portfolio that goes long firms that had an insider sell at a gain relative to the FIFO reference price in month $$(t-1)$$ and goes short firms that had an insider sell at a loss relative to the FIFO reference price in month $$(t-1)$$. I require at least 5 stocks in both the long and short sides of the portfolio. The sample period runs from 1986 to 2015. I drop all observations in which the month-end price is less than one dollar. I put t-statistics in parentheses and multiply all coefficients by 100. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. 3.3 Earnings Having established that sales at a loss predict more negative returns than sales at a gain, a natural follow-up question is to determine why sales at a loss have more negative predictive power than sales at a gain. I consider a sample of all firm-months that are either one month or two months before the month of an earnings announcement and follow the previous earnings announcement. I examine whether sales at a loss predict more negative earnings surprises than sales at a gain. Specifically, I run regressions of the earnings surprise on dummy variables that indicate whether a firm-month had the associated sale at a loss or sale at a gain. I estimate regressions of the following form for firm $$i$$ in fiscal year-quarter $$t$$ during month $$t'$$: \begin{align} \begin{split} \textit{Earnings}\;\textit{Surprise}_{i,t} & = \beta_{0}+\beta_{1}\:\textit{Sale}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t,t'}\\ &\quad +\beta_{2}\:\textit{Sale}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t,t'}\\ &\quad +\gamma_{1}\:\textit{Controls}+\gamma_{2}\:\textit{Fiscal}\;\textit{year-quarter}+\epsilon_{i,t}.\label{eq:5} \end{split} \end{align} (5) The dependent variable is equal to actual earnings per share minus the median earnings per share estimate. Sale below reference price – a dummy which takes a value of one if an insider in firm $$i$$ sold stock at a loss relative to the reference price in the current month – and the analogous Sale above reference price are my independent variables. My prediction is that $$\beta_{1}$$ is less than $$\beta_{2}$$. In other words, I predict that a sale at a loss has more negative predictive power for the subsequent earnings announcement than a sale at a gain. I conduct my analysis with common earnings surprise controls – a control for each of the past four earnings surprises, the dispersion of earnings estimates, the number of earnings estimates, and a dummy that equals one if there was only one earnings estimate – and a dummy that equals one if there was an insider purchase.9 The evidence confirms my prediction that realized losses predict more negative earnings surprises than realized gains; I present the results in Table 6. A sale of stock at a loss relative to the FIFO reference price predicts a future earnings surprise about one cent lower than a sale at a gain. Clustering standard errors by fiscal year-quarter and firm, I find that this result is statistically significant at the 1% level $$(F-\mbox{statistic}=52.56)$$. I find similar, though weaker, results when I include common controls for stock characteristics – the return in the month before the earnings announcement, the return from month $$t-12$$ to month $$t-1$$ before the earnings announcement, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang. Namely, a sale of stock at a loss relative to the FIFO reference price predicts a future earnings surprise about 0.3 cents lower than a sale at a gain. For comparison purposes, this difference is slightly larger than the magnitude of the positive predictive power of an insider purchase for the upcoming earnings announcement. I find that this difference is statistically significant at the 5% level $$(F-\mbox{statistic}=4.03)$$. Finally, I add firm fixed effects. A sale of stock at a loss predicts a future earnings surprise about 0.4 cents lower than a sale at a gain. Again, this result is significant at the 5% level $$(F-\mbox{statistic}=5.32)$$. Table 6 Earnings surprises Earnings surprise Earnings surprise Earnings surprise Earnings surprise Earnings surprise Sale below FIFO reference price –0.801*** –0.963*** –0.405*** –0.168 –0.125 (–4.42) (–5.65) (–2.89) (–1.15) (–0.87) Sale above FIFO reference price 2.107*** 1.555*** 0.774*** 0.160** 0.268*** (16.74) (14.34) (9.15) (1.98) (3.11) Buy –0.055 0.168** 0.284*** 0.423*** (–0.58) (2.03) (3.33) (5.32) Single estimate –3.183*** –2.173*** –1.541 –1.83*** (–15.24) (–12.33) (–8.60) (–8.80) Dispersion –39.15*** –27.28*** –27.04*** –40.24*** (–10.72) (–9.52) (–9.13) (–12.13) Number of estimates 0.172*** 0.104*** 0.0135 0.022 (14.53) (13.37) (1.30) (1.31) Earnings surprise (t-1) 0.213*** 0.198*** 0.145*** (22.54) (21.05) (15.37) Earnings surprise (t-2) 0.102*** 0.0950*** 0.0572*** (15.48) (13.69) (8.00) Earnings surprise (t-3) 0.0649*** 0.0601*** 0.0290*** (12.41) (11.00) (5.08) Earnings surprise (t-4) 0.0764*** 0.0704*** 0.0389*** (12.80) (11.45) (6.33) Stock characteristics No No No Yes Yes Fiscal year-quarter FEs Yes Yes Yes Yes Yes Firm FEs No No No No Yes F-statistic 186.77*** 171.07*** 52.56*** 4.03** 5.32** R-squared 0.0182 0.0384 0.1089 0.1181 0.1720 Number of observations 652,492 652,492 524,390 450,613 450,521 Earnings surprise Earnings surprise Earnings surprise Earnings surprise Earnings surprise Sale below FIFO reference price –0.801*** –0.963*** –0.405*** –0.168 –0.125 (–4.42) (–5.65) (–2.89) (–1.15) (–0.87) Sale above FIFO reference price 2.107*** 1.555*** 0.774*** 0.160** 0.268*** (16.74) (14.34) (9.15) (1.98) (3.11) Buy –0.055 0.168** 0.284*** 0.423*** (–0.58) (2.03) (3.33) (5.32) Single estimate –3.183*** –2.173*** –1.541 –1.83*** (–15.24) (–12.33) (–8.60) (–8.80) Dispersion –39.15*** –27.28*** –27.04*** –40.24*** (–10.72) (–9.52) (–9.13) (–12.13) Number of estimates 0.172*** 0.104*** 0.0135 0.022 (14.53) (13.37) (1.30) (1.31) Earnings surprise (t-1) 0.213*** 0.198*** 0.145*** (22.54) (21.05) (15.37) Earnings surprise (t-2) 0.102*** 0.0950*** 0.0572*** (15.48) (13.69) (8.00) Earnings surprise (t-3) 0.0649*** 0.0601*** 0.0290*** (12.41) (11.00) (5.08) Earnings surprise (t-4) 0.0764*** 0.0704*** 0.0389*** (12.80) (11.45) (6.33) Stock characteristics No No No Yes Yes Fiscal year-quarter FEs Yes Yes Yes Yes Yes Firm FEs No No No No Yes F-statistic 186.77*** 171.07*** 52.56*** 4.03** 5.32** R-squared 0.0182 0.0384 0.1089 0.1181 0.1720 Number of observations 652,492 652,492 524,390 450,613 450,521 This table shows the regressions of earnings surprises on insider sales and control variables. The dependent variable is the quarterly earnings surprise measured as the number of cents that the actual earnings per share differs from the median earnings per share estimate. All buy and sell variables are dummies that equal one if the stated trade occurred in the associated month. I consider all firm-months that are in the 2 months before an earnings announcement and follow the previous earnings announcement. Single estimate is a dummy that equals one if there was only one forecast for the associated firm-quarter. Dispersion equals the standard deviation of earnings forecasts. Number of estimates equals the number of analyst forecasts. I winsorize all earnings surprise variables, dispersion, and coverage at the 1% level. Stock characteristics include the return in the month before the earnings announcement, the return from month $$t-12$$ to month $$t-1$$ before the earnings announcement, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. The F-statistic tests whether the coefficient on sale below FIFO reference price equals the coefficient on sale above FIFO reference price. Standard errors are clustered by fiscal year-quarter and firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 6 Earnings surprises Earnings surprise Earnings surprise Earnings surprise Earnings surprise Earnings surprise Sale below FIFO reference price –0.801*** –0.963*** –0.405*** –0.168 –0.125 (–4.42) (–5.65) (–2.89) (–1.15) (–0.87) Sale above FIFO reference price 2.107*** 1.555*** 0.774*** 0.160** 0.268*** (16.74) (14.34) (9.15) (1.98) (3.11) Buy –0.055 0.168** 0.284*** 0.423*** (–0.58) (2.03) (3.33) (5.32) Single estimate –3.183*** –2.173*** –1.541 –1.83*** (–15.24) (–12.33) (–8.60) (–8.80) Dispersion –39.15*** –27.28*** –27.04*** –40.24*** (–10.72) (–9.52) (–9.13) (–12.13) Number of estimates 0.172*** 0.104*** 0.0135 0.022 (14.53) (13.37) (1.30) (1.31) Earnings surprise (t-1) 0.213*** 0.198*** 0.145*** (22.54) (21.05) (15.37) Earnings surprise (t-2) 0.102*** 0.0950*** 0.0572*** (15.48) (13.69) (8.00) Earnings surprise (t-3) 0.0649*** 0.0601*** 0.0290*** (12.41) (11.00) (5.08) Earnings surprise (t-4) 0.0764*** 0.0704*** 0.0389*** (12.80) (11.45) (6.33) Stock characteristics No No No Yes Yes Fiscal year-quarter FEs Yes Yes Yes Yes Yes Firm FEs No No No No Yes F-statistic 186.77*** 171.07*** 52.56*** 4.03** 5.32** R-squared 0.0182 0.0384 0.1089 0.1181 0.1720 Number of observations 652,492 652,492 524,390 450,613 450,521 Earnings surprise Earnings surprise Earnings surprise Earnings surprise Earnings surprise Sale below FIFO reference price –0.801*** –0.963*** –0.405*** –0.168 –0.125 (–4.42) (–5.65) (–2.89) (–1.15) (–0.87) Sale above FIFO reference price 2.107*** 1.555*** 0.774*** 0.160** 0.268*** (16.74) (14.34) (9.15) (1.98) (3.11) Buy –0.055 0.168** 0.284*** 0.423*** (–0.58) (2.03) (3.33) (5.32) Single estimate –3.183*** –2.173*** –1.541 –1.83*** (–15.24) (–12.33) (–8.60) (–8.80) Dispersion –39.15*** –27.28*** –27.04*** –40.24*** (–10.72) (–9.52) (–9.13) (–12.13) Number of estimates 0.172*** 0.104*** 0.0135 0.022 (14.53) (13.37) (1.30) (1.31) Earnings surprise (t-1) 0.213*** 0.198*** 0.145*** (22.54) (21.05) (15.37) Earnings surprise (t-2) 0.102*** 0.0950*** 0.0572*** (15.48) (13.69) (8.00) Earnings surprise (t-3) 0.0649*** 0.0601*** 0.0290*** (12.41) (11.00) (5.08) Earnings surprise (t-4) 0.0764*** 0.0704*** 0.0389*** (12.80) (11.45) (6.33) Stock characteristics No No No Yes Yes Fiscal year-quarter FEs Yes Yes Yes Yes Yes Firm FEs No No No No Yes F-statistic 186.77*** 171.07*** 52.56*** 4.03** 5.32** R-squared 0.0182 0.0384 0.1089 0.1181 0.1720 Number of observations 652,492 652,492 524,390 450,613 450,521 This table shows the regressions of earnings surprises on insider sales and control variables. The dependent variable is the quarterly earnings surprise measured as the number of cents that the actual earnings per share differs from the median earnings per share estimate. All buy and sell variables are dummies that equal one if the stated trade occurred in the associated month. I consider all firm-months that are in the 2 months before an earnings announcement and follow the previous earnings announcement. Single estimate is a dummy that equals one if there was only one forecast for the associated firm-quarter. Dispersion equals the standard deviation of earnings forecasts. Number of estimates equals the number of analyst forecasts. I winsorize all earnings surprise variables, dispersion, and coverage at the 1% level. Stock characteristics include the return in the month before the earnings announcement, the return from month $$t-12$$ to month $$t-1$$ before the earnings announcement, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. The F-statistic tests whether the coefficient on sale below FIFO reference price equals the coefficient on sale above FIFO reference price. Standard errors are clustered by fiscal year-quarter and firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Overall, this evidence suggests that part of the negative return predictability of insider sales at a loss relative to insider sales at a gain is driven by information regarding the upcoming earnings announcement. 3.4 Individual-level regressions Having confirmed my prediction that sales at a loss have more predictive power than sales at a gain, I next examine whether an aversion to realizing losses is driving my result. That is, does my basic result stem from the fact that insiders require a stronger negative signal to sell at a loss than to sell at a gain? Over the next three subsections, I test alternatives to this view. In this subsection, I examine investor heterogeneity and insider position heterogeneity. In the next subsection, I look at whether my basic result is a consequence of the short swing rule. Finally, I examine whether my result stems from a correlation with the price path. Insiders surely differ in the sophistication of their trading. It seems possible that more sophisticated traders will make trades that are more informative about future returns, and also that they will exhibit the disposition effect less because they know it to be a mistake. If this is the case, we would then observe that sales at a loss are more informative than sales at a gain, but this would not be driven by any aversion to realizing losses. To address this concern, I estimate models with individual-firm fixed effects. That is, I compare the predictive power of realized losses relative to realized gains within each insider-firm combination. I consider a panel of all insider-firm-months where the insider makes a sale; doing so ensures that individual-firm fixed effects subtract the average predictability of a sale by the insider. Again, I control for common shocks within a particular month by using month fixed effects.10 I cluster standard errors by month when the dependent variable is the 1-month horizon. When the dependent variable is a return horizon longer than 1-month, I cluster standard errors by firm and by month. Specifically, I estimate the following equation for individual-firm $$i$$ in month $$t$$: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}\\ &\quad +\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}\\ &\quad +\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}\\ & \quad+\beta_{6}\:\textit{Sale}\;\textit{below}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{7}\:\textit{Sale}\;\textit{above}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{8}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}\\ & \quad+\gamma_{1}\:\textit{Individual-firm}+\gamma_{2}\:\textit{Month}+\epsilon_{i,t.}\label{eq:6} \end{split} \end{align} (6) I predict that $$\beta_{5}<\beta_{6}.$$ That is, I expect to see evidence within-insider that sales at a loss have more negative predictive power than sales at a gain. I present the results in the first two columns of Table 7. Within insider, I find that insider sales at a loss predict 1-month returns 33 bps lower than insider sales at a gain. The difference is statistically significant at the 10% level $$(F-\mbox{statistic}=3.03)$$. At the 1-year return horizon, the difference in return predictability between insider sales at a loss and insider sales at a gain within insider is almost 2%. This difference is statistically significant at the 5% level $$(F-\mbox{statistic}=5.62)$$. Table 7 Individual-level analysis 1-month return 1-year return 1-month return 1-year return Sale below FIFO reference price (t-1) –0.022 –0.72 (–0.11) (–0.89) Sale above FIFO reference price (t-1) 0.306** 1.16* (2.18) (1.92) Sale below moving average (t-1) –0.12 –0.68 (–1.10) (–1.65) Sale below moving average by purchaser (t-1) –0.45*** –1.17* (–2.96) (–1.87) Sale above moving average (t-1) –0.44*** –1.30*** (–5.79) (–4.48) Sale above moving average by purchaser (t-1) 0.19* 0.65 (1.82) (1.50) Purchaser (t-1) 0.05 –0.63 (0.51) (–1.47) F-statistic 3.03* 5.61** 2.84* 2.47 Controls Yes Yes Yes Yes R-squared 0.4080 0.6040 0.1749 0.3779 Adjusted R-squared 0.1941 0.4598 0.1573 0.3646 Number of observations 140,859 133,710 6,937,616 6,625,582 1-month return 1-year return 1-month return 1-year return Sale below FIFO reference price (t-1) –0.022 –0.72 (–0.11) (–0.89) Sale above FIFO reference price (t-1) 0.306** 1.16* (2.18) (1.92) Sale below moving average (t-1) –0.12 –0.68 (–1.10) (–1.65) Sale below moving average by purchaser (t-1) –0.45*** –1.17* (–2.96) (–1.87) Sale above moving average (t-1) –0.44*** –1.30*** (–5.79) (–4.48) Sale above moving average by purchaser (t-1) 0.19* 0.65 (1.82) (1.50) Purchaser (t-1) 0.05 –0.63 (0.51) (–1.47) F-statistic 3.03* 5.61** 2.84* 2.47 Controls Yes Yes Yes Yes R-squared 0.4080 0.6040 0.1749 0.3779 Adjusted R-squared 0.1941 0.4598 0.1573 0.3646 Number of observations 140,859 133,710 6,937,616 6,625,582 This table presents regressions of returns on insider sales with controls at the individual level. I consider a panel of all insider-firm combinations. In the first two columns, I include all insider-firm-month observations in my sample where the insider executed a sale. In the final two columns, I include all insider-firm-month observations for insiders from the month of their first transaction at the firm to the month of their last transaction in my sample. All sale variables are dummies that equal one if there was the relevant trade in the associated individual-firm-month. The FIFO reference price is a reference price constructed using FIFO (first-in first-out) accounting. The moving average reference price equals the average of the previous 6-month-end prices. A purchaser is an insider who has made a purchase. I include insider-firm and month fixed effects. Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, a dummy that equals one if the insider made a purchase in month $$t-1$$, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. In the final four columns, I also include a dummy that equals one if month $$(t-2)$$’s closing price was below the moving-average reference price. I multiply all coefficient estimates by 100. In the first two columns, the F-statistic tests whether the coefficient on sale below FIFO reference price $$(t-1)$$ equals the coefficient on sale above FIFO reference $$(t-1)$$ price. In the last two columns, the F-statistic tests whether the predictive power of a sale below the moving average reference price by a purchaser $$(t-1)$$ equals the predictive power of a sale above the moving average by a purchaser $$(t-1)$$. When the return horizon is 1 month, I cluster standard errors by month; when the return horizon is longer, I cluster standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 7 Individual-level analysis 1-month return 1-year return 1-month return 1-year return Sale below FIFO reference price (t-1) –0.022 –0.72 (–0.11) (–0.89) Sale above FIFO reference price (t-1) 0.306** 1.16* (2.18) (1.92) Sale below moving average (t-1) –0.12 –0.68 (–1.10) (–1.65) Sale below moving average by purchaser (t-1) –0.45*** –1.17* (–2.96) (–1.87) Sale above moving average (t-1) –0.44*** –1.30*** (–5.79) (–4.48) Sale above moving average by purchaser (t-1) 0.19* 0.65 (1.82) (1.50) Purchaser (t-1) 0.05 –0.63 (0.51) (–1.47) F-statistic 3.03* 5.61** 2.84* 2.47 Controls Yes Yes Yes Yes R-squared 0.4080 0.6040 0.1749 0.3779 Adjusted R-squared 0.1941 0.4598 0.1573 0.3646 Number of observations 140,859 133,710 6,937,616 6,625,582 1-month return 1-year return 1-month return 1-year return Sale below FIFO reference price (t-1) –0.022 –0.72 (–0.11) (–0.89) Sale above FIFO reference price (t-1) 0.306** 1.16* (2.18) (1.92) Sale below moving average (t-1) –0.12 –0.68 (–1.10) (–1.65) Sale below moving average by purchaser (t-1) –0.45*** –1.17* (–2.96) (–1.87) Sale above moving average (t-1) –0.44*** –1.30*** (–5.79) (–4.48) Sale above moving average by purchaser (t-1) 0.19* 0.65 (1.82) (1.50) Purchaser (t-1) 0.05 –0.63 (0.51) (–1.47) F-statistic 3.03* 5.61** 2.84* 2.47 Controls Yes Yes Yes Yes R-squared 0.4080 0.6040 0.1749 0.3779 Adjusted R-squared 0.1941 0.4598 0.1573 0.3646 Number of observations 140,859 133,710 6,937,616 6,625,582 This table presents regressions of returns on insider sales with controls at the individual level. I consider a panel of all insider-firm combinations. In the first two columns, I include all insider-firm-month observations in my sample where the insider executed a sale. In the final two columns, I include all insider-firm-month observations for insiders from the month of their first transaction at the firm to the month of their last transaction in my sample. All sale variables are dummies that equal one if there was the relevant trade in the associated individual-firm-month. The FIFO reference price is a reference price constructed using FIFO (first-in first-out) accounting. The moving average reference price equals the average of the previous 6-month-end prices. A purchaser is an insider who has made a purchase. I include insider-firm and month fixed effects. Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, a dummy that equals one if the insider made a purchase in month $$t-1$$, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. In the final four columns, I also include a dummy that equals one if month $$(t-2)$$’s closing price was below the moving-average reference price. I multiply all coefficient estimates by 100. In the first two columns, the F-statistic tests whether the coefficient on sale below FIFO reference price $$(t-1)$$ equals the coefficient on sale above FIFO reference $$(t-1)$$ price. In the last two columns, the F-statistic tests whether the predictive power of a sale below the moving average reference price by a purchaser $$(t-1)$$ equals the predictive power of a sale above the moving average by a purchaser $$(t-1)$$. When the return horizon is 1 month, I cluster standard errors by month; when the return horizon is longer, I cluster standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. It is likely that insider characteristics are not static and that these characteristics could relate to the insider’s position in the firm. It is possible that different insider types exhibit less of a disposition effect, and have more predictive trades. In particular, this is a concern as CFO trades are more informative for future returns than CEO trades (Wang et al. 2012). I investigate this alternative explanation by considering the behavior of different types of insiders. In Table 8, I present the propensity to sell at a gain and the propensity to sell at a loss for different types of insiders. I consider insider types that occupy at least 5% of my sample, and all exhibit the disposition effect. Beneficial owners exhibit the weakest disposition effect. When I exclude the sales of beneficial owners from my analysis, the results from estimating Equation (1) are very similar to the results from estimating Equation (1) when I include the sales of beneficial owners.11 Table 8 Insider role in the firm Number of Acquisitions, Sales and Purchase Days PGR PLR Ratio Director 556,714 0.0132 0.0073 1.80 Officer 301,046 0.0347 0.0215 1.61 Vice president 103,951 0.0229 0.0134 1.70 Beneficial owner 122,548 0.0410 0.0349 1.17 CEO 91,376 0.0302 0.0148 2.03 CFO 71,351 0.0317 0.0163 1.94 Number of Acquisitions, Sales and Purchase Days PGR PLR Ratio Director 556,714 0.0132 0.0073 1.80 Officer 301,046 0.0347 0.0215 1.61 Vice president 103,951 0.0229 0.0134 1.70 Beneficial owner 122,548 0.0410 0.0349 1.17 CEO 91,376 0.0302 0.0148 2.03 CFO 71,351 0.0317 0.0163 1.94 This table looks at the characteristics of insiders by their position in the firm. I present statistics for types of insider positions that occupy at least 5% of the sample. The table presents the number of acquisitions, sales, and purchase days by insider type. PGR, the proportion of gains realized; PLR, the proportion of losses realized; Ratio, PGR divided by PLR. Table 8 Insider role in the firm Number of Acquisitions, Sales and Purchase Days PGR PLR Ratio Director 556,714 0.0132 0.0073 1.80 Officer 301,046 0.0347 0.0215 1.61 Vice president 103,951 0.0229 0.0134 1.70 Beneficial owner 122,548 0.0410 0.0349 1.17 CEO 91,376 0.0302 0.0148 2.03 CFO 71,351 0.0317 0.0163 1.94 Number of Acquisitions, Sales and Purchase Days PGR PLR Ratio Director 556,714 0.0132 0.0073 1.80 Officer 301,046 0.0347 0.0215 1.61 Vice president 103,951 0.0229 0.0134 1.70 Beneficial owner 122,548 0.0410 0.0349 1.17 CEO 91,376 0.0302 0.0148 2.03 CFO 71,351 0.0317 0.0163 1.94 This table looks at the characteristics of insiders by their position in the firm. I present statistics for types of insider positions that occupy at least 5% of the sample. The table presents the number of acquisitions, sales, and purchase days by insider type. PGR, the proportion of gains realized; PLR, the proportion of losses realized; Ratio, PGR divided by PLR. Next, I look at the difference in return predictability between realized gains and realized losses for different types of insiders. I again consider a sample of individual-firm-months with a sale and reestimate Equation (6) for different types of insiders. Table 9 presents the difference in return predictability between sales at a loss and sales at a gain for different types of insiders. There is limited identification because I am looking within insider in a small subsample based on insider type. As such, it is not surprising that I do not find much statistical significance. That being said, I find that sales at a loss predict more negative returns than sales at a gain for every insider type at the 1-year horizon, except for beneficial owners. It is not shocking that a sale at a loss by a beneficial owner does not predict more negative returns than a sale at a gain by a beneficial owner as beneficial owners exhibit the weakest disposition effect in my sample. It is surprising that we see a large difference in return predictability between a sale at a loss by a director and a sale at a gain by a director as we typically think of directors as being at the bottom of the information pyramid in a firm (Seyhun 1998). However, a deeper examination suggests that the trades of directors in my sample have predictive power. A purchase by a director in the lagged month predicts a 1-year return about 2.47% higher than all other firm-months. Table 9 Return predictability by insider type Difference at a 1-month horizon Difference at 1-year horizon Director 0.44 –1.78 Officer –0.33 –2.01* Vice president –1.25** –0.52 Beneficial owner –0.45 4.35 CEO –0.20 –1.01 CFO –0.53 –2.53 Difference at a 1-month horizon Difference at 1-year horizon Director 0.44 –1.78 Officer –0.33 –2.01* Vice president –1.25** –0.52 Beneficial owner –0.45 4.35 CEO –0.20 –1.01 CFO –0.53 –2.53 In this table, we demonstrate the difference in return predictability between sales at a loss relative to the FIFO reference price and sales at a gain relative to the FIFO reference price by insider type. I include all insider-firm-month observations in my sample where the insider executed a sale. I then create subsamples based on insider type. I present statistics for insider position types that occupy at least 5% of the sample. Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, a dummy that equals one if the insider made a purchase in month $$t-1$$, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. I include insider-firm and month fixed effects. When the return horizon is 1 month, I cluster standard errors by month; when the return horizon is longer, I cluster standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 9 Return predictability by insider type Difference at a 1-month horizon Difference at 1-year horizon Director 0.44 –1.78 Officer –0.33 –2.01* Vice president –1.25** –0.52 Beneficial owner –0.45 4.35 CEO –0.20 –1.01 CFO –0.53 –2.53 Difference at a 1-month horizon Difference at 1-year horizon Director 0.44 –1.78 Officer –0.33 –2.01* Vice president –1.25** –0.52 Beneficial owner –0.45 4.35 CEO –0.20 –1.01 CFO –0.53 –2.53 In this table, we demonstrate the difference in return predictability between sales at a loss relative to the FIFO reference price and sales at a gain relative to the FIFO reference price by insider type. I include all insider-firm-month observations in my sample where the insider executed a sale. I then create subsamples based on insider type. I present statistics for insider position types that occupy at least 5% of the sample. Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, a dummy that equals one if the insider made a purchase in month $$t-1$$, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. I include insider-firm and month fixed effects. When the return horizon is 1 month, I cluster standard errors by month; when the return horizon is longer, I cluster standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. I interpret this evidence as suggestive that my result is not being driven by investor heterogeneity or insider position heterogeneity. 3.5 Short swing rule Having considered insider heterogeneity as the driver of my basic result, I now examine whether it may instead be driven by the short swing rule. This rule (15 U.S. Code § 78p) states that all profits realized by an insider from executing two offsetting transactions within a 6-month period (a buy and a subsequent sale or a sale and a subsequent buy) are recoverable by the issuer.12 As such, it is unlikely that an insider would purchase or sell stock if she had any intention to complete an offsetting transaction in the near future. Therefore, in the six months following a purchase, it is unlikely that an insider will feel the need to sell shares for liquidity or diversification reasons. As such, sales executed shortly after purchases are presumably very informative. And, as a consequence of the regulatory environment (i.e., the short swing rule), insiders are strictly penalized for realizing gains. This could drive my result as the legal environment prevents realized gains and permits realized losses during a window where insiders likely only make informed trades. I test this alternative explanation by looking at the information content of realized losses made at least six months after the most recent purchase and the information content of realized gains made at least six months after the most recent purchase. I construct a dummy variable that equals one if there was a realized loss (gain) at least 180 days after the most recent purchase in month $$(t-1)$$. For comparison purposes (the short swing rule assumes a sort of LIFO accounting), I use the most recent purchase price—the transaction that an insider would be most worried about offsetting—as the reference price. I also consider this a natural reference price, because it is likely very salient to the insider due to the recency bias. I find minimal difference in the return predictability difference between sales at a loss and sales at a gain when I consider sales that happen long after the purchase, and the return predictability difference when I consider all sales. I present the results in Table 10. Table 10 Short swing rule 1-month return 1-month return 6-month return 6-month return 1-year return 1-year return Buy (t-1) 0.651*** 0.642*** 1.87*** 1.84*** 2.86*** 2.82*** (7.86) (7.77) (8.45) (8.32) (6.71) (6.64) Sale below purchase (t-1) –0.624*** –2.67*** –3.29*** (–3.87) (–5.01) (–3.66) Sale above purchase (t-1) –0.394*** –0.644*** –0.615 (–4.16) (–2.64) (–1.54) Slow sale below purchase (t-1) –0.620*** –2.67*** –3.30*** (–3.45) (–4.54) (–3.21) Slow sale above purchase (t-1) –0.403*** –0.686*** –0.615 (–4.16) (–2.64) (–1.49) Controls Yes Yes Yes Yes Yes Yes R-squared 0.1101 0.1101 0.1215 0.1215 0.0965 0.0965 F-statistic 1.32 1.05 12.02*** 9.52*** 6.91*** 5.48** Number of observations 1,352,175 1,352,175 1,297,278 1,297,278 1,229,697 1,229,697 1-month return 1-month return 6-month return 6-month return 1-year return 1-year return Buy (t-1) 0.651*** 0.642*** 1.87*** 1.84*** 2.86*** 2.82*** (7.86) (7.77) (8.45) (8.32) (6.71) (6.64) Sale below purchase (t-1) –0.624*** –2.67*** –3.29*** (–3.87) (–5.01) (–3.66) Sale above purchase (t-1) –0.394*** –0.644*** –0.615 (–4.16) (–2.64) (–1.54) Slow sale below purchase (t-1) –0.620*** –2.67*** –3.30*** (–3.45) (–4.54) (–3.21) Slow sale above purchase (t-1) –0.403*** –0.686*** –0.615 (–4.16) (–2.64) (–1.49) Controls Yes Yes Yes Yes Yes Yes R-squared 0.1101 0.1101 0.1215 0.1215 0.0965 0.0965 F-statistic 1.32 1.05 12.02*** 9.52*** 6.91*** 5.48** Number of observations 1,352,175 1,352,175 1,297,278 1,297,278 1,229,697 1,229,697 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade occurred in the associated firm-month. The reference price Purchase equals the most recent purchase price. To indicate that the sale happened at least 180 days after the most recent purchase, I use the adjective “slow.” Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. I multiply all coefficient estimates by 100. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than one month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 10 Short swing rule 1-month return 1-month return 6-month return 6-month return 1-year return 1-year return Buy (t-1) 0.651*** 0.642*** 1.87*** 1.84*** 2.86*** 2.82*** (7.86) (7.77) (8.45) (8.32) (6.71) (6.64) Sale below purchase (t-1) –0.624*** –2.67*** –3.29*** (–3.87) (–5.01) (–3.66) Sale above purchase (t-1) –0.394*** –0.644*** –0.615 (–4.16) (–2.64) (–1.54) Slow sale below purchase (t-1) –0.620*** –2.67*** –3.30*** (–3.45) (–4.54) (–3.21) Slow sale above purchase (t-1) –0.403*** –0.686*** –0.615 (–4.16) (–2.64) (–1.49) Controls Yes Yes Yes Yes Yes Yes R-squared 0.1101 0.1101 0.1215 0.1215 0.0965 0.0965 F-statistic 1.32 1.05 12.02*** 9.52*** 6.91*** 5.48** Number of observations 1,352,175 1,352,175 1,297,278 1,297,278 1,229,697 1,229,697 1-month return 1-month return 6-month return 6-month return 1-year return 1-year return Buy (t-1) 0.651*** 0.642*** 1.87*** 1.84*** 2.86*** 2.82*** (7.86) (7.77) (8.45) (8.32) (6.71) (6.64) Sale below purchase (t-1) –0.624*** –2.67*** –3.29*** (–3.87) (–5.01) (–3.66) Sale above purchase (t-1) –0.394*** –0.644*** –0.615 (–4.16) (–2.64) (–1.54) Slow sale below purchase (t-1) –0.620*** –2.67*** –3.30*** (–3.45) (–4.54) (–3.21) Slow sale above purchase (t-1) –0.403*** –0.686*** –0.615 (–4.16) (–2.64) (–1.49) Controls Yes Yes Yes Yes Yes Yes R-squared 0.1101 0.1101 0.1215 0.1215 0.0965 0.0965 F-statistic 1.32 1.05 12.02*** 9.52*** 6.91*** 5.48** Number of observations 1,352,175 1,352,175 1,297,278 1,297,278 1,229,697 1,229,697 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade occurred in the associated firm-month. The reference price Purchase equals the most recent purchase price. To indicate that the sale happened at least 180 days after the most recent purchase, I use the adjective “slow.” Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. I multiply all coefficient estimates by 100. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than one month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. 3.6 Correlation with the price path 3.6.1 Opportunistic and routine traders As the final alternative to the hypothesis that my basic result is driven by insider aversion to realizing losses, I examine whether this result stems from a correlation with the price path. For example, the timing of an insider trade at a loss may be correlated with the recent path of the stock price in some way that is not fully captured by my controls for short-term reversals, momentum, and capital gains overhang. To examine this explanation, I first consider a placebo test. I test whether there is differential predictive power for future returns between uninformed realized losses and uninformed realized gains. Cohen et al. (2012) use a filter to distinguish informative trades from uninformative trades. The authors label informative trades “opportunistic” and uninformative trades “routine”. They show that “opportunistic” sales contain all the information content in their universe of sales. Since routine sales do not have any informational content, I do not expect there to be a difference in predictive power for future returns between routine sales at a loss and routine sales at a gain. Of course, I still expect there to be a significant difference between the information content of opportunistic realized losses and opportunistic realized gains. For this test, I use the 6-month moving average as the reference price because it does not require a trading history and therefore does not interfere with Cohen et al.’s (2012) identification of routine and opportunistic traders. Also, I use the sample of all insiders instead of limiting the sample to “inactive” insiders as I do not want to interfere with Cohen et al.’s (2012) classification schema. I consider a sample of open-market purchases and sales.13 Following Cohen et al. (2012), I classify insiders at the beginning of each year as opportunistic or routine. An insider is classified as a routine trader if she made a trade (a purchase or sale) in the same month for three consecutive years (e.g. an insider who made a trade in April 2001, April 2002, and April 2003 would be considered a routine trader from January 2004 to present). All insiders who are not classified as routine are classified as opportunistic. I only consider trades by insiders who have three consecutive years of trading history. This makes it possible for an insider to be classified as a routine trader. To allow a 3-year trading history to be built, I restrict my sample to firm-months after 1989. By definition, opportunistic traders can be reclassified as routine traders at the beginning of each year, but routine traders stay routine. Again, I run pooled OLS regressions with month fixed effects and cluster standard errors by month. For firm $$i$$ in month $$t$$, I estimate the model: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}\\ &\quad +\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}\\ &\quad +\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Opportunistic}\;\textit{buy}_{i,t-1}+\beta_{6}\:\textit{Routine}\;\textit{buy}_{i,t-1}\\ &\quad +\beta_{7}\:\textit{Routine}\;\textit{realized}\;\textit{gain}_{i,t-1}+\beta_{8}\:\textit{Routine}\;\textit{realized}\;\textit{loss}{}_{i,t-1}\\ &\quad +\beta_{9}\:\textit{Previous}\;\textit{month-end}\;\textit{price}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{10}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma \textit{Month}+\epsilon_{i,t}.\label{eq:7} \end{split} \end{align} (7) My prediction is that $$\beta_{7}=\beta_{8}.$$ That is, I do not expect there to be any difference between the predictive power of future returns for realized losses relative to realized gains among routine trades. I present the results in Table 11. Like Cohen et al. (2012), opportunistic sales predict more negative returns than routine sales and the difference is significant at the 1% level $$(F-\mbox{statistic}=13.65)$$. Consistent with my result reflecting the information content of insider sales, I fail to reject the null hypothesis that the predictive power of routine realized gains equals the predictive power of routine realized losses. If anything, routine sales below the reference price predict more positive returns than routine sales above the reference price. On the other hand, the difference in predictive power between opportunistic realized gains, $$-$$15 basis points, and opportunistic realized losses, $$-$$74 bps, is statistically significant at the 1% level $$(F-\mbox{statistic}=10.10)$$. Table 11 Opportunistic and routine trades 1-month return 1-month return 1-month return Opportunistic buy (t-1) 0.459*** 0.392*** 0.404*** (3.24) (2.88) (2.97) Routine buy (t-1) –0.145 –0.226 –0.222 (–0.72) (–1.22) (–1.20) Opportunistic sale (t-1) –0.233** (–2.96) Routine sale (t-1) 0.244** (2.16) Routine sale below moving average (t-1) 0.292 (1.37) Routine sale above moving average (t-1) 0.012 (0.10) Opportunistic sale below moving average (t-1) –0.742*** (–4.47) Opportunistic sale above moving average (t-1) –0.151* (–1.79) Below moving average (t-1) –0.391*** –0.386*** (–3.65) (–3.60) Capital gains overhang 0.213 0.213 (0.70) (0.70) Standard stock characteristic controls Yes Yes Yes R-squared 0.0992 0.1025 0.1025 F-statistic 13.65*** 1.25 10.10*** Number of observations 1,444,601 1,230,813 1,230,813 1-month return 1-month return 1-month return Opportunistic buy (t-1) 0.459*** 0.392*** 0.404*** (3.24) (2.88) (2.97) Routine buy (t-1) –0.145 –0.226 –0.222 (–0.72) (–1.22) (–1.20) Opportunistic sale (t-1) –0.233** (–2.96) Routine sale (t-1) 0.244** (2.16) Routine sale below moving average (t-1) 0.292 (1.37) Routine sale above moving average (t-1) 0.012 (0.10) Opportunistic sale below moving average (t-1) –0.742*** (–4.47) Opportunistic sale above moving average (t-1) –0.151* (–1.79) Below moving average (t-1) –0.391*** –0.386*** (–3.65) (–3.60) Capital gains overhang 0.213 0.213 (0.70) (0.70) Standard stock characteristic controls Yes Yes Yes R-squared 0.0992 0.1025 0.1025 F-statistic 13.65*** 1.25 10.10*** Number of observations 1,444,601 1,230,813 1,230,813 This table presents regressions of 1-month returns on different types of insider sales at the firm level. I consider a sample of firm-months from 1989 to 2015. The dependent variable in these panel regressions is the future 1-month return $$(t,t+1)$$. I consider insiders who have traded for 3 consecutive years. Insiders are labeled routine if they made a trade in the same month for 3 consecutive years. All remaining insiders are classified as opportunistic. Dummies for opportunistic (routine) buys and sales take a value of one for a specific firm-month entry if a firm’s opportunistic (routine) insider bought or sold, respectively, in that month. Standard stock characteristic controls include the return from $$t-1$$ to $$t$$, the return from $$t-12$$ to $$t-1$$, the log value of common equity divided by market capitalization, and the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I include month fixed effects. Standard errors are clustered by month. Moving average is the reference price, and is equal to the average of the previous 6-month-end prices. Below moving average$$(t-1)$$ is a dummy that equals one if the previous month-end price is below the moving-average. The F-statistic in the first column tests whether the coefficient on opportunistic sale $$(t-1)$$ equals the coefficient on routine sale $$(t-1)$$. The F-statistic in the second column tests whether the coefficient on routine sale below moving average $$(t-1)$$ equals the coefficient on routine sale above moving average $$(t-1)$$. The F-statistic for the third column tests whether the coefficient on opportunistic sale below moving average $$(t-1)$$ equals the coefficient on opportunistic sale above moving average $$(t-1)$$. I multiply all coefficient estimates by 100. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 11 Opportunistic and routine trades 1-month return 1-month return 1-month return Opportunistic buy (t-1) 0.459*** 0.392*** 0.404*** (3.24) (2.88) (2.97) Routine buy (t-1) –0.145 –0.226 –0.222 (–0.72) (–1.22) (–1.20) Opportunistic sale (t-1) –0.233** (–2.96) Routine sale (t-1) 0.244** (2.16) Routine sale below moving average (t-1) 0.292 (1.37) Routine sale above moving average (t-1) 0.012 (0.10) Opportunistic sale below moving average (t-1) –0.742*** (–4.47) Opportunistic sale above moving average (t-1) –0.151* (–1.79) Below moving average (t-1) –0.391*** –0.386*** (–3.65) (–3.60) Capital gains overhang 0.213 0.213 (0.70) (0.70) Standard stock characteristic controls Yes Yes Yes R-squared 0.0992 0.1025 0.1025 F-statistic 13.65*** 1.25 10.10*** Number of observations 1,444,601 1,230,813 1,230,813 1-month return 1-month return 1-month return Opportunistic buy (t-1) 0.459*** 0.392*** 0.404*** (3.24) (2.88) (2.97) Routine buy (t-1) –0.145 –0.226 –0.222 (–0.72) (–1.22) (–1.20) Opportunistic sale (t-1) –0.233** (–2.96) Routine sale (t-1) 0.244** (2.16) Routine sale below moving average (t-1) 0.292 (1.37) Routine sale above moving average (t-1) 0.012 (0.10) Opportunistic sale below moving average (t-1) –0.742*** (–4.47) Opportunistic sale above moving average (t-1) –0.151* (–1.79) Below moving average (t-1) –0.391*** –0.386*** (–3.65) (–3.60) Capital gains overhang 0.213 0.213 (0.70) (0.70) Standard stock characteristic controls Yes Yes Yes R-squared 0.0992 0.1025 0.1025 F-statistic 13.65*** 1.25 10.10*** Number of observations 1,444,601 1,230,813 1,230,813 This table presents regressions of 1-month returns on different types of insider sales at the firm level. I consider a sample of firm-months from 1989 to 2015. The dependent variable in these panel regressions is the future 1-month return $$(t,t+1)$$. I consider insiders who have traded for 3 consecutive years. Insiders are labeled routine if they made a trade in the same month for 3 consecutive years. All remaining insiders are classified as opportunistic. Dummies for opportunistic (routine) buys and sales take a value of one for a specific firm-month entry if a firm’s opportunistic (routine) insider bought or sold, respectively, in that month. Standard stock characteristic controls include the return from $$t-1$$ to $$t$$, the return from $$t-12$$ to $$t-1$$, the log value of common equity divided by market capitalization, and the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I include month fixed effects. Standard errors are clustered by month. Moving average is the reference price, and is equal to the average of the previous 6-month-end prices. Below moving average$$(t-1)$$ is a dummy that equals one if the previous month-end price is below the moving-average. The F-statistic in the first column tests whether the coefficient on opportunistic sale $$(t-1)$$ equals the coefficient on routine sale $$(t-1)$$. The F-statistic in the second column tests whether the coefficient on routine sale below moving average $$(t-1)$$ equals the coefficient on routine sale above moving average $$(t-1)$$. The F-statistic for the third column tests whether the coefficient on opportunistic sale below moving average $$(t-1)$$ equals the coefficient on opportunistic sale above moving average $$(t-1)$$. I multiply all coefficient estimates by 100. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. 3.6.2 Difference in return predictability around the break-even point I perform another test to see whether I am capturing price path effects. Specifically, I look at the difference in return predictability in a narrow band around the reference point. Losses and gains are still fundamentally different, but there will not be a lot of variation, by construction, in the distance from the reference price. As such, any difference in return predictability, in this narrow band, can more plausibly be attributed to insider information than the price path. Specifically, I consider a regression of the following form: \begin{align*} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}+\\ &\quad \beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}+\\ &\quad \beta_{6}\:\textit{Sale}\;\textit{slightly}\;\textit{below}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+ \end{split} \end{align*} \begin{align} \begin{split} &\quad \beta_{7}\:\textit{Sale}\;\textit{slightly}\;\textit{above}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{8}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t}.\label{eq:narrow_band} \end{split} \end{align} (8) The band I consider runs from 10% below the reference price to 10% above the reference price. As such, Sale slightly below FIFO reference price is a sale below the reference price, but no less than 10% below the reference price. My prediction is that $$\beta_{6}<\beta_{7}.$$ I find evidence consistent with this hypothesis. I present the results in Table 12. At the 1-month horizon a sale slightly below the reference price predicts a 1-month return 67 bps lower, whereas a sale slightly above the reference price predicts a 1-month return only 19 basis points lower. The difference is statistically significant at the 1% level. At the 1-year horizon a sale slightly below the reference price predicts a return 258 bps lower, whereas a sale slightly above the reference price predicts a 1-month return only 28 basis points lower. Again, the difference is statistically significant at the 1% level. In Figure 1, I present the coefficient estimates from a regression of future 1-month returns on ten different buckets measuring distance from the FIFO reference price. (I use the same controls used in Equation (8).) The figure also offers evidence of a jump at the reference price. Table 12 Difference in a narrow band 1-month return 6-month return 1-year return Short-term reversal –2.38** –1.16 1.83 (–2.71) (–0.86) (0.78) Momentum 0.308 0.617 –0.836 (1.61) (1.51) (–1.42) Book-to-market 0.321*** 1.68*** 3.35*** (3.59) (5.14) (6.96) Size –0.0207 –0.154 –0.416** (–0.47) (–1.22) (–2.36) Buy (t-1) 0.639*** 1.83*** 2.82*** (7.75) (8.28) (6.61) Sale slightly below FIFO (t-1) –0.665*** –1.78*** –2.58*** (–4.69) (–4.65) (–4.07) Sale slightly above FIFO (t-1) –0.188* –0.176 –0.276 (–1.73) (–0.48) (–0.52) Capital gains overhang 0.327 0.799 –1.99 (1.14) (1.11) (–1.54) R-squared 0.1101 0.1214 0.0965 F-statistic 7.44*** 10.52*** 10.27*** Number of observations 1,352,175 1,297,278 1,229,697 1-month return 6-month return 1-year return Short-term reversal –2.38** –1.16 1.83 (–2.71) (–0.86) (0.78) Momentum 0.308 0.617 –0.836 (1.61) (1.51) (–1.42) Book-to-market 0.321*** 1.68*** 3.35*** (3.59) (5.14) (6.96) Size –0.0207 –0.154 –0.416** (–0.47) (–1.22) (–2.36) Buy (t-1) 0.639*** 1.83*** 2.82*** (7.75) (8.28) (6.61) Sale slightly below FIFO (t-1) –0.665*** –1.78*** –2.58*** (–4.69) (–4.65) (–4.07) Sale slightly above FIFO (t-1) –0.188* –0.176 –0.276 (–1.73) (–0.48) (–0.52) Capital gains overhang 0.327 0.799 –1.99 (1.14) (1.11) (–1.54) R-squared 0.1101 0.1214 0.0965 F-statistic 7.44*** 10.52*** 10.27*** Number of observations 1,352,175 1,297,278 1,229,697 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade(s) occurred in the associated firm-month. “Slightly” indicates that the sale occurred no greater than 10% from the reference price. The reference price FIFO is based on FIFO accounting. Short-term reversal equals the return from $$t-1$$ to $$t$$, and Momentum equals the return from $$t-12$$ to $$t-1$$. Book-to-market is the log value of common equity divided by market capitalization. Size is the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I drop all observations in which the month-end price is less than one. I multiply all coefficient estimates by 100. The F-statistic tests whether the coefficient on sale slightly below FIFO $$(t-1)$$ equals the coefficient on sale slightly above FIFO $$(t-1)$$. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than one month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 12 Difference in a narrow band 1-month return 6-month return 1-year return Short-term reversal –2.38** –1.16 1.83 (–2.71) (–0.86) (0.78) Momentum 0.308 0.617 –0.836 (1.61) (1.51) (–1.42) Book-to-market 0.321*** 1.68*** 3.35*** (3.59) (5.14) (6.96) Size –0.0207 –0.154 –0.416** (–0.47) (–1.22) (–2.36) Buy (t-1) 0.639*** 1.83*** 2.82*** (7.75) (8.28) (6.61) Sale slightly below FIFO (t-1) –0.665*** –1.78*** –2.58*** (–4.69) (–4.65) (–4.07) Sale slightly above FIFO (t-1) –0.188* –0.176 –0.276 (–1.73) (–0.48) (–0.52) Capital gains overhang 0.327 0.799 –1.99 (1.14) (1.11) (–1.54) R-squared 0.1101 0.1214 0.0965 F-statistic 7.44*** 10.52*** 10.27*** Number of observations 1,352,175 1,297,278 1,229,697 1-month return 6-month return 1-year return Short-term reversal –2.38** –1.16 1.83 (–2.71) (–0.86) (0.78) Momentum 0.308 0.617 –0.836 (1.61) (1.51) (–1.42) Book-to-market 0.321*** 1.68*** 3.35*** (3.59) (5.14) (6.96) Size –0.0207 –0.154 –0.416** (–0.47) (–1.22) (–2.36) Buy (t-1) 0.639*** 1.83*** 2.82*** (7.75) (8.28) (6.61) Sale slightly below FIFO (t-1) –0.665*** –1.78*** –2.58*** (–4.69) (–4.65) (–4.07) Sale slightly above FIFO (t-1) –0.188* –0.176 –0.276 (–1.73) (–0.48) (–0.52) Capital gains overhang 0.327 0.799 –1.99 (1.14) (1.11) (–1.54) R-squared 0.1101 0.1214 0.0965 F-statistic 7.44*** 10.52*** 10.27*** Number of observations 1,352,175 1,297,278 1,229,697 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade(s) occurred in the associated firm-month. “Slightly” indicates that the sale occurred no greater than 10% from the reference price. The reference price FIFO is based on FIFO accounting. Short-term reversal equals the return from $$t-1$$ to $$t$$, and Momentum equals the return from $$t-12$$ to $$t-1$$. Book-to-market is the log value of common equity divided by market capitalization. Size is the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I drop all observations in which the month-end price is less than one. I multiply all coefficient estimates by 100. The F-statistic tests whether the coefficient on sale slightly below FIFO $$(t-1)$$ equals the coefficient on sale slightly above FIFO $$(t-1)$$. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than one month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. 3.7 Rebalancing My maintained hypothesis is that the result in Section 4.2 is driven by the aversion to realizing losses inherent in the disposition effect. In the next three sections, I explore possible sources of this aversion to realizing losses, an aversion that makes sales at a loss much more informative than sales at a gain. I first examine a portfolio rebalancing explanation. Rebalancing motives are an important motive when an investor decides to sell a stock. Unlike taxes, rebalancing motives could differentially encourage investors to sell more at a gain compared to at a loss.14 After a fall in the share price, the stock likely makes up a smaller fraction of the investor’s portfolio than before. After a rise in the share price, the stock likely makes up a larger fraction of the investor’s portfolio than before. Therefore, rebalancing motives may encourage the investor to sell more after a rise in the share price compared to after a fall in the share price. If so, my result may stem from the fact that realized gains are more likely to be rebalancing trades than realized losses.15 To test this hypothesis, I compare the predictive power for future returns of complete liquidations at a loss with the predictive power for future returns of complete liquidations at a gain. Complete liquidations are likely not driven by rebalancing motives; as a result, if complete liquidations at a loss predict returns more negatively than complete liquidations at a gain, this would cast doubt on the rebalancing view. Specifically, I estimate the following equation for firm $$i$$ in month $$t$$, clustering standard errors by month: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}\\ &\quad+\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}\\ &\quad +\beta_{6}\:\textit{Complete}\;\textit{liquidation}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{7}\:\textit{Complete}\;\textit{liquidation}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{8}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma\:\textit{Month}_{t}+\epsilon_{i,t}.\label{eq:8} \end{split} \end{align} (9) If rebalancing motives are driving my result, the prediction is that $$\beta_{6}=\beta_{7}.$$ I present the results in Table 4. I find that a complete liquidation at a loss predicts a 1-month return 483 bps lower than all other firm-months in my sample, whereas a complete liquidation at a gain predicts a one month return 121 bps lower. The F-statistic testing the difference between the two is 9.81, which is significant at the 1% level. I interpret this as evidence against a rebalancing explanation of my result. 3.8 An examination of realization utility Realization utility refers to the simple idea that investors derive utility directly from the act of realizing a gain or loss on an asset. Unlike the other explanations I have considered, this simple idea is consistent with the evidence presented thus far.16 Still, the underlying source of realization utility is unclear: why do investors experience disutility when they close out a position at a loss? In this section, I examine two proposed answers to this question – a heuristic-based explanation and a cognitive dissonance-based explanation. An investor who sells all of her positions at a gain (loss) makes (loses) money. This observation is the foundation for the heuristic that Barberis and Xiong (2012) argue underlies realization utility. Selling at a gain is good; selling at a loss is bad. The investor derives utility from doing something good (realizing gains) and derives disutility from doing something bad (realizing losses). Another possible source of realization disutility is a reluctance to admit that an earlier purchase decision was a mistake. Instead of admitting a mistake, investors may manipulate malleable, but negative, signals to preserve a positive self-image. This self-delusion relieves cognitive dissonance. An investor experiences cognitive dissonance, or “the discomfort that arises when a person recognizes that he or she makes choices and/or holds beliefs that are dissonant with each other,” if she thinks she is a talented investor, purchases stock, and then receives a signal indicating that the purchase was a mistake. Chang et al. (2016) highlight that investors can reduce the discomfort by either admitting their mistake, finding a third-party scapegoat, blaming bad luck, or explaining the bad performance as a temporary setback that will soon be reversed. As many people enjoy holding a positive self-image, admitting a mistake can be painful and a source of realization disutility. Company insiders are somewhat unique in that they acquire stock in two ways, either by actively purchasing shares or by being endowed with shares.17 This distinction allows me to shed light on the underlying source of realization utility. Under the heuristic explanation, a company insider will experience disutility if she sells shares at a loss regardless of whether she purchased the shares or was simply endowed with them. This predicts that a sale at a loss by an insider will be a more negative signal of future returns than a sale at a gain, regardless of whether the shares were purchased or endowed. On the other hand, under the cognitive dissonance view, the company insider will only feel pain from selling at a loss if the shares that she sells are shares that she actively purchased; if she was merely endowed with the shares, she has little reason to blame herself for their poor performance. This predicts that a sale of shares by an insider at a loss will be a more negative signal about the stock’s future return than a sale at a gain only if the shares that the insider sold were actively purchased shares. I test these competing theories by examining an individual-level panel data set. I include monthly observations for each insider from the month of her first transaction at the firm to the month of her last transaction in my data set. I examine the difference in return predictability between realized gains and realized losses for purchasers, or individuals who have previously purchased shares, and compare that to the difference in return predictability between realized gains and realized losses for nonpurchasers. To do this, I construct an interaction term between the realized loss dummy and the purchaser dummy, and an interaction term between the realized gain dummy and the purchaser dummy. If the “cognitive dissonance” explanation holds, the difference between realized losses and realized gains should be larger for purchasers than for nonpurchasers. Frydman and Rangel (2014) find that one can debias the disposition effect by reducing the saliency of the purchase price. It is plausible that purchasers will have a more salient reference price than those that have received shares. While reference price salience effects could still influence the results, I mitigate these effects by using a reference price that does not depend on purchase history, namely I use the 6-month moving-average as the reference price. To look at returns within insider, I include insider-firm level fixed effects. I use month fixed effects as I want to capture the firm-specific component of return predictability. Specifically, I estimate the following equation for individual-firm $$i$$ in month $$t$$, and cluster standard errors by month: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}\\ &\quad+\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}\\ &\quad +\beta_{5}\:\textit{Buy}_{i,t-1}+\beta_{6}\:\textit{Purchaser}_{i,t-1}\\ & \quad+\beta_{7}\:\textit{Sale}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{8}\:\textit{Sale}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{9}\:\textit{Sale}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}*\textit{Purchaser}_{i,t-1}\\ &\quad +\beta_{10}\:\textit{Sale}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}*\textit{Purchaser}_{i,t-1}\\ &\quad +\beta_{11\:}\textit{Previous}\;\textit{month-end}\;\textit{price}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma_{1}\:\textit{Individual-firm} +\gamma_{2}\:\textit{Month}+\epsilon_{i,t.}\label{eq:9} \end{split} \end{align} (10) The cognitive dissonance view predicts that $$\beta_{7}+\beta_{9}<\beta_{8}+\beta_{10}.$$ I find evidence that sales at a loss by a purchaser are a more negative signal for future returns than sales at a gain by a purchaser. Additionally, I find little evidence that sales at a loss by nonpurchasers are a more negative signal for future returns than sales at a gain for nonpurchasers. I present the results in the third and fourth columns of Table 7. At the 1-month horizon, a sale at a loss by a purchaser predicts a 1-month return about 32 bps lower than a sale at a gain by a purchaser. This difference is statistically significant at the 1% level. At the 1-year horizon, a sale at a loss by a purchaser predicts a 1-year return about 120 bps lower than a sale at a gain by a purchaser. This difference is not statistically significant, but if I consider the entire sample, where there is more identification, I see similar, but statistically significant, results. I present these results in the Online Appendix. Importantly, a sale at a gain by a nonpurchaser has, if anything, more negative predictive power than a sale at a loss by a nonpurchaser. I interpret these findings as evidence for a “cognitive dissonance” view of realization utility. 4. Conclusion I provide evidence that insiders require a strong negative signal to realize a loss. Specifically, I document that a sale of stock at a loss by a company insider is a much more negative signal for future stock returns than a sale of stock by a company insider at a gain. This simple filtering technique allows us to extract information from the trades of informed investors in a more efficient way. I consider a range of explanations for my results, including investor heterogeneity, the short swing rule, and rebalancing motives, but find that the evidence is most consistent with the idea that investors derive direct disutility from selling a stock at a loss. Since selling at a loss is painful, an investor who does so must have particularly negative information, information that manifests itself in a poor stock return over the next few months. By comparing the predictive power of sales of endowed shares at a loss with the predictive power of purchased shares at a loss, I am able to shed light on the source of this disutility. I find evidence that selling at a loss is painful because it forces the investor to admit that an earlier purchase decision was a mistake. I would like to thank my adviser Nicholas Barberis for his support and guidance. In addition, I am grateful for helpful comments from Ian Ayres, Robert Battalio, Tom Chang, Aytekin Ertan, Cary Frydman, Robin Greenwood (editor), Sam Hartzmark, Lawrence Jin, Andrew Karolyi, Stephen Karolyi, Andrew Metrick, and Justin Murfin; two anonymous referees; and the seminar participants at Cornell University, Cubist Systematic Strategies, Emory University, the London School of Economics, the Ohio State University, Purdue University, UC Irvine, the University of Florida, the University of Notre Dame, the University of Utah, and Yale University. I also acknowledge financial support from a Whitebox Advisors grant. Supplementary data can be found on The Review of Financial Studies web site. Footnotes 1 Odean (1998) finds this behavior in a data set of 10,000 trading accounts from a large discount brokerage. Genesove and Mayer (2001) document an aversion to realizing losses in the downtown Boston housing market. Grinblatt and Keloharju (2001) obtain data on the trading of people and institutions in the Finnish stock market and find that investors are reluctant to sell at a loss. Frazzini (2006) documents this behavior in mutual fund managers. Hartzmark and Solomon (2012) look at a set of NFL betting contracts at Tradesports.com and uncover evidence consistent with the disposition effect. 2 Seyhun (1998) concludes that several different trading rules based on the trades of company insiders lead to profits. Jeng et al. (2003) highlight that insider sales do not earn abnormal returns. Lakonishok and Lee (2001) argue that insider selling appears to have no predictive ability. Gao et al. (2016) argue that the absence of insider selling predicts negative future returns. 3 Shefrin and Statman (1985) suggest that investors open (close) a mental account when purchasing (selling) a stock and then evaluate the transaction at the moment of sale. As such, the realization of gains/losses becomes a determinant of overall utility. Similarly, Thaler (1999, p.189) writes “one clear intuition is that a realized loss is more painful than a paper loss. When a stock is sold, the gain or loss has to be ‘declared’ both to the tax authorities and to the investor (and spouse).” An implicit assumption of my analysis relevant in these examples is that insiders view investments in isolation: they exhibit narrow framing (e.g., see Tversky and Kahneman 1981). Barberis and Xiong (2012) and Ingersoll and Jin (2013) incorporate these ideas into formal models to explain a number of puzzling facts. Frydman et al. (2014) find neuroscientific evidence largely consistent with realization utility. 4 I also include private transactions, because, after May 1991, private transactions use the same codes as open-market transactions. I do not consider sales associated with the exercise of an option. 5 This is Fidelity’s default method for nonmutual fund securities. I do not consider the effects of derivative transactions. Since insiders are not allowed to short sell their own company stock, I also ignore situations in which total holdings would go negative. 6 When the weekly volume is greater than the number of shares outstanding, the weekly turnover equals one. 7 Following Grinblatt and Han (2005), I lag the weekly closing price to mitigate microstructure effects. In all regressions, I control for the capital gains overhang by using the value that corresponds to the last week of the month. 8 I consider return horizons as long as 3 years. At the 3-year horizon, sales at a loss predict returns about 2.5 percentage points lower, but the difference is no longer statistically significant. 9 I do not scale the earnings surprise by price because Cheong and Thomas (2011) find that the earnings per share (EPS) forecast error and dispersion do not vary with scale. To minimize the effect of outliers, I winsorize all earnings surprise variables, dispersion, and coverage at the 1% level. 10 Instead of subtracting an average monthly return, or index return, from the dependent variable, I use the fixed effects estimator as this estimator is consistent (Gormley and Matsa 2014). I overcome the computational difficulties that arise when estimating models with high-dimensional fixed effects by using the Stata procedure “reg2hdfe” designed by Guimaraes and Portugal (2010). 11 In the Online Appendix, I estimate Equation (1) by insider type. I show that there is a strong difference between sales at a loss and sales at a gain for a number of insider types. 12 “For the purpose of preventing the unfair use of information which may have been obtained by such beneficial owner, director, or officer by reason of his relationship to the issuer, any profit realized by him from any purchase and sale, or any sale and purchase, of any equity security of such issuer (other than an exempted security) or a security-based swap agreement involving any such equity security within any period of less than six months, unless such security or security-based swap agreement was acquired in good faith in connection with a debt previously contracted, shall inure to and be recoverable by the issuer, irrespective of any intention on the part of such beneficial owner, director, or officer in entering into such transaction $${\ldots}$$” (15 U.S. Code § 78p). 13 I still do not include sales associated with option exercises. However, I find very similar results when I include sales associated with option exercises. 14 In the Online Appendix, I document weak results when tax considerations are strongest. Odean (1998) shows that tax-motivated selling is most evident in December; I find little difference in return predictability between sales at a loss in December and sales at a gain in December. 15 In contrast, Kallunki et al. (2009) look at data on Swedish insiders and show that insider selling is most informative for Swedish insiders who have the greatest proportion of wealth allocated to insider stocks. They explain this finding by noting that economic incentives are strongest for these insiders. 16 Realization utility might still predict a difference in return predictability in a narrow band—the results from Section 4.6.2—as insiders could still be suffering a large loss (gain) relative to their overall wealth by realizing a small percentage loss (gain). Data limitations prevent thorough analysis of this point. 17 Jin and Scherbina (2011) make a similar distinction for mutual fund holdings. The authors compare the holdings of managers who take over a fund with the holdings of continuing fund managers. References Ali, U., and Hirshleifer. D. A. 2017 . Opportunism as a firm and managerial trait: Predicting insider trading profits and misconduct. Journal of Financial Economics 126 : 490 – 515 . Google Scholar CrossRef Search ADS Barberis, N., and Xiong. W. 2012 . Realization utility. Journal of Financial Economics 104 : 251 – 71 . Google Scholar CrossRef Search ADS Bettis, C., Vickrey, D. and Vickrey. D. 1997 . Mimickers of corporate insiders who make large-volume trades. Financial Analysts Journal 53 : 57 – 66 . Google Scholar CrossRef Search ADS Chang, T., Solomon, D. and Westerfield. M. 2016 . Looking for someone to blame: Delegation, cognitive dissonance, and the disposition effect. Journal of Finance 71 : 267 – 302 . Google Scholar CrossRef Search ADS Cheng, S., Nagar, V. and Rajan. M. 2007 . Insider trades and private information: The special case of delayed-disclosure trades. Review of Financial Studies 20 : 1833 – 64 . Google Scholar CrossRef Search ADS Cheong, F. S., and Thomas. J. 2011 . Why do EPS forecast error and dispersion not vary with scale? Implications for analyst and managerial behavior. Journal of Accounting Research 49 : 359 – 401 . Google Scholar CrossRef Search ADS Cohen, L., Malloy, C. and Pomorski. L. 2012 . Decoding inside information. Journal of Finance 67 : 1009 – 43 . Google Scholar CrossRef Search ADS DellaVigna, S., and Pollet. J. 2009 . Investor inattention and Friday earnings announcements. Journal of Finance 64 : 709 – 49 . Google Scholar CrossRef Search ADS Frazzini, A. 2006 . The disposition effect and underreaction to news. Journal of Finance 61 : 2017 – 46 . Google Scholar CrossRef Search ADS Frydman, C., and Rangel. A. 2014 . Debiasing the disposition effect by reducing the saliency of information about a stock’s purchase price. Journal of Economic Behavior and Organization 107 : 541 – 52 . Google Scholar CrossRef Search ADS PubMed Frydman, C., Barberis, N. Camerer, C. Bossaerts, P. and Rangel. A. 2014 . Using neural data to test a theory of investor behavior: An application to realization utility. Journal of Finance 69 : 907 – 46 . Google Scholar CrossRef Search ADS PubMed Gao, G., Ma, Q. and Ng. D. 2016 . The sound of silence: What do we know when insiders do not trade? Working Paper . Genesove, D., and Mayer. C. 2001 . Loss aversion and seller behavior: Evidence from the housing market. Quarterly Journal of Economics 116 : 1233 – 60 . Google Scholar CrossRef Search ADS Gormley, T., and Matsa. D. 2014 . Common errors: How to (and not to) control for unobserved heterogeneity. Review of Finacial Studies 27 : 617 – 61 . Google Scholar CrossRef Search ADS Grinblatt, M., and Han. B. 2005 . Prospect theory, mental accounting, and momentum. Journal of Financial Economics 78 : 311 – 39 . Google Scholar CrossRef Search ADS Grinblatt, M., and Keloharju. M. 2001 . What makes investors trade? Journal of Finance 56 : 589 – 616 . Google Scholar CrossRef Search ADS Guimaraes, P., and Portugal. P. 2010 . A simple feasible procedure to fit models with high-dimensional fixed effects. Stata Journal 10 : 628 – 49 . Hartzmark, S., and Solomon. D. 2012 . Efficiency and the disposition effect in NFL prediction markets. Quarterly Journal of Finance 2 : 125 – 200 . Google Scholar CrossRef Search ADS Heath, C., Huddart, S. and Lang. M. 1999 . Psychological factors and stock option exercises. Quarterly Journal of Economics 114 : 601 – 27 . Google Scholar CrossRef Search ADS Heimer, R. 2016 . Peer pressure: Social interaction and the disposition effect. Review of Financial Studies 29 : 3177 – 209 . Google Scholar CrossRef Search ADS Huddart, S., and Lang. M. 2003 . Information distribution within firms: Evidence from stock option exercises. Journal of Accounting and Economics 34 : 3 – 31 . Google Scholar CrossRef Search ADS Ingersoll, J., and Jin. L. 2013 . Realization utility with reference-dependent preferences. Review of Financial Studies 26 : 723 – 67 . Google Scholar CrossRef Search ADS Jagolinzer, A. D. 2009 . SEC rule 10b5-1 and insiders’ strategic trade. Management Science 55 : 224 – 39 . Google Scholar CrossRef Search ADS Jeng, L., Metrick, A. and Zeckhauser. R. 2003 . Evidence of predictable behavior of security returns. Review of Economics and Statistics 85 : 453 – 71 . Google Scholar CrossRef Search ADS Jin, L., and Scherbina. A. 2011 . Inheriting losers. Review of Financial Studies 24 : 786 – 820 . Google Scholar CrossRef Search ADS Kallunki, J., Nillson, H. and Hellstrom. J. 2009 . Why do insiders trade? Evidence based on unique data on Swedish insiders. Journal of Accounting and Economics 48 : 37 – 53 . Google Scholar CrossRef Search ADS Lakonishok, J., and Lee. I. 2001 . Are insider trades informative? Review of Financial Studies 14 : 79 – 111 . Google Scholar CrossRef Search ADS Lin, J., and Howe. J. 1990 . Insider trading in the OTC market. Journal of Finance 45 : 1273 – 84 . Google Scholar CrossRef Search ADS Lorie, J., and Niederhoffer. V. 1968 . Predictive and statistical properties of insider trading. Journal of Law and Economics 11 : 35 – 53 . Google Scholar CrossRef Search ADS Marin, J., and Olivier. J. 2008 . The dog that did not bark: Insider trading and crashes. Journal of Finance 63 : 2429 – 76 . Google Scholar CrossRef Search ADS Massa, M., and Von Beschwitz. B. 2015 . Biased shorts: Stock market implications of short sellers’ disposition effect. Working Paper . Odean, T. 1998 . Are investors reluctant to realize their losses? Journal of Finance 53 : 1775 – 98 . Google Scholar CrossRef Search ADS Petersen, M. 2008 . Estimating standard errors in finance panel data sets: Comparing approaches. Review of Financial Studies 22 : 435 – 80 . Google Scholar CrossRef Search ADS Rozeff, M., and Zaman. M. 1988 . Market efficiency and insider trading: New evidence. Journal of Business 61 : 25 – 44 . Google Scholar CrossRef Search ADS Seyhun, H. N. 1998 . Investment intelligence from insider trading. Cambridge : MIT Press . Shefrin, H., and Statman. M. 1985 . The disposition to sell winners too early and ride losers too long: Theory and evidence. Journal of Finance 40 : 777 – 90 . Google Scholar CrossRef Search ADS Shumway, T. 1997 . The delisting bias in CRSP data. Journal of Finance 52 : 327 – 40 . Google Scholar CrossRef Search ADS Thaler, R. 1999 . Mental accounting matters. Journal of Behavioral Decision Making 12 : 183 – 206 . Google Scholar CrossRef Search ADS Tversky, A., and Kahneman. D. 1981 . The framing of decisions and the psychology of choice. Science 211 : 453 – 58 . Google Scholar CrossRef Search ADS PubMed Wang, W., Shin, Y. and Francis. B. 2012 . Are CFOs’ trades more informative than CEO’s trades? Journal of Financial and Quantitative Analysis 47 : 743 – 62 . Google Scholar CrossRef Search ADS Weber, M., and Camerer. C. 1998 . The disposition effect in securities trading: An experimental analysis. Journal of Economic Behavior and Organization 33 : 167 – 84 . Google Scholar CrossRef Search ADS © The Author(s) 2018. Published by Oxford University Press on behalf of The Society for Financial Studies. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png The Review of Financial Studies Oxford University Press

# The Information Content of Realized Losses

, Volume Advance Article (7) – Feb 7, 2018
31 pages

/lp/ou_press/the-information-content-of-realized-losses-MODGhcI4gm
Publisher
Oxford University Press
ISSN
0893-9454
eISSN
1465-7368
D.O.I.
10.1093/rfs/hhy013
Publisher site
See Article on Publisher Site

### Abstract

Abstract Examining the trades of company insiders, I find that a sale of stock at a loss is a much more negative signal about future returns than is a sale of stock at a gain. I consider a range of explanations for my results and find that the evidence is most consistent with the idea that investors derive direct disutility from selling a stock at a loss. Since selling a stock at a loss is painful, an investor who sells at a loss must have particularly negative information. This result offers a novel measurement of the strength of the disposition effect. Received December 7, 2015; editorial decision December 18, 2017 by Editor Robin Greenwood. Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online. The disposition effect is the robust empirical fact that many groups of investors—even sophisticated ones—have a greater propensity to sell assets at a gain, relative to the reference price, rather than at a loss.1 One way of summarizing the disposition effect is to say that investors are averse to realizing losses. This aversion has several possible sources, both economic and psychological. For example, it may result from rebalancing motives: a desire to rebalance their positions means that investors will have a greater propensity to sell assets at a gain rather than at a loss. Alternatively, it may be due to “realization utility,” in other words, to a direct preference for realizing gains and for not realizing losses. In this paper, I investigate how the disposition effect interacts with informative signals. If an informed investor does not like realizing losses, then observing her realize a loss suggests that she must have especially negative information about the future returns of the stock. Indeed, the information must be so negative that it leads her to override her inclinations to not realize losses. As such, I predict that a realized loss by an insider—an informed one—is a more negative signal for future returns than is a realized gain by an insider.2 To test this prediction, I must define the terms “realized gain” and “realized loss.” Presumably, an investor views a sale of stock as a realized gain if the price when sold exceeds some reference price she had in mind. In my analysis, I consider three reference prices. I focus on the price at which the stock was purchased, based on FIFO, or first in, first out, accounting. I also consider a reference price based on a time-series average of recent month-end prices and a reference price based on the most recent purchase price. With this definition of realized gain/loss in hand, I test my prediction using data from the U.S. Securities and Exchange Commission (SEC) on the trades of company insiders. I find that when an insider sells company stock at a loss relative to the FIFO reference price in the lagged month, the subsequent 6-month return for the company is 188 basis points (bps) lower than all other firm-months. On the other hand, when an insider sells shares at a gain relative to the FIFO reference price in the lagged month, the subsequent 6-month return is 9 bps higher than all other firm months. This suggests that insiders require a strong negative signal to realize a loss. The predictive power of realized losses allows for the construction of portfolio strategies that, at least before transaction costs, generate excess returns. An equal-weighted portfolio that buys stocks that have been recently sold at a gain by insiders and sells stocks that have been recently sold at a loss by insiders (using the FIFO reference price as the reference price) earns a four-factor alpha of 67 bps per month. In short, stocks perform more poorly after a sale by an insider at a loss than after a sale by an insider at a gain. Having confirmed my basic prediction, I then try to understand, at a deeper level, what is driving the result. My maintained hypothesis is the one described earlier: given the aversion to realizing losses that is inherent in the disposition effect, a sale of stock by an insider at a loss should be particularly informative. The investor must have a particularly negative signal about future returns if she is willing to sell despite a typical aversion to doing so. I first offer evidence against some alternative explanations for my result. For example, it is plausible that insiders who have a stronger propensity to sell at a loss execute more informative trades. Then, we would observe more negative return predictability from realized losses than from realized gains, but this would not be driven by an individual aversion to realizing losses. Rather, it is driven by the fact that informed traders tend to realize more losses. By comparing realized losses and realized gains within insider, I show that this is an unlikely explanation if the insider’s propensity to sell at a loss is static. It is, however, possible that this aversion is not static. For example, an aversion to realizing losses could change with the insider’s position in the firm. Furthermore, access to information likely changes as an insider’s position in the firm changes. Then we might observe more negative return predictability from realized losses than from realized gains, but this would not be driven by an individual aversion to realizing losses. By comparing the propensity to realize gains with the propensity to realize losses for each insider position type and comparing the return predictability of realized gains with the return predictability of realized losses for each insider position type, I show that this is an unlikely explanation of my result. The evidence suggests that my result is driven by an aversion to realizing losses. I examine this suggestion more closely: what exactly is driving insiders’ aversion to realizing losses – an aversion that makes their sales at a loss particularly informative? The evidence I offer is consistent with an explanation rooted in realization utility. Under this view, investors derive utility directly from realizing gains and losses: they experience a positive burst of utility when they realize a gain, a burst whose size depends on the size of the gain realized, and a negative burst of utility when they realize a loss.3 A sale of stock at a loss is then particularly informative about future returns: the investor must have negative information that overrides the pain felt when selling at a loss. I am able to go one step further by using my data on company insiders to comment on the source of realization utility. One view, the “heuristic” view, is that insiders feel pain when they sell an asset at a loss because they have in mind a rule of thumb, namely that selling assets at a loss is a bad idea. Another view, the “cognitive dissonance” view, is that investors feel pain when they sell an asset at a loss because doing so forces them to admit that their earlier purchase decision was a mistake. The company insider data allows me to examine these alternative hypotheses. When insiders acquire a position in a stock, they can do so in one of two ways: they may actively purchase shares or they may be endowed with shares. Under the dissonance view, an investor will not find it painful to sell shares she has been endowed with: since there was no active purchase decision, she does not have to blame herself for a trade gone bad. As a result, a sale of endowed shares at a loss should be no more informative about future returns than a sale of endowed shares at a gain. Under the heuristic view, however, a sale of endowed shares at a loss will be more informative than a sale of endowed shares at a gain: the investor finds it painful to sell even endowed shares at a loss. I test this prediction in the data and find evidence in favor of the dissonance view. My basic finding – that a sale of company stock by an insider at a loss has more negative predictive power for future returns than a sale at a gain – is most consistent with realization utility, in other words, with the view that people experience pain when they close out a position at a loss. This, in turn, means that when they do close out a position at a loss, they must have particularly negative information, information which manifests itself in low stock returns over the next few months. 1. Related Literature A large empirical literature investigates whether the trading activity of corporate insiders predicts returns in the cross-section (e.g., Lorie and Niederhoffer 1968; Rozeff and Zaman 1988; Lin and Howe 1990; Bettis et al. 1997). Seyhun (1998) reviews the evidence and concludes that insider trades contain predictive power for future returns. He notes that the information content of insider trades is higher when insiders purchase. Similarly, Lakonishok and Lee (2001) argue that the informativeness of insider trades comes from purchases, not sales. Finally, Jeng et al. (2003) highlight that whereas insider purchases earn abnormal returns, insider sales do not. In general, the literature attributes the weak predictive power of sales to the fact that sales are often driven by liquidity and diversification concerns, not private information. There are some papers that relate insider sales to private information. Cheng et al. (2007) show that sales filed under Form 5, which allow for delayed disclosure, predict negative returns. Marin and Olivier (2008) show that insider sales peak months before a large drop in the stock price. Relatedly, Gao et al. (2016) find evidence that insider silence predicts negative returns. In contrast, Huddart and Lang (2003) document that returns are higher when option exercises are low. Ali and Hirshleifer (2017) show that the sales of insiders, who make profitable trades prior to quarterly earnings announcements, have information content. Jagolinzer (2009) looks at the trades of insiders who pre-specify sales using 10b5-1 plans and finds that sale plans are terminated before good performance. Cohen et al. (2012) develop a filter based on trading patterns that decodes whether a sale is likely to be informative (opportunistic) or not (routine). The authors show that opportunistic sales have predictive ability for future returns. I extend our understanding of the information signals that insider trades reveal. The conditioning variable I suggest is motivated by the large literature on the aversion to realizing losses. Numerous researchers have documented this investor tendency. For example, Odean (1998) documents an aversion to realizing losses in a data set of 10,000 trading accounts from a large discount brokerage. Genesove and Mayer (2001) find an aversion to realizing losses in the downtown Boston housing market. Grinblatt and Keloharju (2001) obtain data on the trading of people and institutions in the Finnish stock market and find that investors are reluctant to sell at a loss. Frazzini (2006) documents this behavior in mutual fund managers. Hartzmark and Solomon (2012) look at a set of NFL betting contracts at Tradesports.com and uncover evidence consistent with the disposition effect. Heimer (2016) shows that social interaction can exacerbate the disposition effect. Kallunki et al. (2009) document the disposition effect in the trades of Swedish insiders. Heath et al. (1999) offer evidence that corporate insiders consider reference price heuristics when timing their option exercises. In particular, since option exercises do not have an associated purchase price, the authors show that corporate insiders consider dynamic reference prices. Specifically, the authors show that employee exercise activity approximately doubles when the stock price exceeds the maximum price attained during the previous year. My research offers a new way of measuring the disposition effect. Namely, by looking at returns following a realized loss, I can proxy for how strong of a negative signal insiders require in order to sell at a loss. In this paper, I argue that, as a consequence of this aversion to realizing losses, insider sales at a loss will have more negative predictive power for returns than sales at a gain. After confirming this prediction, I examine the data to understand the source of the aversion to realizing losses. There are a number of theoretical explanations for this aversion. Odean (1998) notes that rebalancing is a possible explanation, but provides evidence against it. Another possible explanation is a belief in mean reversion. However, Weber and Camerer (1998) and Hartzmark and Solomon (2012) provide evidence against this view. Barberis and Xiong (2012) show that realization utility—the notion that investors directly derive utility from realizing gains and losses—and time discounting can explain the disposition effect and a number of other puzzling facts. Finally, using brokerage and experimental data, Chang et al. (2016) provide evidence for the view that investors are averse to realizing losses because it is painful to admit that a prior purchase decision was a mistake. I find evidence consistent with the theory that it is painful to admit that a prior purchase decision was a mistake. 2. Data and Reference Price Construction 2.1 Data set construction To conduct the analysis, I collect data from several sources. I obtain return information, monthly closing prices, split-adjustments, and daily low/high/closing prices from CRSP. Other firm-level information comes from the CRSP/Compustat Merged database. I consider ordinary common shares listed on the AMEX, NASDAQ, or NYSE. I do not consider REITs, closed-end funds, ETFs, or Americus Trust Components. Following Shumway (1997), I replace missing delisting returns with a return equal to $$-$$0.3 for performance-related delistings. I construct a book-to-market control equal to the log value of common equity divided by market capitalization, where market capitalization is equal to the quarterly closing price times the number of common shares outstanding. I only consider observations that have a pre-log book-to-market ratio greater than 0 and less than or equal to 100. Size is calculated as the log value of market capitalization. I construct a control for momentum by calculating the previous year’s return, excluding the most recent month (i.e., I use the return from $$(t-12)$$ to $$(t-1)$$). To construct a control for the capital gains overhang, I need a weekly return series. I use CRSP’s daily stock return data file to construct a weekly return series. To mitigate any microstructure effects, I exclude observations where the month-end share price is below one dollar from my return-predictability analysis. I download factor data from Kenneth French’s website. I obtain insider data from Thomson-Reuters. The SEC requires corporate insiders, or “a company’s officers and directors, and any beneficial owners of more than 10% of a class of the company’s equity securities $${\ldots}$$”, to file their trades. I exclude observations with a cleanse indicator equal to A or S, as these indicate a failed cleansing attempt. I consider insider trades with a transaction code equal to P, S, or A from 1986 to 2015. That is, I consider open market purchases, open market sales, and grants or award transactions (hereafter, endowed shares), respectively.4 From here on, when I refer to a transaction by an insider, I am referring to one of these three types of transactions. I aggregate trades by personid, firm, transaction date, and transaction code. I use a share-weighted average, split-adjusted transaction price to compute the daily transaction price. If the transaction/endowed price is unreported, lower than the daily low price, or higher than the daily high price, I use the split-adjusted closing price for the corresponding day. I present summary statistics for the Number of trades in Table 2. As it may be difficult for “active” traders to keep track of their reference price, I drop “active” insiders from the sample. Specifically, I drop any insider that has more than 10 sale days, more than 10 buy days, or more than 10 acquisition days. I present similar results from the full sample in the Online Appendix. Table 2 Number of transactions by insider Number of acquisition days Number of purchase days Number of sale days 5th percentile 1 1 1 10th percentile 1 1 1 25th percentile 1 1 1 50th percentile 2 2 2 75th percentile 5 4 6 90th percentile 10 8 13 95th percentile 15 15 22 99th percentile 41 45 58 Max 430 1,057 1,551 Number of acquisition days Number of purchase days Number of sale days 5th percentile 1 1 1 10th percentile 1 1 1 25th percentile 1 1 1 50th percentile 2 2 2 75th percentile 5 4 6 90th percentile 10 8 13 95th percentile 15 15 22 99th percentile 41 45 58 Max 430 1,057 1,551 This table looks at the number of trading days for each insider and firm combination. I present the percentiles for number of acquisitions, purchases, and sale days by insider conditional on having at least one acquisition, purchase, or sale, respectively. I drop insider-firm combinations from the sample that have more than 10 purchases, 10 sales, or 10 acquisitions upon entering the month. Table 2 Number of transactions by insider Number of acquisition days Number of purchase days Number of sale days 5th percentile 1 1 1 10th percentile 1 1 1 25th percentile 1 1 1 50th percentile 2 2 2 75th percentile 5 4 6 90th percentile 10 8 13 95th percentile 15 15 22 99th percentile 41 45 58 Max 430 1,057 1,551 Number of acquisition days Number of purchase days Number of sale days 5th percentile 1 1 1 10th percentile 1 1 1 25th percentile 1 1 1 50th percentile 2 2 2 75th percentile 5 4 6 90th percentile 10 8 13 95th percentile 15 15 22 99th percentile 41 45 58 Max 430 1,057 1,551 This table looks at the number of trading days for each insider and firm combination. I present the percentiles for number of acquisitions, purchases, and sale days by insider conditional on having at least one acquisition, purchase, or sale, respectively. I drop insider-firm combinations from the sample that have more than 10 purchases, 10 sales, or 10 acquisitions upon entering the month. I construct earnings controls using financial data from Compustat Fundamentals and IBES. Considering Compustat Fundamentals quarterly data, I first drop duplicates that are a byproduct of fiscal year-end changes and then I drop duplicates that are a byproduct of multiple issues by the firm. The data from IBES is constructed from IBES actuals and IBES’ summary of analyst forecasts. I collect the unadjusted median estimate, the unadjusted actual earnings, the standard deviation of earnings estimates and the number of earnings estimates. I drop observations where the stated quarter end is one day before, on, or after the earnings announcement. I remove duplicates by retaining the most recent median forecast. I consider firms that use U.S. dollars as the currency. To merge Compustat Fundamentals data and IBES data I use the iclink linking table. I drop duplicate observations after merging with the iclink table by only keeping observations with the highest matching scores. Following DellaVigna and Pollet (2009), I define the earnings announcement date as the earlier of the Compustat and IBES dates. I calculate the earnings surprise as equal to the actual estimate minus the median estimate times 100. I winsorize the earnings surprise, the lagged earnings surprise, the standard deviation of earnings estimates, and the number of earnings estimates at the 1% level. 2.2 Reference prices My prediction is that a sale of stock by an informed investor that represents a realized loss is a more negative signal for the stock’s future return than a sale of stock by an informed investor that represents a realized gain. To test this hypothesis, I need to define “realized gain” and “realized loss.” Presumably, an insider thinks of a sale as a realized gain if the sale price is higher than some reference price she has in mind, and as a realized loss if the price is below the reference price. Therefore, to define “realized gain” and “realized loss,” I need to specify what this reference price is. In this paper, I consider three plausible reference prices. I list them below, along with details of their construction, and then discuss them. Recall that an insider can acquire a position in her firm’s stock in one of two ways: she can actively purchase shares, or she can be endowed with them. When I use the term “purchase price,” I am referring to the weighted-average split-adjusted purchase price from the day of purchase; when I use the term “endowed price,” I am referring to the weighted-average split-adjusted price recorded on Form 4 for grants or awards on the day of endowment; and, finally, when I use the term “acquisition price,” I mean either a purchase price or an endowed price. The reference prices I use are described here. FIFO reference price. I focus on this reference price throughout the paper. I determine this reference price by calculating a weighted average of past acquisition prices. Each acquisition price is weighted by the number of shares acquired that day (from the associated method) that have not been sold. I determine which shares have been sold based on FIFO accounting. That is, I will consider the oldest possible shares acquired to be the ones sold in a given period. Most recent purchase price. Moving-average price. I take the average of the previous six split-adjusted month-end prices. I focus on the FIFO reference price because it is likely the reference price used for inventory considerations.5 This reference price depends on previous purchase and previous endowed prices. I also consider a reference price based on the most recent purchase price. This reference price allows for a more reasonable discussion when I consider the short swing rule. Additionally, the most recent purchase price is plausibly the most salient purchase price due to recency effects. Several papers have suggested that purchase prices are natural reference prices. For example, Shefrin and Statman (1985) suggest that, when an investor buys a stock, a mental account is opened, one that closes only when she eventually sells the stock, at which point she evaluates the transaction by comparing the sale price to the purchase price. Similarly, Barberis and Xiong (2012) suggest that people think of their investing history as a series of investing episodes, each characterized by the name of the asset, the purchase price, and the sale price (“I bought IBM for 80 dollars and sold it for 120 dollars”). The last reference price I consider is different: it is a time-series average of recent monthly stock prices. I use this reference price for a number of reasons. First, and most importantly, it strikes me as a plausible proxy for when an insider might consider herself to be trading at a gain or loss. Suppose that an insider bought stock of Company X 10 years ago and enjoyed a meteoric rise in Company X’s stock price over the first 9.5 years. However, over the past 6 months, Company X’s stock is down 50%. I suspect that this insider might consider the stock to be trading at a loss. Specifically, if there are recent higher prices, the individual may regret not selling at the recent higher prices, and may consider herself to be trading at a loss relative to those prices. As attentive people, insiders may have more than one reference price; “I’m doing worse than I was doing last month, but I’m doing better than I was two months ago.” The moving average reference price tries to average a few of these recent reference prices. Of course, this reference price is most relevant for those who have held the stock for a long time. In Table 3, I show that the average holding periods of insiders in my sample are quite long. Specifically, the average time from the most recent purchase to a sale is 1,165 days and the median period is 782 days. The average time from the most recent acquisition to sale is 592 days and the median period is 294 days. Second, the moving average reference price does not require a history of purchases or acquisitions – as such, it is easy to consider people who have acquired shares through derivatives/who have acquired shares in an unclear way. Relatedly, it does not interfere with Cohen et al.’s (2012) selection criteria for opportunistic/routine insiders, which I will make use of later in the paper. Finally, the moving average reference price is a firm-wide reference price. This allows me to easily proxy for the “ability” to sell at a loss on a firm wide basis. Table 3 Average holding periods Average holding period Median holding period From most recent purchase to sale 1,165 days 782 days From most recent acquisition to sale 592 days 294 days Average holding period Median holding period From most recent purchase to sale 1,165 days 782 days From most recent acquisition to sale 592 days 294 days This table presents insiders’ average and median holding periods in my sample. The most recent acquisition refers to either the most recent purchase or the most recent grant or award of stock. Table 3 Average holding periods Average holding period Median holding period From most recent purchase to sale 1,165 days 782 days From most recent acquisition to sale 592 days 294 days Average holding period Median holding period From most recent purchase to sale 1,165 days 782 days From most recent acquisition to sale 592 days 294 days This table presents insiders’ average and median holding periods in my sample. The most recent acquisition refers to either the most recent purchase or the most recent grant or award of stock. 3. Company Insiders In this section, I examine the trades of company insiders. I first test my prediction that a sale at a loss predicts more negative returns than a sale at a gain. Having confirmed my prediction, I examine alternatives to my hypothesis that my result is driven by an aversion to realizing losses. After finding evidence that the result is driven by an aversion to realizing losses, I examine potential drivers of this aversion. 3.1 Disposition effect Central to my prediction is the premise that insiders are averse to selling at a loss. If this is the case, the proportion of losses realized by insiders should be less than the proportion of gains realized by insiders. I consider a panel of insider-firm combinations. I include monthly observations for each insider from the month of her first transaction at the firm to the last month she appears in my data set. I consider it to be possible for an insider to sell at a loss if the previous month-end price is below the insider’s reference price. I calculate the proportion of losses realized, or PLR, by dividing the number of months when an insider sold at a loss by the number of months when the insider could have sold at a loss. In Table 1, I present the PLR and the PGR, or the proportion of gains realized, for the three reference prices that I focus on throughout my analysis - the reference price using FIFO accounting, the most recent purchase price, and the previous 6-month moving-average price. I find significant evidence of the disposition effect. For example, I find that, relative to the FIFO reference price, the PLR is 1.35% and the PGR is 2.28%. The difference is statistically significant at the 1% level. This result adds to a recent literature which shows that even “sophisticated” investors exhibit a disposition effect. Frazzini (2006) documents this behavior in mutual fund managers. Massa and Von Beschwitz (2015) show that short sellers exhibit the disposition effect. Kallunki et al. (2009) document the disposition effect in the trades of Swedish insiders. Table 1 Disposition effect Gain Loss Gain Loss No sale 4,494,639 3,642,119 No sale 4,442,653 2,065,232 Sale 210,164 101,302 Sale 103,529 38,539 PGR=0.0447 PLR=0.0271 PGR=0.0228 PLR=0.0135 Gain Loss Gain Loss No sale 4,494,639 3,642,119 No sale 4,442,653 2,065,232 Sale 210,164 101,302 Sale 103,529 38,539 PGR=0.0447 PLR=0.0271 PGR=0.0228 PLR=0.0135 Gain Loss No sale 2,909,502 1,577,226 Sale 68,671 25,049 PGR=0.0231 PLR=0.0156 Gain Loss No sale 2,909,502 1,577,226 Sale 68,671 25,049 PGR=0.0231 PLR=0.0156 The first row of the left table shows the disposition effect for the moving average. The first row of the right table shows the disposition effect for the FIFO reference price, and the second row of the table shows the disposition effect for the most recent purchase price. PGR, the proportion of gains realized; PLR, the proportion of losses realized. Table 1 Disposition effect Gain Loss Gain Loss No sale 4,494,639 3,642,119 No sale 4,442,653 2,065,232 Sale 210,164 101,302 Sale 103,529 38,539 PGR=0.0447 PLR=0.0271 PGR=0.0228 PLR=0.0135 Gain Loss Gain Loss No sale 4,494,639 3,642,119 No sale 4,442,653 2,065,232 Sale 210,164 101,302 Sale 103,529 38,539 PGR=0.0447 PLR=0.0271 PGR=0.0228 PLR=0.0135 Gain Loss No sale 2,909,502 1,577,226 Sale 68,671 25,049 PGR=0.0231 PLR=0.0156 Gain Loss No sale 2,909,502 1,577,226 Sale 68,671 25,049 PGR=0.0231 PLR=0.0156 The first row of the left table shows the disposition effect for the moving average. The first row of the right table shows the disposition effect for the FIFO reference price, and the second row of the table shows the disposition effect for the most recent purchase price. PGR, the proportion of gains realized; PLR, the proportion of losses realized. 3.2 Firm-level regressions The central prediction of this paper is that insiders’ realized losses have more information content than their realized gains. To test this, I examine the predictive power of realized losses relative to realized gains for the future one month return. I run pooled ordinary least squares (OLS) regressions with month fixed effects. Specifically, I estimate the following equation for firm $$i$$ in month $$t$$: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1}&= \beta_{0}+\beta_{1}\:\textit{Short-term}\ \textit{reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}\\ &\quad +\beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}\\ & \quad +\beta_{6}\:\textit{Sale}\ \textit{below}\ \textit{reference}\ \textit{price}_{i,t-1} \\ & \quad+\beta_{7}\:\textit{Sale}\ \textit{above}\ \textit{reference}\ \textit{price}_{i,t-1}\\ & \quad+\beta_{8}\:\textit{Capital}\ \textit{gains}\ \textit{overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t}.\label{eq:1} \end{split} \end{align} (1) I estimate the equation using a sample of all firm-month observations since 1986. I include all firm-months as Ali and Hirshleifer (2017) argue that limiting the sample to firms with insider trades may bias the results in favor of finding return predictability in sales. The dependent variable is the return on stock $$i$$ from the end of month $$t$$ to the end of month $$t+1$$. The most important independent variables are Sale below reference price – a dummy that takes a value of one if an insider in firm $$i$$ sold stock at a loss relative to the reference price in month $$t-1$$ – and the analogous Sale above reference price. I lag the variables of interest as I do not want to pick up a contemporaneous correlation between firm returns and insiders’ trade signals. My prediction is that $$\beta_{6}$$ is less than $$\beta_{7}$$. In other words, I predict that a sale at a loss has more negative predictive power for subsequent returns than a sale at a gain. Before turning to the results, I first explain the other independent variables in the regression. I control for the previous month return, the previous year return, excluding the most recent month, the book-to-market ratio, and the size of the firm to ensure that my results are not driven by the short-term reversal phenomenon, the momentum anomaly, the value anomaly, or the size anomaly, respectively. I also control for the capital gains overhang (CGO), which, for a given firm, is a measure of the average capital gain embedded in investors’ holdings of the firm’s stock. Grinblatt and Han (2005) and Frazzini (2006) argue that, as a consequence of the disposition effect, CGO will predict subsequent stock returns; they confirm this prediction in the data. Since CGO is likely correlated with my Sale below reference price and Sale above reference price variables, it is important to control for it in my tests. Following Grinblatt and Han (2005), I construct a capital gains overhang control using weekly stock return data. To calculate the value of this control, I first calculate an aggregate reference price equal to: $$R_{t-1}=\frac{1}{k}\sum_{n=1}^{156}\left(V_{t-1-n}\prod_{j=1}^{n-1}(1-V_{t-1-n+j})\right)P_{t-1-n},\label{eq:2}$$ (2) where $$V_t$$ is the week $$t$$ volume turnover, which is calculated as the weekly volume divided by the number of shares outstanding.6$$P_t$$ is the week $$t$$ closing price, or the last daily closing price in the week. The coefficient on $$P_t$$ equals the probability that a share was purchased in period $$t$$ and has not been sold since then. The scaling factor $$k$$ ensures the sum of coefficients on $$P_t$$ equal one. As such, this reference price attempts to proxy for the cost basis of the typical investor in the market. Following Grinblatt and Han (2005), I then define capital gains overhang for firm $$i$$ in week $$t$$ as:7 $$Capital\;Gains\;Overhang_{i,t}=\frac{P_{i,t-2}-R_{t-1}}{P_{t-2}}.\label{eq:3}$$ (3) Finally, I also include month fixed effects as there could be common shocks within a period (e.g., changes in investor discount rates or investor sentiment) as well as within-month correlation across firms in my dummy variables. Some months may have a higher number of realized losses and those months may have negative 1-month returns as a result of aggregate market shocks. For example, tax considerations likely motivate insiders to realize losses in December. Additionally, there is likely correlation across firms as to when bonuses are paid/stock compensation vests, which would influence selling behavior, and also likely correlation as to when firms’ insiders are trading at a loss. In general, I am more interested in the firm-specific information content of insider sales. Finally, to address correlation within an industry/investment category in a particular month, I cluster standard errors by month. Having defined the control variables, I turn to the results. I focus on the FIFO reference price, which is defined in Section 3.2, and Table 4 presents the results. I look at a number of different return horizons, and find robust evidence that sales at a loss have more negative predictive power than sales at a gain. At the one month horizon, I find that a sale below the reference price predicts a return 61 $$(t=-5.32)$$ bps lower than all other firm-months in my sample. On the other hand, a sale above the reference price predicts a 1-month return only 21 bps lower $$(t=-2.58)$$. I reject the null hypothesis that the two are equal at the 1% level $$(F-\mbox{statistic}=7.44)$$. Table 4 Predictive power of sales at a loss relative to sales at a gain 1-month return 6-month return 1-year return 1-month return 1-month return Short-term reversal –2.38*** –1.16 1.84 –2.38*** –2.38*** (–2.71) (–0.86) (0.78) (–2.71) (–2.70) Momentum 0.311 0.611 –0.853 0.311 0.309 (1.63) (1.49) (–1.45) (1.63) (1.62) Book-to-market 0.321*** 1.68*** 3.35*** 0.321*** 0.321*** (3.58) (5.15) (6.97) (3.58) (3.59) Size –0.018 –0.154 –0.422** –0.018 –0.022 (–0.42) (–1.23) (–2.39) (–0.42) (–0.51) Buy (t-1) 0.647*** 1.84*** 2.82*** 0.646*** 0.635*** (7.81) (8.31) (6.61) (7.81) (7.70) Sale below FIFO (t-1) –0.609*** –1.88*** –2.31*** –0.510*** (–5.32) (–5.25) (–3.96) (–4.31) Sale above FIFO (t-1) –0.206*** 0.0898 0.592 –0.204*** (–2.58) (0.42) (1.60) (–2.58) Capital gains overhang 0.326 0.790 –2.01 0.326 0.326 (1.14) (1.09) (–1.55) (1.14) (1.14) More than one person selling below FIFO (t-1) –0.760*** (–2.74) More than one person selling above FIFO (t-1) 0.002 (0.01) Complete liquidation below FIFO (t-1) –4.83*** (–4.29) Complete liquidation above FIFO (t-1) –1.21** (–2.16) R-squared 0.1101 0.1215 0.0965 0.1101 0.1101 F-statistic 7.44*** 20.98*** 16.82*** 11.22*** 9.81*** Number of observations 1,352,175 1,297,278 1,229,697 1,352,175 1,352,175 1-month return 6-month return 1-year return 1-month return 1-month return Short-term reversal –2.38*** –1.16 1.84 –2.38*** –2.38*** (–2.71) (–0.86) (0.78) (–2.71) (–2.70) Momentum 0.311 0.611 –0.853 0.311 0.309 (1.63) (1.49) (–1.45) (1.63) (1.62) Book-to-market 0.321*** 1.68*** 3.35*** 0.321*** 0.321*** (3.58) (5.15) (6.97) (3.58) (3.59) Size –0.018 –0.154 –0.422** –0.018 –0.022 (–0.42) (–1.23) (–2.39) (–0.42) (–0.51) Buy (t-1) 0.647*** 1.84*** 2.82*** 0.646*** 0.635*** (7.81) (8.31) (6.61) (7.81) (7.70) Sale below FIFO (t-1) –0.609*** –1.88*** –2.31*** –0.510*** (–5.32) (–5.25) (–3.96) (–4.31) Sale above FIFO (t-1) –0.206*** 0.0898 0.592 –0.204*** (–2.58) (0.42) (1.60) (–2.58) Capital gains overhang 0.326 0.790 –2.01 0.326 0.326 (1.14) (1.09) (–1.55) (1.14) (1.14) More than one person selling below FIFO (t-1) –0.760*** (–2.74) More than one person selling above FIFO (t-1) 0.002 (0.01) Complete liquidation below FIFO (t-1) –4.83*** (–4.29) Complete liquidation above FIFO (t-1) –1.21** (–2.16) R-squared 0.1101 0.1215 0.0965 0.1101 0.1101 F-statistic 7.44*** 20.98*** 16.82*** 11.22*** 9.81*** Number of observations 1,352,175 1,297,278 1,229,697 1,352,175 1,352,175 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade(s) occurred in the associated firm-month. The reference price FIFO is based on FIFO accounting. Short-term reversal equals the return from $$t-1$$ to $$t$$, and Momentum equals the return from $$t-12$$ to $$t-1$$. Book-to-market is the log value of common equity divided by market capitalization. Size is the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I drop all observations in which the month-end price is less than one. I multiply all coefficient estimates by 100. In the first three columns, the F-statistic tests whether the coefficient on sale below FIFO $$(t-1)$$ equals the coefficient on sale above FIFO $$(t-1)$$. In the fourth column, the F-statistic tests whether multiple sales at a loss predict the same return as multiple sales at a gain. In the fifth column, the F-statistic tests whether the coefficient on complete liquidation below FIFO $$(t-1)$$ equals the coefficient on complete liquidation above FIFO $$(t-1)$$. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than 1 month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 4 Predictive power of sales at a loss relative to sales at a gain 1-month return 6-month return 1-year return 1-month return 1-month return Short-term reversal –2.38*** –1.16 1.84 –2.38*** –2.38*** (–2.71) (–0.86) (0.78) (–2.71) (–2.70) Momentum 0.311 0.611 –0.853 0.311 0.309 (1.63) (1.49) (–1.45) (1.63) (1.62) Book-to-market 0.321*** 1.68*** 3.35*** 0.321*** 0.321*** (3.58) (5.15) (6.97) (3.58) (3.59) Size –0.018 –0.154 –0.422** –0.018 –0.022 (–0.42) (–1.23) (–2.39) (–0.42) (–0.51) Buy (t-1) 0.647*** 1.84*** 2.82*** 0.646*** 0.635*** (7.81) (8.31) (6.61) (7.81) (7.70) Sale below FIFO (t-1) –0.609*** –1.88*** –2.31*** –0.510*** (–5.32) (–5.25) (–3.96) (–4.31) Sale above FIFO (t-1) –0.206*** 0.0898 0.592 –0.204*** (–2.58) (0.42) (1.60) (–2.58) Capital gains overhang 0.326 0.790 –2.01 0.326 0.326 (1.14) (1.09) (–1.55) (1.14) (1.14) More than one person selling below FIFO (t-1) –0.760*** (–2.74) More than one person selling above FIFO (t-1) 0.002 (0.01) Complete liquidation below FIFO (t-1) –4.83*** (–4.29) Complete liquidation above FIFO (t-1) –1.21** (–2.16) R-squared 0.1101 0.1215 0.0965 0.1101 0.1101 F-statistic 7.44*** 20.98*** 16.82*** 11.22*** 9.81*** Number of observations 1,352,175 1,297,278 1,229,697 1,352,175 1,352,175 1-month return 6-month return 1-year return 1-month return 1-month return Short-term reversal –2.38*** –1.16 1.84 –2.38*** –2.38*** (–2.71) (–0.86) (0.78) (–2.71) (–2.70) Momentum 0.311 0.611 –0.853 0.311 0.309 (1.63) (1.49) (–1.45) (1.63) (1.62) Book-to-market 0.321*** 1.68*** 3.35*** 0.321*** 0.321*** (3.58) (5.15) (6.97) (3.58) (3.59) Size –0.018 –0.154 –0.422** –0.018 –0.022 (–0.42) (–1.23) (–2.39) (–0.42) (–0.51) Buy (t-1) 0.647*** 1.84*** 2.82*** 0.646*** 0.635*** (7.81) (8.31) (6.61) (7.81) (7.70) Sale below FIFO (t-1) –0.609*** –1.88*** –2.31*** –0.510*** (–5.32) (–5.25) (–3.96) (–4.31) Sale above FIFO (t-1) –0.206*** 0.0898 0.592 –0.204*** (–2.58) (0.42) (1.60) (–2.58) Capital gains overhang 0.326 0.790 –2.01 0.326 0.326 (1.14) (1.09) (–1.55) (1.14) (1.14) More than one person selling below FIFO (t-1) –0.760*** (–2.74) More than one person selling above FIFO (t-1) 0.002 (0.01) Complete liquidation below FIFO (t-1) –4.83*** (–4.29) Complete liquidation above FIFO (t-1) –1.21** (–2.16) R-squared 0.1101 0.1215 0.0965 0.1101 0.1101 F-statistic 7.44*** 20.98*** 16.82*** 11.22*** 9.81*** Number of observations 1,352,175 1,297,278 1,229,697 1,352,175 1,352,175 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade(s) occurred in the associated firm-month. The reference price FIFO is based on FIFO accounting. Short-term reversal equals the return from $$t-1$$ to $$t$$, and Momentum equals the return from $$t-12$$ to $$t-1$$. Book-to-market is the log value of common equity divided by market capitalization. Size is the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I drop all observations in which the month-end price is less than one. I multiply all coefficient estimates by 100. In the first three columns, the F-statistic tests whether the coefficient on sale below FIFO $$(t-1)$$ equals the coefficient on sale above FIFO $$(t-1)$$. In the fourth column, the F-statistic tests whether multiple sales at a loss predict the same return as multiple sales at a gain. In the fifth column, the F-statistic tests whether the coefficient on complete liquidation below FIFO $$(t-1)$$ equals the coefficient on complete liquidation above FIFO $$(t-1)$$. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than 1 month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. I then consider the same regression model as Equation (1), except that I consider returns at longer horizons. Since many observations are overlapping, and there is more autocorrelation in returns at longer time horizons, I cluster standard errors in two dimensions. Namely, I cluster standard errors by firm and by month (Petersen 2008). At longer horizons, the difference between the return predictability of realized losses and realized gains is even more striking. A sale below the reference price predicts a 6-month return 188 bps lower than all other firm-months, whereas a sale above the reference price predicts a return 9 bps higher. The difference is statistically significant at the 1% level $$(F-\mbox{statistic}=20.98)$$. The difference is even bigger at the 1-year horizon. A sale below the reference price predicts a 1-year return 231 bps lower than all other firm-months, whereas a sale above the reference price predicts a return 59 bps higher. The difference is also statistically significant at the 1% level $$(F-\mbox{statistic}=16.82)$$.8 I present coefficient estimates from various return horizons in Figure 2. Figure 1 View largeDownload slide Magnitude of the sale at a loss or gain This graph charts the return predictability of sales based on distance from the FIFO reference price. The x-axis is a measure of the 1-month return predictability of the associated sale. The y-axis indicates the distance from the reference price. Figure 1 View largeDownload slide Magnitude of the sale at a loss or gain This graph charts the return predictability of sales based on distance from the FIFO reference price. The x-axis is a measure of the 1-month return predictability of the associated sale. The y-axis indicates the distance from the reference price. Figure 2 View largeDownload slide Predictive power of sales based on relation to the FIFO reference price This graph plots the coefficients on the dummy variables Sale below purchase price and Sale above purchase price from estimating: \begin{align*} \begin{split} \textit{Return}_{i,t\rightarrow t+j} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{Reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}\\ &\quad +\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}+\beta_{6}\:\textit{Sale}\;\textit{below}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{7}\:\textit{Sale}\;\textit{above}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{8}\:\textit{Capital}\;\textit{Gains}\;\textit{Overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t.} \end{split} \end{align*} over different return horizons $$j$$. The x-axis displays the time horizon, in months, over which returns are predicted. The y-axis displays the expected difference in returns when the dummy variable equals one compared to when the dummy variable equals zero. Figure 2 View largeDownload slide Predictive power of sales based on relation to the FIFO reference price This graph plots the coefficients on the dummy variables Sale below purchase price and Sale above purchase price from estimating: \begin{align*} \begin{split} \textit{Return}_{i,t\rightarrow t+j} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{Reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}\\ &\quad +\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}+\beta_{6}\:\textit{Sale}\;\textit{below}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{7}\:\textit{Sale}\;\textit{above}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{8}\:\textit{Capital}\;\textit{Gains}\;\textit{Overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t.} \end{split} \end{align*} over different return horizons $$j$$. The x-axis displays the time horizon, in months, over which returns are predicted. The y-axis displays the expected difference in returns when the dummy variable equals one compared to when the dummy variable equals zero. Under my hypothesis, it is plausible that firm-months which have more than one insider sell at a loss will predict more negative returns than firm-months which have exactly one insider sell at a loss. To test this, I add dummy variables to Equation (1). Specifically, I add dummy variables that equal one if there was more than one insider who sold at a loss (gain) relative to the reference price in the associated firm-month. I estimate the following equation for firm $$i$$ in month $$t$$: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}\\ &\quad +\beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}\\ &\quad +\beta_{6}\:\textit{Sale}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ & \quad+\beta_{7}\:\textit{Sale}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{8}\:\textit{Multiple}\;\textit{insiders}\;\textit{selling}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{9}\:\textit{Multiple}\;\textit{insiders}\;\textit{selling}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{10}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t}.\label{eq:4} \end{split} \end{align} (4) My prediction is that firm-months which have more than one insider realizing a loss will have more negative predictive power than firm-months with only one insider realizing a loss. That is, I predict that $$\beta_8<0$$. The results are consistent with this prediction. I estimate $$\beta_{8}$$ equal to $$-$$0.76 $$(t=-2.74)$$. That is, a firm-month which has more than one insider sell below the reference price is associated with a 1-month return 76 bps lower than a firm-month that has exactly one insider sell below the reference price. I also find that if more than one insider sells above the reference price, this predicts a 1-month return of the same magnitude as if only one insider sold at a gain. The predictive power of realized losses allows for the construction of portfolio strategies that, at least before transaction costs, earn alpha. At the end of each month $$t$$, I construct an equal-weighted portfolio that goes long a firm if one of its insiders sold shares at a gain relative to the FIFO reference price in month $$(t-1)$$ and go short a firm if one of its insiders sold at a loss relative to the FIFO reference price in month $$(t-1)$$. As shown in Table 5, I find that an equal-weighted portfolio earns a four-factor alpha of 67 bps per month with a t-statistic equal to 3.85, significant at the 1% level. I list other equal-weighted portfolio alphas in Table 5. Value-weighted portfolios do not earn alpha. This suggests that the greater negative predictive power of sales at a loss relative to sales at a gain is concentrated within small firms. Table 5 Portfolio returns 1-month return 1-month return 1-month return Intercept 0.60*** 0.67*** 0.67*** (3.63) (4.05) (3.85) Excess market return –0.09** –0.09** (–2.35) (–2.29) SMB –0.07 –0.07 (–1.35) (–1.35) HML –0.06 –0.06 (–1.05) (–1.03) Momentum –0.00 (–0.06) R-squared 0.000 0.0240 0.0240 N 351 351 351 1-month return 1-month return 1-month return Intercept 0.60*** 0.67*** 0.67*** (3.63) (4.05) (3.85) Excess market return –0.09** –0.09** (–2.35) (–2.29) SMB –0.07 –0.07 (–1.35) (–1.35) HML –0.06 –0.06 (–1.05) (–1.03) Momentum –0.00 (–0.06) R-squared 0.000 0.0240 0.0240 N 351 351 351 This table presents regressions of portfolio returns on return factors from Kenneth French’s Web site. The portfolio returns are monthly—end of month t to the end of month $$(t+1)$$—returns earned by an equal-weighted portfolio that goes long firms that had an insider sell at a gain relative to the FIFO reference price in month $$(t-1)$$ and goes short firms that had an insider sell at a loss relative to the FIFO reference price in month $$(t-1)$$. I require at least 5 stocks in both the long and short sides of the portfolio. The sample period runs from 1986 to 2015. I drop all observations in which the month-end price is less than one dollar. I put t-statistics in parentheses and multiply all coefficients by 100. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 5 Portfolio returns 1-month return 1-month return 1-month return Intercept 0.60*** 0.67*** 0.67*** (3.63) (4.05) (3.85) Excess market return –0.09** –0.09** (–2.35) (–2.29) SMB –0.07 –0.07 (–1.35) (–1.35) HML –0.06 –0.06 (–1.05) (–1.03) Momentum –0.00 (–0.06) R-squared 0.000 0.0240 0.0240 N 351 351 351 1-month return 1-month return 1-month return Intercept 0.60*** 0.67*** 0.67*** (3.63) (4.05) (3.85) Excess market return –0.09** –0.09** (–2.35) (–2.29) SMB –0.07 –0.07 (–1.35) (–1.35) HML –0.06 –0.06 (–1.05) (–1.03) Momentum –0.00 (–0.06) R-squared 0.000 0.0240 0.0240 N 351 351 351 This table presents regressions of portfolio returns on return factors from Kenneth French’s Web site. The portfolio returns are monthly—end of month t to the end of month $$(t+1)$$—returns earned by an equal-weighted portfolio that goes long firms that had an insider sell at a gain relative to the FIFO reference price in month $$(t-1)$$ and goes short firms that had an insider sell at a loss relative to the FIFO reference price in month $$(t-1)$$. I require at least 5 stocks in both the long and short sides of the portfolio. The sample period runs from 1986 to 2015. I drop all observations in which the month-end price is less than one dollar. I put t-statistics in parentheses and multiply all coefficients by 100. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. 3.3 Earnings Having established that sales at a loss predict more negative returns than sales at a gain, a natural follow-up question is to determine why sales at a loss have more negative predictive power than sales at a gain. I consider a sample of all firm-months that are either one month or two months before the month of an earnings announcement and follow the previous earnings announcement. I examine whether sales at a loss predict more negative earnings surprises than sales at a gain. Specifically, I run regressions of the earnings surprise on dummy variables that indicate whether a firm-month had the associated sale at a loss or sale at a gain. I estimate regressions of the following form for firm $$i$$ in fiscal year-quarter $$t$$ during month $$t'$$: \begin{align} \begin{split} \textit{Earnings}\;\textit{Surprise}_{i,t} & = \beta_{0}+\beta_{1}\:\textit{Sale}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t,t'}\\ &\quad +\beta_{2}\:\textit{Sale}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t,t'}\\ &\quad +\gamma_{1}\:\textit{Controls}+\gamma_{2}\:\textit{Fiscal}\;\textit{year-quarter}+\epsilon_{i,t}.\label{eq:5} \end{split} \end{align} (5) The dependent variable is equal to actual earnings per share minus the median earnings per share estimate. Sale below reference price – a dummy which takes a value of one if an insider in firm $$i$$ sold stock at a loss relative to the reference price in the current month – and the analogous Sale above reference price are my independent variables. My prediction is that $$\beta_{1}$$ is less than $$\beta_{2}$$. In other words, I predict that a sale at a loss has more negative predictive power for the subsequent earnings announcement than a sale at a gain. I conduct my analysis with common earnings surprise controls – a control for each of the past four earnings surprises, the dispersion of earnings estimates, the number of earnings estimates, and a dummy that equals one if there was only one earnings estimate – and a dummy that equals one if there was an insider purchase.9 The evidence confirms my prediction that realized losses predict more negative earnings surprises than realized gains; I present the results in Table 6. A sale of stock at a loss relative to the FIFO reference price predicts a future earnings surprise about one cent lower than a sale at a gain. Clustering standard errors by fiscal year-quarter and firm, I find that this result is statistically significant at the 1% level $$(F-\mbox{statistic}=52.56)$$. I find similar, though weaker, results when I include common controls for stock characteristics – the return in the month before the earnings announcement, the return from month $$t-12$$ to month $$t-1$$ before the earnings announcement, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang. Namely, a sale of stock at a loss relative to the FIFO reference price predicts a future earnings surprise about 0.3 cents lower than a sale at a gain. For comparison purposes, this difference is slightly larger than the magnitude of the positive predictive power of an insider purchase for the upcoming earnings announcement. I find that this difference is statistically significant at the 5% level $$(F-\mbox{statistic}=4.03)$$. Finally, I add firm fixed effects. A sale of stock at a loss predicts a future earnings surprise about 0.4 cents lower than a sale at a gain. Again, this result is significant at the 5% level $$(F-\mbox{statistic}=5.32)$$. Table 6 Earnings surprises Earnings surprise Earnings surprise Earnings surprise Earnings surprise Earnings surprise Sale below FIFO reference price –0.801*** –0.963*** –0.405*** –0.168 –0.125 (–4.42) (–5.65) (–2.89) (–1.15) (–0.87) Sale above FIFO reference price 2.107*** 1.555*** 0.774*** 0.160** 0.268*** (16.74) (14.34) (9.15) (1.98) (3.11) Buy –0.055 0.168** 0.284*** 0.423*** (–0.58) (2.03) (3.33) (5.32) Single estimate –3.183*** –2.173*** –1.541 –1.83*** (–15.24) (–12.33) (–8.60) (–8.80) Dispersion –39.15*** –27.28*** –27.04*** –40.24*** (–10.72) (–9.52) (–9.13) (–12.13) Number of estimates 0.172*** 0.104*** 0.0135 0.022 (14.53) (13.37) (1.30) (1.31) Earnings surprise (t-1) 0.213*** 0.198*** 0.145*** (22.54) (21.05) (15.37) Earnings surprise (t-2) 0.102*** 0.0950*** 0.0572*** (15.48) (13.69) (8.00) Earnings surprise (t-3) 0.0649*** 0.0601*** 0.0290*** (12.41) (11.00) (5.08) Earnings surprise (t-4) 0.0764*** 0.0704*** 0.0389*** (12.80) (11.45) (6.33) Stock characteristics No No No Yes Yes Fiscal year-quarter FEs Yes Yes Yes Yes Yes Firm FEs No No No No Yes F-statistic 186.77*** 171.07*** 52.56*** 4.03** 5.32** R-squared 0.0182 0.0384 0.1089 0.1181 0.1720 Number of observations 652,492 652,492 524,390 450,613 450,521 Earnings surprise Earnings surprise Earnings surprise Earnings surprise Earnings surprise Sale below FIFO reference price –0.801*** –0.963*** –0.405*** –0.168 –0.125 (–4.42) (–5.65) (–2.89) (–1.15) (–0.87) Sale above FIFO reference price 2.107*** 1.555*** 0.774*** 0.160** 0.268*** (16.74) (14.34) (9.15) (1.98) (3.11) Buy –0.055 0.168** 0.284*** 0.423*** (–0.58) (2.03) (3.33) (5.32) Single estimate –3.183*** –2.173*** –1.541 –1.83*** (–15.24) (–12.33) (–8.60) (–8.80) Dispersion –39.15*** –27.28*** –27.04*** –40.24*** (–10.72) (–9.52) (–9.13) (–12.13) Number of estimates 0.172*** 0.104*** 0.0135 0.022 (14.53) (13.37) (1.30) (1.31) Earnings surprise (t-1) 0.213*** 0.198*** 0.145*** (22.54) (21.05) (15.37) Earnings surprise (t-2) 0.102*** 0.0950*** 0.0572*** (15.48) (13.69) (8.00) Earnings surprise (t-3) 0.0649*** 0.0601*** 0.0290*** (12.41) (11.00) (5.08) Earnings surprise (t-4) 0.0764*** 0.0704*** 0.0389*** (12.80) (11.45) (6.33) Stock characteristics No No No Yes Yes Fiscal year-quarter FEs Yes Yes Yes Yes Yes Firm FEs No No No No Yes F-statistic 186.77*** 171.07*** 52.56*** 4.03** 5.32** R-squared 0.0182 0.0384 0.1089 0.1181 0.1720 Number of observations 652,492 652,492 524,390 450,613 450,521 This table shows the regressions of earnings surprises on insider sales and control variables. The dependent variable is the quarterly earnings surprise measured as the number of cents that the actual earnings per share differs from the median earnings per share estimate. All buy and sell variables are dummies that equal one if the stated trade occurred in the associated month. I consider all firm-months that are in the 2 months before an earnings announcement and follow the previous earnings announcement. Single estimate is a dummy that equals one if there was only one forecast for the associated firm-quarter. Dispersion equals the standard deviation of earnings forecasts. Number of estimates equals the number of analyst forecasts. I winsorize all earnings surprise variables, dispersion, and coverage at the 1% level. Stock characteristics include the return in the month before the earnings announcement, the return from month $$t-12$$ to month $$t-1$$ before the earnings announcement, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. The F-statistic tests whether the coefficient on sale below FIFO reference price equals the coefficient on sale above FIFO reference price. Standard errors are clustered by fiscal year-quarter and firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 6 Earnings surprises Earnings surprise Earnings surprise Earnings surprise Earnings surprise Earnings surprise Sale below FIFO reference price –0.801*** –0.963*** –0.405*** –0.168 –0.125 (–4.42) (–5.65) (–2.89) (–1.15) (–0.87) Sale above FIFO reference price 2.107*** 1.555*** 0.774*** 0.160** 0.268*** (16.74) (14.34) (9.15) (1.98) (3.11) Buy –0.055 0.168** 0.284*** 0.423*** (–0.58) (2.03) (3.33) (5.32) Single estimate –3.183*** –2.173*** –1.541 –1.83*** (–15.24) (–12.33) (–8.60) (–8.80) Dispersion –39.15*** –27.28*** –27.04*** –40.24*** (–10.72) (–9.52) (–9.13) (–12.13) Number of estimates 0.172*** 0.104*** 0.0135 0.022 (14.53) (13.37) (1.30) (1.31) Earnings surprise (t-1) 0.213*** 0.198*** 0.145*** (22.54) (21.05) (15.37) Earnings surprise (t-2) 0.102*** 0.0950*** 0.0572*** (15.48) (13.69) (8.00) Earnings surprise (t-3) 0.0649*** 0.0601*** 0.0290*** (12.41) (11.00) (5.08) Earnings surprise (t-4) 0.0764*** 0.0704*** 0.0389*** (12.80) (11.45) (6.33) Stock characteristics No No No Yes Yes Fiscal year-quarter FEs Yes Yes Yes Yes Yes Firm FEs No No No No Yes F-statistic 186.77*** 171.07*** 52.56*** 4.03** 5.32** R-squared 0.0182 0.0384 0.1089 0.1181 0.1720 Number of observations 652,492 652,492 524,390 450,613 450,521 Earnings surprise Earnings surprise Earnings surprise Earnings surprise Earnings surprise Sale below FIFO reference price –0.801*** –0.963*** –0.405*** –0.168 –0.125 (–4.42) (–5.65) (–2.89) (–1.15) (–0.87) Sale above FIFO reference price 2.107*** 1.555*** 0.774*** 0.160** 0.268*** (16.74) (14.34) (9.15) (1.98) (3.11) Buy –0.055 0.168** 0.284*** 0.423*** (–0.58) (2.03) (3.33) (5.32) Single estimate –3.183*** –2.173*** –1.541 –1.83*** (–15.24) (–12.33) (–8.60) (–8.80) Dispersion –39.15*** –27.28*** –27.04*** –40.24*** (–10.72) (–9.52) (–9.13) (–12.13) Number of estimates 0.172*** 0.104*** 0.0135 0.022 (14.53) (13.37) (1.30) (1.31) Earnings surprise (t-1) 0.213*** 0.198*** 0.145*** (22.54) (21.05) (15.37) Earnings surprise (t-2) 0.102*** 0.0950*** 0.0572*** (15.48) (13.69) (8.00) Earnings surprise (t-3) 0.0649*** 0.0601*** 0.0290*** (12.41) (11.00) (5.08) Earnings surprise (t-4) 0.0764*** 0.0704*** 0.0389*** (12.80) (11.45) (6.33) Stock characteristics No No No Yes Yes Fiscal year-quarter FEs Yes Yes Yes Yes Yes Firm FEs No No No No Yes F-statistic 186.77*** 171.07*** 52.56*** 4.03** 5.32** R-squared 0.0182 0.0384 0.1089 0.1181 0.1720 Number of observations 652,492 652,492 524,390 450,613 450,521 This table shows the regressions of earnings surprises on insider sales and control variables. The dependent variable is the quarterly earnings surprise measured as the number of cents that the actual earnings per share differs from the median earnings per share estimate. All buy and sell variables are dummies that equal one if the stated trade occurred in the associated month. I consider all firm-months that are in the 2 months before an earnings announcement and follow the previous earnings announcement. Single estimate is a dummy that equals one if there was only one forecast for the associated firm-quarter. Dispersion equals the standard deviation of earnings forecasts. Number of estimates equals the number of analyst forecasts. I winsorize all earnings surprise variables, dispersion, and coverage at the 1% level. Stock characteristics include the return in the month before the earnings announcement, the return from month $$t-12$$ to month $$t-1$$ before the earnings announcement, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. The F-statistic tests whether the coefficient on sale below FIFO reference price equals the coefficient on sale above FIFO reference price. Standard errors are clustered by fiscal year-quarter and firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Overall, this evidence suggests that part of the negative return predictability of insider sales at a loss relative to insider sales at a gain is driven by information regarding the upcoming earnings announcement. 3.4 Individual-level regressions Having confirmed my prediction that sales at a loss have more predictive power than sales at a gain, I next examine whether an aversion to realizing losses is driving my result. That is, does my basic result stem from the fact that insiders require a stronger negative signal to sell at a loss than to sell at a gain? Over the next three subsections, I test alternatives to this view. In this subsection, I examine investor heterogeneity and insider position heterogeneity. In the next subsection, I look at whether my basic result is a consequence of the short swing rule. Finally, I examine whether my result stems from a correlation with the price path. Insiders surely differ in the sophistication of their trading. It seems possible that more sophisticated traders will make trades that are more informative about future returns, and also that they will exhibit the disposition effect less because they know it to be a mistake. If this is the case, we would then observe that sales at a loss are more informative than sales at a gain, but this would not be driven by any aversion to realizing losses. To address this concern, I estimate models with individual-firm fixed effects. That is, I compare the predictive power of realized losses relative to realized gains within each insider-firm combination. I consider a panel of all insider-firm-months where the insider makes a sale; doing so ensures that individual-firm fixed effects subtract the average predictability of a sale by the insider. Again, I control for common shocks within a particular month by using month fixed effects.10 I cluster standard errors by month when the dependent variable is the 1-month horizon. When the dependent variable is a return horizon longer than 1-month, I cluster standard errors by firm and by month. Specifically, I estimate the following equation for individual-firm $$i$$ in month $$t$$: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}\\ &\quad +\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}\\ &\quad +\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}\\ & \quad+\beta_{6}\:\textit{Sale}\;\textit{below}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{7}\:\textit{Sale}\;\textit{above}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{8}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}\\ & \quad+\gamma_{1}\:\textit{Individual-firm}+\gamma_{2}\:\textit{Month}+\epsilon_{i,t.}\label{eq:6} \end{split} \end{align} (6) I predict that $$\beta_{5}<\beta_{6}.$$ That is, I expect to see evidence within-insider that sales at a loss have more negative predictive power than sales at a gain. I present the results in the first two columns of Table 7. Within insider, I find that insider sales at a loss predict 1-month returns 33 bps lower than insider sales at a gain. The difference is statistically significant at the 10% level $$(F-\mbox{statistic}=3.03)$$. At the 1-year return horizon, the difference in return predictability between insider sales at a loss and insider sales at a gain within insider is almost 2%. This difference is statistically significant at the 5% level $$(F-\mbox{statistic}=5.62)$$. Table 7 Individual-level analysis 1-month return 1-year return 1-month return 1-year return Sale below FIFO reference price (t-1) –0.022 –0.72 (–0.11) (–0.89) Sale above FIFO reference price (t-1) 0.306** 1.16* (2.18) (1.92) Sale below moving average (t-1) –0.12 –0.68 (–1.10) (–1.65) Sale below moving average by purchaser (t-1) –0.45*** –1.17* (–2.96) (–1.87) Sale above moving average (t-1) –0.44*** –1.30*** (–5.79) (–4.48) Sale above moving average by purchaser (t-1) 0.19* 0.65 (1.82) (1.50) Purchaser (t-1) 0.05 –0.63 (0.51) (–1.47) F-statistic 3.03* 5.61** 2.84* 2.47 Controls Yes Yes Yes Yes R-squared 0.4080 0.6040 0.1749 0.3779 Adjusted R-squared 0.1941 0.4598 0.1573 0.3646 Number of observations 140,859 133,710 6,937,616 6,625,582 1-month return 1-year return 1-month return 1-year return Sale below FIFO reference price (t-1) –0.022 –0.72 (–0.11) (–0.89) Sale above FIFO reference price (t-1) 0.306** 1.16* (2.18) (1.92) Sale below moving average (t-1) –0.12 –0.68 (–1.10) (–1.65) Sale below moving average by purchaser (t-1) –0.45*** –1.17* (–2.96) (–1.87) Sale above moving average (t-1) –0.44*** –1.30*** (–5.79) (–4.48) Sale above moving average by purchaser (t-1) 0.19* 0.65 (1.82) (1.50) Purchaser (t-1) 0.05 –0.63 (0.51) (–1.47) F-statistic 3.03* 5.61** 2.84* 2.47 Controls Yes Yes Yes Yes R-squared 0.4080 0.6040 0.1749 0.3779 Adjusted R-squared 0.1941 0.4598 0.1573 0.3646 Number of observations 140,859 133,710 6,937,616 6,625,582 This table presents regressions of returns on insider sales with controls at the individual level. I consider a panel of all insider-firm combinations. In the first two columns, I include all insider-firm-month observations in my sample where the insider executed a sale. In the final two columns, I include all insider-firm-month observations for insiders from the month of their first transaction at the firm to the month of their last transaction in my sample. All sale variables are dummies that equal one if there was the relevant trade in the associated individual-firm-month. The FIFO reference price is a reference price constructed using FIFO (first-in first-out) accounting. The moving average reference price equals the average of the previous 6-month-end prices. A purchaser is an insider who has made a purchase. I include insider-firm and month fixed effects. Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, a dummy that equals one if the insider made a purchase in month $$t-1$$, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. In the final four columns, I also include a dummy that equals one if month $$(t-2)$$’s closing price was below the moving-average reference price. I multiply all coefficient estimates by 100. In the first two columns, the F-statistic tests whether the coefficient on sale below FIFO reference price $$(t-1)$$ equals the coefficient on sale above FIFO reference $$(t-1)$$ price. In the last two columns, the F-statistic tests whether the predictive power of a sale below the moving average reference price by a purchaser $$(t-1)$$ equals the predictive power of a sale above the moving average by a purchaser $$(t-1)$$. When the return horizon is 1 month, I cluster standard errors by month; when the return horizon is longer, I cluster standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 7 Individual-level analysis 1-month return 1-year return 1-month return 1-year return Sale below FIFO reference price (t-1) –0.022 –0.72 (–0.11) (–0.89) Sale above FIFO reference price (t-1) 0.306** 1.16* (2.18) (1.92) Sale below moving average (t-1) –0.12 –0.68 (–1.10) (–1.65) Sale below moving average by purchaser (t-1) –0.45*** –1.17* (–2.96) (–1.87) Sale above moving average (t-1) –0.44*** –1.30*** (–5.79) (–4.48) Sale above moving average by purchaser (t-1) 0.19* 0.65 (1.82) (1.50) Purchaser (t-1) 0.05 –0.63 (0.51) (–1.47) F-statistic 3.03* 5.61** 2.84* 2.47 Controls Yes Yes Yes Yes R-squared 0.4080 0.6040 0.1749 0.3779 Adjusted R-squared 0.1941 0.4598 0.1573 0.3646 Number of observations 140,859 133,710 6,937,616 6,625,582 1-month return 1-year return 1-month return 1-year return Sale below FIFO reference price (t-1) –0.022 –0.72 (–0.11) (–0.89) Sale above FIFO reference price (t-1) 0.306** 1.16* (2.18) (1.92) Sale below moving average (t-1) –0.12 –0.68 (–1.10) (–1.65) Sale below moving average by purchaser (t-1) –0.45*** –1.17* (–2.96) (–1.87) Sale above moving average (t-1) –0.44*** –1.30*** (–5.79) (–4.48) Sale above moving average by purchaser (t-1) 0.19* 0.65 (1.82) (1.50) Purchaser (t-1) 0.05 –0.63 (0.51) (–1.47) F-statistic 3.03* 5.61** 2.84* 2.47 Controls Yes Yes Yes Yes R-squared 0.4080 0.6040 0.1749 0.3779 Adjusted R-squared 0.1941 0.4598 0.1573 0.3646 Number of observations 140,859 133,710 6,937,616 6,625,582 This table presents regressions of returns on insider sales with controls at the individual level. I consider a panel of all insider-firm combinations. In the first two columns, I include all insider-firm-month observations in my sample where the insider executed a sale. In the final two columns, I include all insider-firm-month observations for insiders from the month of their first transaction at the firm to the month of their last transaction in my sample. All sale variables are dummies that equal one if there was the relevant trade in the associated individual-firm-month. The FIFO reference price is a reference price constructed using FIFO (first-in first-out) accounting. The moving average reference price equals the average of the previous 6-month-end prices. A purchaser is an insider who has made a purchase. I include insider-firm and month fixed effects. Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, a dummy that equals one if the insider made a purchase in month $$t-1$$, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. In the final four columns, I also include a dummy that equals one if month $$(t-2)$$’s closing price was below the moving-average reference price. I multiply all coefficient estimates by 100. In the first two columns, the F-statistic tests whether the coefficient on sale below FIFO reference price $$(t-1)$$ equals the coefficient on sale above FIFO reference $$(t-1)$$ price. In the last two columns, the F-statistic tests whether the predictive power of a sale below the moving average reference price by a purchaser $$(t-1)$$ equals the predictive power of a sale above the moving average by a purchaser $$(t-1)$$. When the return horizon is 1 month, I cluster standard errors by month; when the return horizon is longer, I cluster standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. It is likely that insider characteristics are not static and that these characteristics could relate to the insider’s position in the firm. It is possible that different insider types exhibit less of a disposition effect, and have more predictive trades. In particular, this is a concern as CFO trades are more informative for future returns than CEO trades (Wang et al. 2012). I investigate this alternative explanation by considering the behavior of different types of insiders. In Table 8, I present the propensity to sell at a gain and the propensity to sell at a loss for different types of insiders. I consider insider types that occupy at least 5% of my sample, and all exhibit the disposition effect. Beneficial owners exhibit the weakest disposition effect. When I exclude the sales of beneficial owners from my analysis, the results from estimating Equation (1) are very similar to the results from estimating Equation (1) when I include the sales of beneficial owners.11 Table 8 Insider role in the firm Number of Acquisitions, Sales and Purchase Days PGR PLR Ratio Director 556,714 0.0132 0.0073 1.80 Officer 301,046 0.0347 0.0215 1.61 Vice president 103,951 0.0229 0.0134 1.70 Beneficial owner 122,548 0.0410 0.0349 1.17 CEO 91,376 0.0302 0.0148 2.03 CFO 71,351 0.0317 0.0163 1.94 Number of Acquisitions, Sales and Purchase Days PGR PLR Ratio Director 556,714 0.0132 0.0073 1.80 Officer 301,046 0.0347 0.0215 1.61 Vice president 103,951 0.0229 0.0134 1.70 Beneficial owner 122,548 0.0410 0.0349 1.17 CEO 91,376 0.0302 0.0148 2.03 CFO 71,351 0.0317 0.0163 1.94 This table looks at the characteristics of insiders by their position in the firm. I present statistics for types of insider positions that occupy at least 5% of the sample. The table presents the number of acquisitions, sales, and purchase days by insider type. PGR, the proportion of gains realized; PLR, the proportion of losses realized; Ratio, PGR divided by PLR. Table 8 Insider role in the firm Number of Acquisitions, Sales and Purchase Days PGR PLR Ratio Director 556,714 0.0132 0.0073 1.80 Officer 301,046 0.0347 0.0215 1.61 Vice president 103,951 0.0229 0.0134 1.70 Beneficial owner 122,548 0.0410 0.0349 1.17 CEO 91,376 0.0302 0.0148 2.03 CFO 71,351 0.0317 0.0163 1.94 Number of Acquisitions, Sales and Purchase Days PGR PLR Ratio Director 556,714 0.0132 0.0073 1.80 Officer 301,046 0.0347 0.0215 1.61 Vice president 103,951 0.0229 0.0134 1.70 Beneficial owner 122,548 0.0410 0.0349 1.17 CEO 91,376 0.0302 0.0148 2.03 CFO 71,351 0.0317 0.0163 1.94 This table looks at the characteristics of insiders by their position in the firm. I present statistics for types of insider positions that occupy at least 5% of the sample. The table presents the number of acquisitions, sales, and purchase days by insider type. PGR, the proportion of gains realized; PLR, the proportion of losses realized; Ratio, PGR divided by PLR. Next, I look at the difference in return predictability between realized gains and realized losses for different types of insiders. I again consider a sample of individual-firm-months with a sale and reestimate Equation (6) for different types of insiders. Table 9 presents the difference in return predictability between sales at a loss and sales at a gain for different types of insiders. There is limited identification because I am looking within insider in a small subsample based on insider type. As such, it is not surprising that I do not find much statistical significance. That being said, I find that sales at a loss predict more negative returns than sales at a gain for every insider type at the 1-year horizon, except for beneficial owners. It is not shocking that a sale at a loss by a beneficial owner does not predict more negative returns than a sale at a gain by a beneficial owner as beneficial owners exhibit the weakest disposition effect in my sample. It is surprising that we see a large difference in return predictability between a sale at a loss by a director and a sale at a gain by a director as we typically think of directors as being at the bottom of the information pyramid in a firm (Seyhun 1998). However, a deeper examination suggests that the trades of directors in my sample have predictive power. A purchase by a director in the lagged month predicts a 1-year return about 2.47% higher than all other firm-months. Table 9 Return predictability by insider type Difference at a 1-month horizon Difference at 1-year horizon Director 0.44 –1.78 Officer –0.33 –2.01* Vice president –1.25** –0.52 Beneficial owner –0.45 4.35 CEO –0.20 –1.01 CFO –0.53 –2.53 Difference at a 1-month horizon Difference at 1-year horizon Director 0.44 –1.78 Officer –0.33 –2.01* Vice president –1.25** –0.52 Beneficial owner –0.45 4.35 CEO –0.20 –1.01 CFO –0.53 –2.53 In this table, we demonstrate the difference in return predictability between sales at a loss relative to the FIFO reference price and sales at a gain relative to the FIFO reference price by insider type. I include all insider-firm-month observations in my sample where the insider executed a sale. I then create subsamples based on insider type. I present statistics for insider position types that occupy at least 5% of the sample. Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, a dummy that equals one if the insider made a purchase in month $$t-1$$, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. I include insider-firm and month fixed effects. When the return horizon is 1 month, I cluster standard errors by month; when the return horizon is longer, I cluster standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 9 Return predictability by insider type Difference at a 1-month horizon Difference at 1-year horizon Director 0.44 –1.78 Officer –0.33 –2.01* Vice president –1.25** –0.52 Beneficial owner –0.45 4.35 CEO –0.20 –1.01 CFO –0.53 –2.53 Difference at a 1-month horizon Difference at 1-year horizon Director 0.44 –1.78 Officer –0.33 –2.01* Vice president –1.25** –0.52 Beneficial owner –0.45 4.35 CEO –0.20 –1.01 CFO –0.53 –2.53 In this table, we demonstrate the difference in return predictability between sales at a loss relative to the FIFO reference price and sales at a gain relative to the FIFO reference price by insider type. I include all insider-firm-month observations in my sample where the insider executed a sale. I then create subsamples based on insider type. I present statistics for insider position types that occupy at least 5% of the sample. Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, a dummy that equals one if the insider made a purchase in month $$t-1$$, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. I include insider-firm and month fixed effects. When the return horizon is 1 month, I cluster standard errors by month; when the return horizon is longer, I cluster standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. I interpret this evidence as suggestive that my result is not being driven by investor heterogeneity or insider position heterogeneity. 3.5 Short swing rule Having considered insider heterogeneity as the driver of my basic result, I now examine whether it may instead be driven by the short swing rule. This rule (15 U.S. Code § 78p) states that all profits realized by an insider from executing two offsetting transactions within a 6-month period (a buy and a subsequent sale or a sale and a subsequent buy) are recoverable by the issuer.12 As such, it is unlikely that an insider would purchase or sell stock if she had any intention to complete an offsetting transaction in the near future. Therefore, in the six months following a purchase, it is unlikely that an insider will feel the need to sell shares for liquidity or diversification reasons. As such, sales executed shortly after purchases are presumably very informative. And, as a consequence of the regulatory environment (i.e., the short swing rule), insiders are strictly penalized for realizing gains. This could drive my result as the legal environment prevents realized gains and permits realized losses during a window where insiders likely only make informed trades. I test this alternative explanation by looking at the information content of realized losses made at least six months after the most recent purchase and the information content of realized gains made at least six months after the most recent purchase. I construct a dummy variable that equals one if there was a realized loss (gain) at least 180 days after the most recent purchase in month $$(t-1)$$. For comparison purposes (the short swing rule assumes a sort of LIFO accounting), I use the most recent purchase price—the transaction that an insider would be most worried about offsetting—as the reference price. I also consider this a natural reference price, because it is likely very salient to the insider due to the recency bias. I find minimal difference in the return predictability difference between sales at a loss and sales at a gain when I consider sales that happen long after the purchase, and the return predictability difference when I consider all sales. I present the results in Table 10. Table 10 Short swing rule 1-month return 1-month return 6-month return 6-month return 1-year return 1-year return Buy (t-1) 0.651*** 0.642*** 1.87*** 1.84*** 2.86*** 2.82*** (7.86) (7.77) (8.45) (8.32) (6.71) (6.64) Sale below purchase (t-1) –0.624*** –2.67*** –3.29*** (–3.87) (–5.01) (–3.66) Sale above purchase (t-1) –0.394*** –0.644*** –0.615 (–4.16) (–2.64) (–1.54) Slow sale below purchase (t-1) –0.620*** –2.67*** –3.30*** (–3.45) (–4.54) (–3.21) Slow sale above purchase (t-1) –0.403*** –0.686*** –0.615 (–4.16) (–2.64) (–1.49) Controls Yes Yes Yes Yes Yes Yes R-squared 0.1101 0.1101 0.1215 0.1215 0.0965 0.0965 F-statistic 1.32 1.05 12.02*** 9.52*** 6.91*** 5.48** Number of observations 1,352,175 1,352,175 1,297,278 1,297,278 1,229,697 1,229,697 1-month return 1-month return 6-month return 6-month return 1-year return 1-year return Buy (t-1) 0.651*** 0.642*** 1.87*** 1.84*** 2.86*** 2.82*** (7.86) (7.77) (8.45) (8.32) (6.71) (6.64) Sale below purchase (t-1) –0.624*** –2.67*** –3.29*** (–3.87) (–5.01) (–3.66) Sale above purchase (t-1) –0.394*** –0.644*** –0.615 (–4.16) (–2.64) (–1.54) Slow sale below purchase (t-1) –0.620*** –2.67*** –3.30*** (–3.45) (–4.54) (–3.21) Slow sale above purchase (t-1) –0.403*** –0.686*** –0.615 (–4.16) (–2.64) (–1.49) Controls Yes Yes Yes Yes Yes Yes R-squared 0.1101 0.1101 0.1215 0.1215 0.0965 0.0965 F-statistic 1.32 1.05 12.02*** 9.52*** 6.91*** 5.48** Number of observations 1,352,175 1,352,175 1,297,278 1,297,278 1,229,697 1,229,697 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade occurred in the associated firm-month. The reference price Purchase equals the most recent purchase price. To indicate that the sale happened at least 180 days after the most recent purchase, I use the adjective “slow.” Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. I multiply all coefficient estimates by 100. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than one month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 10 Short swing rule 1-month return 1-month return 6-month return 6-month return 1-year return 1-year return Buy (t-1) 0.651*** 0.642*** 1.87*** 1.84*** 2.86*** 2.82*** (7.86) (7.77) (8.45) (8.32) (6.71) (6.64) Sale below purchase (t-1) –0.624*** –2.67*** –3.29*** (–3.87) (–5.01) (–3.66) Sale above purchase (t-1) –0.394*** –0.644*** –0.615 (–4.16) (–2.64) (–1.54) Slow sale below purchase (t-1) –0.620*** –2.67*** –3.30*** (–3.45) (–4.54) (–3.21) Slow sale above purchase (t-1) –0.403*** –0.686*** –0.615 (–4.16) (–2.64) (–1.49) Controls Yes Yes Yes Yes Yes Yes R-squared 0.1101 0.1101 0.1215 0.1215 0.0965 0.0965 F-statistic 1.32 1.05 12.02*** 9.52*** 6.91*** 5.48** Number of observations 1,352,175 1,352,175 1,297,278 1,297,278 1,229,697 1,229,697 1-month return 1-month return 6-month return 6-month return 1-year return 1-year return Buy (t-1) 0.651*** 0.642*** 1.87*** 1.84*** 2.86*** 2.82*** (7.86) (7.77) (8.45) (8.32) (6.71) (6.64) Sale below purchase (t-1) –0.624*** –2.67*** –3.29*** (–3.87) (–5.01) (–3.66) Sale above purchase (t-1) –0.394*** –0.644*** –0.615 (–4.16) (–2.64) (–1.54) Slow sale below purchase (t-1) –0.620*** –2.67*** –3.30*** (–3.45) (–4.54) (–3.21) Slow sale above purchase (t-1) –0.403*** –0.686*** –0.615 (–4.16) (–2.64) (–1.49) Controls Yes Yes Yes Yes Yes Yes R-squared 0.1101 0.1101 0.1215 0.1215 0.0965 0.0965 F-statistic 1.32 1.05 12.02*** 9.52*** 6.91*** 5.48** Number of observations 1,352,175 1,352,175 1,297,278 1,297,278 1,229,697 1,229,697 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade occurred in the associated firm-month. The reference price Purchase equals the most recent purchase price. To indicate that the sale happened at least 180 days after the most recent purchase, I use the adjective “slow.” Controls include the previous month’s return, the return from month $$t-12$$ to month $$t-1$$, the log value of common equity divided by market capitalization, the log value of market capitalization, and the capital gains overhang, which is a measure of how the price relates to the average investor’s reference price. I multiply all coefficient estimates by 100. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than one month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. 3.6 Correlation with the price path 3.6.1 Opportunistic and routine traders As the final alternative to the hypothesis that my basic result is driven by insider aversion to realizing losses, I examine whether this result stems from a correlation with the price path. For example, the timing of an insider trade at a loss may be correlated with the recent path of the stock price in some way that is not fully captured by my controls for short-term reversals, momentum, and capital gains overhang. To examine this explanation, I first consider a placebo test. I test whether there is differential predictive power for future returns between uninformed realized losses and uninformed realized gains. Cohen et al. (2012) use a filter to distinguish informative trades from uninformative trades. The authors label informative trades “opportunistic” and uninformative trades “routine”. They show that “opportunistic” sales contain all the information content in their universe of sales. Since routine sales do not have any informational content, I do not expect there to be a difference in predictive power for future returns between routine sales at a loss and routine sales at a gain. Of course, I still expect there to be a significant difference between the information content of opportunistic realized losses and opportunistic realized gains. For this test, I use the 6-month moving average as the reference price because it does not require a trading history and therefore does not interfere with Cohen et al.’s (2012) identification of routine and opportunistic traders. Also, I use the sample of all insiders instead of limiting the sample to “inactive” insiders as I do not want to interfere with Cohen et al.’s (2012) classification schema. I consider a sample of open-market purchases and sales.13 Following Cohen et al. (2012), I classify insiders at the beginning of each year as opportunistic or routine. An insider is classified as a routine trader if she made a trade (a purchase or sale) in the same month for three consecutive years (e.g. an insider who made a trade in April 2001, April 2002, and April 2003 would be considered a routine trader from January 2004 to present). All insiders who are not classified as routine are classified as opportunistic. I only consider trades by insiders who have three consecutive years of trading history. This makes it possible for an insider to be classified as a routine trader. To allow a 3-year trading history to be built, I restrict my sample to firm-months after 1989. By definition, opportunistic traders can be reclassified as routine traders at the beginning of each year, but routine traders stay routine. Again, I run pooled OLS regressions with month fixed effects and cluster standard errors by month. For firm $$i$$ in month $$t$$, I estimate the model: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}\\ &\quad +\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}\\ &\quad +\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Opportunistic}\;\textit{buy}_{i,t-1}+\beta_{6}\:\textit{Routine}\;\textit{buy}_{i,t-1}\\ &\quad +\beta_{7}\:\textit{Routine}\;\textit{realized}\;\textit{gain}_{i,t-1}+\beta_{8}\:\textit{Routine}\;\textit{realized}\;\textit{loss}{}_{i,t-1}\\ &\quad +\beta_{9}\:\textit{Previous}\;\textit{month-end}\;\textit{price}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{10}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma \textit{Month}+\epsilon_{i,t}.\label{eq:7} \end{split} \end{align} (7) My prediction is that $$\beta_{7}=\beta_{8}.$$ That is, I do not expect there to be any difference between the predictive power of future returns for realized losses relative to realized gains among routine trades. I present the results in Table 11. Like Cohen et al. (2012), opportunistic sales predict more negative returns than routine sales and the difference is significant at the 1% level $$(F-\mbox{statistic}=13.65)$$. Consistent with my result reflecting the information content of insider sales, I fail to reject the null hypothesis that the predictive power of routine realized gains equals the predictive power of routine realized losses. If anything, routine sales below the reference price predict more positive returns than routine sales above the reference price. On the other hand, the difference in predictive power between opportunistic realized gains, $$-$$15 basis points, and opportunistic realized losses, $$-$$74 bps, is statistically significant at the 1% level $$(F-\mbox{statistic}=10.10)$$. Table 11 Opportunistic and routine trades 1-month return 1-month return 1-month return Opportunistic buy (t-1) 0.459*** 0.392*** 0.404*** (3.24) (2.88) (2.97) Routine buy (t-1) –0.145 –0.226 –0.222 (–0.72) (–1.22) (–1.20) Opportunistic sale (t-1) –0.233** (–2.96) Routine sale (t-1) 0.244** (2.16) Routine sale below moving average (t-1) 0.292 (1.37) Routine sale above moving average (t-1) 0.012 (0.10) Opportunistic sale below moving average (t-1) –0.742*** (–4.47) Opportunistic sale above moving average (t-1) –0.151* (–1.79) Below moving average (t-1) –0.391*** –0.386*** (–3.65) (–3.60) Capital gains overhang 0.213 0.213 (0.70) (0.70) Standard stock characteristic controls Yes Yes Yes R-squared 0.0992 0.1025 0.1025 F-statistic 13.65*** 1.25 10.10*** Number of observations 1,444,601 1,230,813 1,230,813 1-month return 1-month return 1-month return Opportunistic buy (t-1) 0.459*** 0.392*** 0.404*** (3.24) (2.88) (2.97) Routine buy (t-1) –0.145 –0.226 –0.222 (–0.72) (–1.22) (–1.20) Opportunistic sale (t-1) –0.233** (–2.96) Routine sale (t-1) 0.244** (2.16) Routine sale below moving average (t-1) 0.292 (1.37) Routine sale above moving average (t-1) 0.012 (0.10) Opportunistic sale below moving average (t-1) –0.742*** (–4.47) Opportunistic sale above moving average (t-1) –0.151* (–1.79) Below moving average (t-1) –0.391*** –0.386*** (–3.65) (–3.60) Capital gains overhang 0.213 0.213 (0.70) (0.70) Standard stock characteristic controls Yes Yes Yes R-squared 0.0992 0.1025 0.1025 F-statistic 13.65*** 1.25 10.10*** Number of observations 1,444,601 1,230,813 1,230,813 This table presents regressions of 1-month returns on different types of insider sales at the firm level. I consider a sample of firm-months from 1989 to 2015. The dependent variable in these panel regressions is the future 1-month return $$(t,t+1)$$. I consider insiders who have traded for 3 consecutive years. Insiders are labeled routine if they made a trade in the same month for 3 consecutive years. All remaining insiders are classified as opportunistic. Dummies for opportunistic (routine) buys and sales take a value of one for a specific firm-month entry if a firm’s opportunistic (routine) insider bought or sold, respectively, in that month. Standard stock characteristic controls include the return from $$t-1$$ to $$t$$, the return from $$t-12$$ to $$t-1$$, the log value of common equity divided by market capitalization, and the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I include month fixed effects. Standard errors are clustered by month. Moving average is the reference price, and is equal to the average of the previous 6-month-end prices. Below moving average$$(t-1)$$ is a dummy that equals one if the previous month-end price is below the moving-average. The F-statistic in the first column tests whether the coefficient on opportunistic sale $$(t-1)$$ equals the coefficient on routine sale $$(t-1)$$. The F-statistic in the second column tests whether the coefficient on routine sale below moving average $$(t-1)$$ equals the coefficient on routine sale above moving average $$(t-1)$$. The F-statistic for the third column tests whether the coefficient on opportunistic sale below moving average $$(t-1)$$ equals the coefficient on opportunistic sale above moving average $$(t-1)$$. I multiply all coefficient estimates by 100. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 11 Opportunistic and routine trades 1-month return 1-month return 1-month return Opportunistic buy (t-1) 0.459*** 0.392*** 0.404*** (3.24) (2.88) (2.97) Routine buy (t-1) –0.145 –0.226 –0.222 (–0.72) (–1.22) (–1.20) Opportunistic sale (t-1) –0.233** (–2.96) Routine sale (t-1) 0.244** (2.16) Routine sale below moving average (t-1) 0.292 (1.37) Routine sale above moving average (t-1) 0.012 (0.10) Opportunistic sale below moving average (t-1) –0.742*** (–4.47) Opportunistic sale above moving average (t-1) –0.151* (–1.79) Below moving average (t-1) –0.391*** –0.386*** (–3.65) (–3.60) Capital gains overhang 0.213 0.213 (0.70) (0.70) Standard stock characteristic controls Yes Yes Yes R-squared 0.0992 0.1025 0.1025 F-statistic 13.65*** 1.25 10.10*** Number of observations 1,444,601 1,230,813 1,230,813 1-month return 1-month return 1-month return Opportunistic buy (t-1) 0.459*** 0.392*** 0.404*** (3.24) (2.88) (2.97) Routine buy (t-1) –0.145 –0.226 –0.222 (–0.72) (–1.22) (–1.20) Opportunistic sale (t-1) –0.233** (–2.96) Routine sale (t-1) 0.244** (2.16) Routine sale below moving average (t-1) 0.292 (1.37) Routine sale above moving average (t-1) 0.012 (0.10) Opportunistic sale below moving average (t-1) –0.742*** (–4.47) Opportunistic sale above moving average (t-1) –0.151* (–1.79) Below moving average (t-1) –0.391*** –0.386*** (–3.65) (–3.60) Capital gains overhang 0.213 0.213 (0.70) (0.70) Standard stock characteristic controls Yes Yes Yes R-squared 0.0992 0.1025 0.1025 F-statistic 13.65*** 1.25 10.10*** Number of observations 1,444,601 1,230,813 1,230,813 This table presents regressions of 1-month returns on different types of insider sales at the firm level. I consider a sample of firm-months from 1989 to 2015. The dependent variable in these panel regressions is the future 1-month return $$(t,t+1)$$. I consider insiders who have traded for 3 consecutive years. Insiders are labeled routine if they made a trade in the same month for 3 consecutive years. All remaining insiders are classified as opportunistic. Dummies for opportunistic (routine) buys and sales take a value of one for a specific firm-month entry if a firm’s opportunistic (routine) insider bought or sold, respectively, in that month. Standard stock characteristic controls include the return from $$t-1$$ to $$t$$, the return from $$t-12$$ to $$t-1$$, the log value of common equity divided by market capitalization, and the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I include month fixed effects. Standard errors are clustered by month. Moving average is the reference price, and is equal to the average of the previous 6-month-end prices. Below moving average$$(t-1)$$ is a dummy that equals one if the previous month-end price is below the moving-average. The F-statistic in the first column tests whether the coefficient on opportunistic sale $$(t-1)$$ equals the coefficient on routine sale $$(t-1)$$. The F-statistic in the second column tests whether the coefficient on routine sale below moving average $$(t-1)$$ equals the coefficient on routine sale above moving average $$(t-1)$$. The F-statistic for the third column tests whether the coefficient on opportunistic sale below moving average $$(t-1)$$ equals the coefficient on opportunistic sale above moving average $$(t-1)$$. I multiply all coefficient estimates by 100. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. 3.6.2 Difference in return predictability around the break-even point I perform another test to see whether I am capturing price path effects. Specifically, I look at the difference in return predictability in a narrow band around the reference point. Losses and gains are still fundamentally different, but there will not be a lot of variation, by construction, in the distance from the reference price. As such, any difference in return predictability, in this narrow band, can more plausibly be attributed to insider information than the price path. Specifically, I consider a regression of the following form: \begin{align*} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}+\beta_{2}\:\textit{Momentum}_{i,t}+\\ &\quad \beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}+\\ &\quad \beta_{6}\:\textit{Sale}\;\textit{slightly}\;\textit{below}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+ \end{split} \end{align*} \begin{align} \begin{split} &\quad \beta_{7}\:\textit{Sale}\;\textit{slightly}\;\textit{above}\;\textit{FIFO}\;\textit{reference}\;\textit{price}_{i,t-1}+\\ &\quad \beta_{8}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma\:\textit{Month}+\epsilon_{i,t}.\label{eq:narrow_band} \end{split} \end{align} (8) The band I consider runs from 10% below the reference price to 10% above the reference price. As such, Sale slightly below FIFO reference price is a sale below the reference price, but no less than 10% below the reference price. My prediction is that $$\beta_{6}<\beta_{7}.$$ I find evidence consistent with this hypothesis. I present the results in Table 12. At the 1-month horizon a sale slightly below the reference price predicts a 1-month return 67 bps lower, whereas a sale slightly above the reference price predicts a 1-month return only 19 basis points lower. The difference is statistically significant at the 1% level. At the 1-year horizon a sale slightly below the reference price predicts a return 258 bps lower, whereas a sale slightly above the reference price predicts a 1-month return only 28 basis points lower. Again, the difference is statistically significant at the 1% level. In Figure 1, I present the coefficient estimates from a regression of future 1-month returns on ten different buckets measuring distance from the FIFO reference price. (I use the same controls used in Equation (8).) The figure also offers evidence of a jump at the reference price. Table 12 Difference in a narrow band 1-month return 6-month return 1-year return Short-term reversal –2.38** –1.16 1.83 (–2.71) (–0.86) (0.78) Momentum 0.308 0.617 –0.836 (1.61) (1.51) (–1.42) Book-to-market 0.321*** 1.68*** 3.35*** (3.59) (5.14) (6.96) Size –0.0207 –0.154 –0.416** (–0.47) (–1.22) (–2.36) Buy (t-1) 0.639*** 1.83*** 2.82*** (7.75) (8.28) (6.61) Sale slightly below FIFO (t-1) –0.665*** –1.78*** –2.58*** (–4.69) (–4.65) (–4.07) Sale slightly above FIFO (t-1) –0.188* –0.176 –0.276 (–1.73) (–0.48) (–0.52) Capital gains overhang 0.327 0.799 –1.99 (1.14) (1.11) (–1.54) R-squared 0.1101 0.1214 0.0965 F-statistic 7.44*** 10.52*** 10.27*** Number of observations 1,352,175 1,297,278 1,229,697 1-month return 6-month return 1-year return Short-term reversal –2.38** –1.16 1.83 (–2.71) (–0.86) (0.78) Momentum 0.308 0.617 –0.836 (1.61) (1.51) (–1.42) Book-to-market 0.321*** 1.68*** 3.35*** (3.59) (5.14) (6.96) Size –0.0207 –0.154 –0.416** (–0.47) (–1.22) (–2.36) Buy (t-1) 0.639*** 1.83*** 2.82*** (7.75) (8.28) (6.61) Sale slightly below FIFO (t-1) –0.665*** –1.78*** –2.58*** (–4.69) (–4.65) (–4.07) Sale slightly above FIFO (t-1) –0.188* –0.176 –0.276 (–1.73) (–0.48) (–0.52) Capital gains overhang 0.327 0.799 –1.99 (1.14) (1.11) (–1.54) R-squared 0.1101 0.1214 0.0965 F-statistic 7.44*** 10.52*** 10.27*** Number of observations 1,352,175 1,297,278 1,229,697 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade(s) occurred in the associated firm-month. “Slightly” indicates that the sale occurred no greater than 10% from the reference price. The reference price FIFO is based on FIFO accounting. Short-term reversal equals the return from $$t-1$$ to $$t$$, and Momentum equals the return from $$t-12$$ to $$t-1$$. Book-to-market is the log value of common equity divided by market capitalization. Size is the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I drop all observations in which the month-end price is less than one. I multiply all coefficient estimates by 100. The F-statistic tests whether the coefficient on sale slightly below FIFO $$(t-1)$$ equals the coefficient on sale slightly above FIFO $$(t-1)$$. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than one month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. Table 12 Difference in a narrow band 1-month return 6-month return 1-year return Short-term reversal –2.38** –1.16 1.83 (–2.71) (–0.86) (0.78) Momentum 0.308 0.617 –0.836 (1.61) (1.51) (–1.42) Book-to-market 0.321*** 1.68*** 3.35*** (3.59) (5.14) (6.96) Size –0.0207 –0.154 –0.416** (–0.47) (–1.22) (–2.36) Buy (t-1) 0.639*** 1.83*** 2.82*** (7.75) (8.28) (6.61) Sale slightly below FIFO (t-1) –0.665*** –1.78*** –2.58*** (–4.69) (–4.65) (–4.07) Sale slightly above FIFO (t-1) –0.188* –0.176 –0.276 (–1.73) (–0.48) (–0.52) Capital gains overhang 0.327 0.799 –1.99 (1.14) (1.11) (–1.54) R-squared 0.1101 0.1214 0.0965 F-statistic 7.44*** 10.52*** 10.27*** Number of observations 1,352,175 1,297,278 1,229,697 1-month return 6-month return 1-year return Short-term reversal –2.38** –1.16 1.83 (–2.71) (–0.86) (0.78) Momentum 0.308 0.617 –0.836 (1.61) (1.51) (–1.42) Book-to-market 0.321*** 1.68*** 3.35*** (3.59) (5.14) (6.96) Size –0.0207 –0.154 –0.416** (–0.47) (–1.22) (–2.36) Buy (t-1) 0.639*** 1.83*** 2.82*** (7.75) (8.28) (6.61) Sale slightly below FIFO (t-1) –0.665*** –1.78*** –2.58*** (–4.69) (–4.65) (–4.07) Sale slightly above FIFO (t-1) –0.188* –0.176 –0.276 (–1.73) (–0.48) (–0.52) Capital gains overhang 0.327 0.799 –1.99 (1.14) (1.11) (–1.54) R-squared 0.1101 0.1214 0.0965 F-statistic 7.44*** 10.52*** 10.27*** Number of observations 1,352,175 1,297,278 1,229,697 This table presents regressions of returns on insider sales with controls at the firm level. All buy and sell variables are dummies that equal one if the stated trade(s) occurred in the associated firm-month. “Slightly” indicates that the sale occurred no greater than 10% from the reference price. The reference price FIFO is based on FIFO accounting. Short-term reversal equals the return from $$t-1$$ to $$t$$, and Momentum equals the return from $$t-12$$ to $$t-1$$. Book-to-market is the log value of common equity divided by market capitalization. Size is the log value of market capitalization. Capital gains overhang is a control that looks at how the price relates to the average investor’s reference price. I drop all observations in which the month-end price is less than one. I multiply all coefficient estimates by 100. The F-statistic tests whether the coefficient on sale slightly below FIFO $$(t-1)$$ equals the coefficient on sale slightly above FIFO $$(t-1)$$. I include month fixed effects. If the dependent variable is the 1-month return, I cluster standard errors by month. If the dependent variable is at a horizon longer than one month, then I cluster the standard errors by month and by firm. * indicates significance at the 10% level; ** indicates significance at the 5% level; and *** indicates significance at the 1% level. 3.7 Rebalancing My maintained hypothesis is that the result in Section 4.2 is driven by the aversion to realizing losses inherent in the disposition effect. In the next three sections, I explore possible sources of this aversion to realizing losses, an aversion that makes sales at a loss much more informative than sales at a gain. I first examine a portfolio rebalancing explanation. Rebalancing motives are an important motive when an investor decides to sell a stock. Unlike taxes, rebalancing motives could differentially encourage investors to sell more at a gain compared to at a loss.14 After a fall in the share price, the stock likely makes up a smaller fraction of the investor’s portfolio than before. After a rise in the share price, the stock likely makes up a larger fraction of the investor’s portfolio than before. Therefore, rebalancing motives may encourage the investor to sell more after a rise in the share price compared to after a fall in the share price. If so, my result may stem from the fact that realized gains are more likely to be rebalancing trades than realized losses.15 To test this hypothesis, I compare the predictive power for future returns of complete liquidations at a loss with the predictive power for future returns of complete liquidations at a gain. Complete liquidations are likely not driven by rebalancing motives; as a result, if complete liquidations at a loss predict returns more negatively than complete liquidations at a gain, this would cast doubt on the rebalancing view. Specifically, I estimate the following equation for firm $$i$$ in month $$t$$, clustering standard errors by month: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}\\ &\quad+\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}+\beta_{5}\:\textit{Buy}_{i,t-1}\\ &\quad +\beta_{6}\:\textit{Complete}\;\textit{liquidation}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{7}\:\textit{Complete}\;\textit{liquidation}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{8}\:\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma\:\textit{Month}_{t}+\epsilon_{i,t}.\label{eq:8} \end{split} \end{align} (9) If rebalancing motives are driving my result, the prediction is that $$\beta_{6}=\beta_{7}.$$ I present the results in Table 4. I find that a complete liquidation at a loss predicts a 1-month return 483 bps lower than all other firm-months in my sample, whereas a complete liquidation at a gain predicts a one month return 121 bps lower. The F-statistic testing the difference between the two is 9.81, which is significant at the 1% level. I interpret this as evidence against a rebalancing explanation of my result. 3.8 An examination of realization utility Realization utility refers to the simple idea that investors derive utility directly from the act of realizing a gain or loss on an asset. Unlike the other explanations I have considered, this simple idea is consistent with the evidence presented thus far.16 Still, the underlying source of realization utility is unclear: why do investors experience disutility when they close out a position at a loss? In this section, I examine two proposed answers to this question – a heuristic-based explanation and a cognitive dissonance-based explanation. An investor who sells all of her positions at a gain (loss) makes (loses) money. This observation is the foundation for the heuristic that Barberis and Xiong (2012) argue underlies realization utility. Selling at a gain is good; selling at a loss is bad. The investor derives utility from doing something good (realizing gains) and derives disutility from doing something bad (realizing losses). Another possible source of realization disutility is a reluctance to admit that an earlier purchase decision was a mistake. Instead of admitting a mistake, investors may manipulate malleable, but negative, signals to preserve a positive self-image. This self-delusion relieves cognitive dissonance. An investor experiences cognitive dissonance, or “the discomfort that arises when a person recognizes that he or she makes choices and/or holds beliefs that are dissonant with each other,” if she thinks she is a talented investor, purchases stock, and then receives a signal indicating that the purchase was a mistake. Chang et al. (2016) highlight that investors can reduce the discomfort by either admitting their mistake, finding a third-party scapegoat, blaming bad luck, or explaining the bad performance as a temporary setback that will soon be reversed. As many people enjoy holding a positive self-image, admitting a mistake can be painful and a source of realization disutility. Company insiders are somewhat unique in that they acquire stock in two ways, either by actively purchasing shares or by being endowed with shares.17 This distinction allows me to shed light on the underlying source of realization utility. Under the heuristic explanation, a company insider will experience disutility if she sells shares at a loss regardless of whether she purchased the shares or was simply endowed with them. This predicts that a sale at a loss by an insider will be a more negative signal of future returns than a sale at a gain, regardless of whether the shares were purchased or endowed. On the other hand, under the cognitive dissonance view, the company insider will only feel pain from selling at a loss if the shares that she sells are shares that she actively purchased; if she was merely endowed with the shares, she has little reason to blame herself for their poor performance. This predicts that a sale of shares by an insider at a loss will be a more negative signal about the stock’s future return than a sale at a gain only if the shares that the insider sold were actively purchased shares. I test these competing theories by examining an individual-level panel data set. I include monthly observations for each insider from the month of her first transaction at the firm to the month of her last transaction in my data set. I examine the difference in return predictability between realized gains and realized losses for purchasers, or individuals who have previously purchased shares, and compare that to the difference in return predictability between realized gains and realized losses for nonpurchasers. To do this, I construct an interaction term between the realized loss dummy and the purchaser dummy, and an interaction term between the realized gain dummy and the purchaser dummy. If the “cognitive dissonance” explanation holds, the difference between realized losses and realized gains should be larger for purchasers than for nonpurchasers. Frydman and Rangel (2014) find that one can debias the disposition effect by reducing the saliency of the purchase price. It is plausible that purchasers will have a more salient reference price than those that have received shares. While reference price salience effects could still influence the results, I mitigate these effects by using a reference price that does not depend on purchase history, namely I use the 6-month moving-average as the reference price. To look at returns within insider, I include insider-firm level fixed effects. I use month fixed effects as I want to capture the firm-specific component of return predictability. Specifically, I estimate the following equation for individual-firm $$i$$ in month $$t$$, and cluster standard errors by month: \begin{align} \begin{split} \textit{Return}_{i,t\rightarrow t+1} & = \beta_{0}+\beta_{1}\:\textit{Short-term}\;\textit{reversal}_{i,t}\\ &\quad+\beta_{2}\:\textit{Momentum}_{i,t}+\beta_{3}\:\textit{Book-to-market}_{i,t}+\beta_{4}\:\textit{Size}_{i,t}\\ &\quad +\beta_{5}\:\textit{Buy}_{i,t-1}+\beta_{6}\:\textit{Purchaser}_{i,t-1}\\ & \quad+\beta_{7}\:\textit{Sale}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{8}\:\textit{Sale}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\beta_{9}\:\textit{Sale}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}*\textit{Purchaser}_{i,t-1}\\ &\quad +\beta_{10}\:\textit{Sale}\;\textit{above}\;\textit{reference}\;\textit{price}_{i,t-1}*\textit{Purchaser}_{i,t-1}\\ &\quad +\beta_{11\:}\textit{Previous}\;\textit{month-end}\;\textit{price}\;\textit{below}\;\textit{reference}\;\textit{price}_{i,t-1}\\ &\quad +\textit{Capital}\;\textit{gains}\;\textit{overhang}_{i,t}+\gamma_{1}\:\textit{Individual-firm} +\gamma_{2}\:\textit{Month}+\epsilon_{i,t.}\label{eq:9} \end{split} \end{align} (10) The cognitive dissonance view predicts that $$\beta_{7}+\beta_{9}<\beta_{8}+\beta_{10}.$$ I find evidence that sales at a loss by a purchaser are a more negative signal for future returns than sales at a gain by a purchaser. Additionally, I find little evidence that sales at a loss by nonpurchasers are a more negative signal for future returns than sales at a gain for nonpurchasers. I present the results in the third and fourth columns of Table 7. At the 1-month horizon, a sale at a loss by a purchaser predicts a 1-month return about 32 bps lower than a sale at a gain by a purchaser. This difference is statistically significant at the 1% level. At the 1-year horizon, a sale at a loss by a purchaser predicts a 1-year return about 120 bps lower than a sale at a gain by a purchaser. This difference is not statistically significant, but if I consider the entire sample, where there is more identification, I see similar, but statistically significant, results. I present these results in the Online Appendix. Importantly, a sale at a gain by a nonpurchaser has, if anything, more negative predictive power than a sale at a loss by a nonpurchaser. I interpret these findings as evidence for a “cognitive dissonance” view of realization utility. 4. Conclusion I provide evidence that insiders require a strong negative signal to realize a loss. Specifically, I document that a sale of stock at a loss by a company insider is a much more negative signal for future stock returns than a sale of stock by a company insider at a gain. This simple filtering technique allows us to extract information from the trades of informed investors in a more efficient way. I consider a range of explanations for my results, including investor heterogeneity, the short swing rule, and rebalancing motives, but find that the evidence is most consistent with the idea that investors derive direct disutility from selling a stock at a loss. Since selling at a loss is painful, an investor who does so must have particularly negative information, information that manifests itself in a poor stock return over the next few months. By comparing the predictive power of sales of endowed shares at a loss with the predictive power of purchased shares at a loss, I am able to shed light on the source of this disutility. I find evidence that selling at a loss is painful because it forces the investor to admit that an earlier purchase decision was a mistake. I would like to thank my adviser Nicholas Barberis for his support and guidance. In addition, I am grateful for helpful comments from Ian Ayres, Robert Battalio, Tom Chang, Aytekin Ertan, Cary Frydman, Robin Greenwood (editor), Sam Hartzmark, Lawrence Jin, Andrew Karolyi, Stephen Karolyi, Andrew Metrick, and Justin Murfin; two anonymous referees; and the seminar participants at Cornell University, Cubist Systematic Strategies, Emory University, the London School of Economics, the Ohio State University, Purdue University, UC Irvine, the University of Florida, the University of Notre Dame, the University of Utah, and Yale University. I also acknowledge financial support from a Whitebox Advisors grant. Supplementary data can be found on The Review of Financial Studies web site. Footnotes 1 Odean (1998) finds this behavior in a data set of 10,000 trading accounts from a large discount brokerage. Genesove and Mayer (2001) document an aversion to realizing losses in the downtown Boston housing market. Grinblatt and Keloharju (2001) obtain data on the trading of people and institutions in the Finnish stock market and find that investors are reluctant to sell at a loss. Frazzini (2006) documents this behavior in mutual fund managers. Hartzmark and Solomon (2012) look at a set of NFL betting contracts at Tradesports.com and uncover evidence consistent with the disposition effect. 2 Seyhun (1998) concludes that several different trading rules based on the trades of company insiders lead to profits. Jeng et al. (2003) highlight that insider sales do not earn abnormal returns. Lakonishok and Lee (2001) argue that insider selling appears to have no predictive ability. Gao et al. (2016) argue that the absence of insider selling predicts negative future returns. 3 Shefrin and Statman (1985) suggest that investors open (close) a mental account when purchasing (selling) a stock and then evaluate the transaction at the moment of sale. As such, the realization of gains/losses becomes a determinant of overall utility. Similarly, Thaler (1999, p.189) writes “one clear intuition is that a realized loss is more painful than a paper loss. When a stock is sold, the gain or loss has to be ‘declared’ both to the tax authorities and to the investor (and spouse).” An implicit assumption of my analysis relevant in these examples is that insiders view investments in isolation: they exhibit narrow framing (e.g., see Tversky and Kahneman 1981). Barberis and Xiong (2012) and Ingersoll and Jin (2013) incorporate these ideas into formal models to explain a number of puzzling facts. Frydman et al. (2014) find neuroscientific evidence largely consistent with realization utility. 4 I also include private transactions, because, after May 1991, private transactions use the same codes as open-market transactions. I do not consider sales associated with the exercise of an option. 5 This is Fidelity’s default method for nonmutual fund securities. I do not consider the effects of derivative transactions. Since insiders are not allowed to short sell their own company stock, I also ignore situations in which total holdings would go negative. 6 When the weekly volume is greater than the number of shares outstanding, the weekly turnover equals one. 7 Following Grinblatt and Han (2005), I lag the weekly closing price to mitigate microstructure effects. In all regressions, I control for the capital gains overhang by using the value that corresponds to the last week of the month. 8 I consider return horizons as long as 3 years. At the 3-year horizon, sales at a loss predict returns about 2.5 percentage points lower, but the difference is no longer statistically significant. 9 I do not scale the earnings surprise by price because Cheong and Thomas (2011) find that the earnings per share (EPS) forecast error and dispersion do not vary with scale. To minimize the effect of outliers, I winsorize all earnings surprise variables, dispersion, and coverage at the 1% level. 10 Instead of subtracting an average monthly return, or index return, from the dependent variable, I use the fixed effects estimator as this estimator is consistent (Gormley and Matsa 2014). I overcome the computational difficulties that arise when estimating models with high-dimensional fixed effects by using the Stata procedure “reg2hdfe” designed by Guimaraes and Portugal (2010). 11 In the Online Appendix, I estimate Equation (1) by insider type. I show that there is a strong difference between sales at a loss and sales at a gain for a number of insider types. 12 “For the purpose of preventing the unfair use of information which may have been obtained by such beneficial owner, director, or officer by reason of his relationship to the issuer, any profit realized by him from any purchase and sale, or any sale and purchase, of any equity security of such issuer (other than an exempted security) or a security-based swap agreement involving any such equity security within any period of less than six months, unless such security or security-based swap agreement was acquired in good faith in connection with a debt previously contracted, shall inure to and be recoverable by the issuer, irrespective of any intention on the part of such beneficial owner, director, or officer in entering into such transaction $${\ldots}$$” (15 U.S. Code § 78p). 13 I still do not include sales associated with option exercises. However, I find very similar results when I include sales associated with option exercises. 14 In the Online Appendix, I document weak results when tax considerations are strongest. Odean (1998) shows that tax-motivated selling is most evident in December; I find little difference in return predictability between sales at a loss in December and sales at a gain in December. 15 In contrast, Kallunki et al. (2009) look at data on Swedish insiders and show that insider selling is most informative for Swedish insiders who have the greatest proportion of wealth allocated to insider stocks. They explain this finding by noting that economic incentives are strongest for these insiders. 16 Realization utility might still predict a difference in return predictability in a narrow band—the results from Section 4.6.2—as insiders could still be suffering a large loss (gain) relative to their overall wealth by realizing a small percentage loss (gain). Data limitations prevent thorough analysis of this point. 17 Jin and Scherbina (2011) make a similar distinction for mutual fund holdings. The authors compare the holdings of managers who take over a fund with the holdings of continuing fund managers. References Ali, U., and Hirshleifer. D. A. 2017 . Opportunism as a firm and managerial trait: Predicting insider trading profits and misconduct. Journal of Financial Economics 126 : 490 – 515 . Google Scholar CrossRef Search ADS Barberis, N., and Xiong. W. 2012 . Realization utility. Journal of Financial Economics 104 : 251 – 71 . Google Scholar CrossRef Search ADS Bettis, C., Vickrey, D. and Vickrey. D. 1997 . Mimickers of corporate insiders who make large-volume trades. Financial Analysts Journal 53 : 57 – 66 . Google Scholar CrossRef Search ADS Chang, T., Solomon, D. and Westerfield. M. 2016 . Looking for someone to blame: Delegation, cognitive dissonance, and the disposition effect. Journal of Finance 71 : 267 – 302 . Google Scholar CrossRef Search ADS Cheng, S., Nagar, V. and Rajan. M. 2007 . Insider trades and private information: The special case of delayed-disclosure trades. Review of Financial Studies 20 : 1833 – 64 . Google Scholar CrossRef Search ADS Cheong, F. S., and Thomas. J. 2011 . Why do EPS forecast error and dispersion not vary with scale? Implications for analyst and managerial behavior. Journal of Accounting Research 49 : 359 – 401 . Google Scholar CrossRef Search ADS Cohen, L., Malloy, C. and Pomorski. L. 2012 . Decoding inside information. Journal of Finance 67 : 1009 – 43 . Google Scholar CrossRef Search ADS DellaVigna, S., and Pollet. J. 2009 . Investor inattention and Friday earnings announcements. Journal of Finance 64 : 709 – 49 . Google Scholar CrossRef Search ADS Frazzini, A. 2006 . The disposition effect and underreaction to news. Journal of Finance 61 : 2017 – 46 . Google Scholar CrossRef Search ADS Frydman, C., and Rangel. A. 2014 . Debiasing the disposition effect by reducing the saliency of information about a stock’s purchase price. Journal of Economic Behavior and Organization 107 : 541 – 52 . Google Scholar CrossRef Search ADS PubMed Frydman, C., Barberis, N. Camerer, C. Bossaerts, P. and Rangel. A. 2014 . Using neural data to test a theory of investor behavior: An application to realization utility. Journal of Finance 69 : 907 – 46 . Google Scholar CrossRef Search ADS PubMed Gao, G., Ma, Q. and Ng. D. 2016 . The sound of silence: What do we know when insiders do not trade? Working Paper . Genesove, D., and Mayer. C. 2001 . Loss aversion and seller behavior: Evidence from the housing market. Quarterly Journal of Economics 116 : 1233 – 60 . Google Scholar CrossRef Search ADS Gormley, T., and Matsa. D. 2014 . Common errors: How to (and not to) control for unobserved heterogeneity. Review of Finacial Studies 27 : 617 – 61 . Google Scholar CrossRef Search ADS Grinblatt, M., and Han. B. 2005 . Prospect theory, mental accounting, and momentum. Journal of Financial Economics 78 : 311 – 39 . Google Scholar CrossRef Search ADS Grinblatt, M., and Keloharju. M. 2001 . What makes investors trade? Journal of Finance 56 : 589 – 616 . Google Scholar CrossRef Search ADS Guimaraes, P., and Portugal. P. 2010 . A simple feasible procedure to fit models with high-dimensional fixed effects. Stata Journal 10 : 628 – 49 . Hartzmark, S., and Solomon. D. 2012 . Efficiency and the disposition effect in NFL prediction markets. Quarterly Journal of Finance 2 : 125 – 200 . Google Scholar CrossRef Search ADS Heath, C., Huddart, S. and Lang. M. 1999 . Psychological factors and stock option exercises. Quarterly Journal of Economics 114 : 601 – 27 . Google Scholar CrossRef Search ADS Heimer, R. 2016 . Peer pressure: Social interaction and the disposition effect. Review of Financial Studies 29 : 3177 – 209 . Google Scholar CrossRef Search ADS Huddart, S., and Lang. M. 2003 . Information distribution within firms: Evidence from stock option exercises. Journal of Accounting and Economics 34 : 3 – 31 . Google Scholar CrossRef Search ADS Ingersoll, J., and Jin. L. 2013 . Realization utility with reference-dependent preferences. Review of Financial Studies 26 : 723 – 67 . Google Scholar CrossRef Search ADS Jagolinzer, A. D. 2009 . SEC rule 10b5-1 and insiders’ strategic trade. Management Science 55 : 224 – 39 . Google Scholar CrossRef Search ADS Jeng, L., Metrick, A. and Zeckhauser. R. 2003 . Evidence of predictable behavior of security returns. Review of Economics and Statistics 85 : 453 – 71 . Google Scholar CrossRef Search ADS Jin, L., and Scherbina. A. 2011 . Inheriting losers. Review of Financial Studies 24 : 786 – 820 . Google Scholar CrossRef Search ADS Kallunki, J., Nillson, H. and Hellstrom. J. 2009 . Why do insiders trade? Evidence based on unique data on Swedish insiders. Journal of Accounting and Economics 48 : 37 – 53 . Google Scholar CrossRef Search ADS Lakonishok, J., and Lee. I. 2001 . Are insider trades informative? Review of Financial Studies 14 : 79 – 111 . Google Scholar CrossRef Search ADS Lin, J., and Howe. J. 1990 . Insider trading in the OTC market. Journal of Finance 45 : 1273 – 84 . Google Scholar CrossRef Search ADS Lorie, J., and Niederhoffer. V. 1968 . Predictive and statistical properties of insider trading. Journal of Law and Economics 11 : 35 – 53 . Google Scholar CrossRef Search ADS Marin, J., and Olivier. J. 2008 . The dog that did not bark: Insider trading and crashes. Journal of Finance 63 : 2429 – 76 . Google Scholar CrossRef Search ADS Massa, M., and Von Beschwitz. B. 2015 . Biased shorts: Stock market implications of short sellers’ disposition effect. Working Paper . Odean, T. 1998 . Are investors reluctant to realize their losses? Journal of Finance 53 : 1775 – 98 . Google Scholar CrossRef Search ADS Petersen, M. 2008 . Estimating standard errors in finance panel data sets: Comparing approaches. Review of Financial Studies 22 : 435 – 80 . Google Scholar CrossRef Search ADS Rozeff, M., and Zaman. M. 1988 . Market efficiency and insider trading: New evidence. Journal of Business 61 : 25 – 44 . Google Scholar CrossRef Search ADS Seyhun, H. N. 1998 . Investment intelligence from insider trading. Cambridge : MIT Press . Shefrin, H., and Statman. M. 1985 . The disposition to sell winners too early and ride losers too long: Theory and evidence. Journal of Finance 40 : 777 – 90 . Google Scholar CrossRef Search ADS Shumway, T. 1997 . The delisting bias in CRSP data. Journal of Finance 52 : 327 – 40 . Google Scholar CrossRef Search ADS Thaler, R. 1999 . Mental accounting matters. Journal of Behavioral Decision Making 12 : 183 – 206 . Google Scholar CrossRef Search ADS Tversky, A., and Kahneman. D. 1981 . The framing of decisions and the psychology of choice. Science 211 : 453 – 58 . Google Scholar CrossRef Search ADS PubMed Wang, W., Shin, Y. and Francis. B. 2012 . Are CFOs’ trades more informative than CEO’s trades? Journal of Financial and Quantitative Analysis 47 : 743 – 62 . Google Scholar CrossRef Search ADS Weber, M., and Camerer. C. 1998 . The disposition effect in securities trading: An experimental analysis. Journal of Economic Behavior and Organization 33 : 167 – 84 . Google Scholar CrossRef Search ADS © The Author(s) 2018. Published by Oxford University Press on behalf of The Society for Financial Studies. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)

### Journal

The Review of Financial StudiesOxford University Press

Published: Feb 7, 2018

## You’re reading a free preview. Subscribe to read the entire article.

### DeepDyve is your personal research library

It’s your single place to instantly
that matters to you.

over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month ### Explore the DeepDyve Library ### Search Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly ### Organize Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place. ### Access Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals. ### Your journals are on DeepDyve Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more. All the latest content is available, no embargo periods. DeepDyve ### Freelancer DeepDyve ### Pro Price FREE$49/month
\$360/year

Save searches from
PubMed

Create lists to

Export lists, citations