Emotional State and Market Behavior

Emotional State and Market Behavior Abstract We consider the relationship between trader emotions and asset market behavior. We create experimental asset markets with the structure first studied by Smith, Suchanek, and Williams (1988), which is known to generate price bubbles and crashes. To track traders’ emotions in real time, we analyze participants’ facial expressions with facereading software before and while the market is operating. We find that a positive emotional state correlates with purchases and overpricing. Fear correlates with selling, low prices, and price decreases. The experiment confirms the intuition that emotions and market dynamics are closely related. 1. Introduction The connection between asset market price movements and emotions has been widely accepted in popular press and commentary. The supposed existence of fear and exuberance as influences on prices is reaffirmed with great frequency in such quarters. Positive emotion is generally associated with booms and high price levels. Alan Greenspan, while chairman of the Federal Reserve, famously remarked that the American stock market exhibited an “irrational exuberance” when it experienced a rapid run up in 1996. The remark betrayed a belief on his part that the increase had, in part, an origin in positive emotions of traders.1Galbraith (1984) describes stock market price bubbles as “speculative euphoria”. On the other hand, fear is associated with anticipated price variability and cited as a force leading to selloffs and price declines. Market volatility indices such as the CBOE’s VIX, an index of option prices, are referred to colloquially as “fear” indices. The legendary investor Warren Buffett (2008) writes, “A simple rule dictates my buying: be fearful when others are greedy and be greedy when others are fearful”, associating the presence of fear in the market with opportunities to make profitable purchases. There is data supporting the contention that traders’ moods can lead to price movements at the market level. Hirshleifer and Shumway (2003) find that good weather is correlated with higher stock returns. They presume that the mechanism whereby this effect operates is through the positive effect that weather has on mood. Kamstra, Kramer, and Levi (2003) observe that returns are relatively low in the dark seasons of fall and winter and appeal to a similar intuition to explain their results. Sports scores seem to matter for financial returns (Edmans, Garcia, and Norli, 2007), with home team losses translating to lower prices. Bollen, Mao, and Zeng (2011) find that Twitter mood predicts subsequent stock market movements. Gilbert and Karahalios (2010) find that the level of anxiety of posts on the blog site Live Journal predicts price declines. In all of this literature, more positive (negative) emotional states are associated with higher (lower) prices. In this paper, we focus on the connection between trader emotions and extreme pricing episodes: asset price bubbles and crashes. We use an experimental approach, which exploits the fact that bubbles and crashes can be reliably created and studied in the laboratory. The bubble and crash pattern was first observed in the laboratory with a paradigm introduced by Smith, Suchanek, and Williams (1988). Subsequent authors have replicated and established the robustness of this price pattern, and the Smith, Suchanek, and Williams (1988) design has become the dominant experimental paradigm for studying bubbles and crashes. We adhere to this design in the work reported here, and it is described in Section 4. Bubbles can be eliminated in this setting when participants are inexperienced with the market environment, but it typically requires either a very strong framing that deemphasizes the importance of speculative possibilities or a considerable degree of specialized instruction (Lei and Vesely, 2009; Kirchler, Huber, and Stöckl, 2012). The magnitude of bubbles is sensitive to environmental parameters such as the amount of liquidity available (Caginalp, Porter, and Smith, 1998), institutional factors such as the ability to sell short and the trading process (Van Boening, Williams, and LaMaster, 1993; Haruvy and Noussair, 2006; Lugovskyy, Puzzello, and Tucker, 2011), and the time path of fundamentals (Noussair, Robin, and Ruffieux, 2001; Noussair and Powell, 2010; Kirchler, Huber, and Stöckl, 2012; Breaban and Noussair, 2015; Giusti, Jiang, and Xu, 2016). Nonetheless, there is considerable variation within all conditions that remains unexplained. That is, some sessions generate larger bubbles than others despite identical economic structure. We consider here whether variation in the emotional state of participants between different cohorts can account for some of this heterogeneity. In our experiment, we use facereading software to track the emotional state of all traders, as captured in their facial expressions. The software provides measures of happiness, surprise, anger, disgust, sadness, fear, neutrality, and overall emotional valence. According to Elster (1998), emotions can be differentiated from other mental states on the basis of six features: cognitive antecedents, intentional objects, arousal, valence, action tendencies, and physiological expressions. The work we report here focuses on the last feature, the physiological, as manifested in facial expressions. We consider several issues. First, at the market level, we study how emotional factors correlate with the magnitude of bubbles. We test the hypotheses that a positive emotional state on the part of traders before a market opens correlates with higher prices, and that greater fear correlates with lower prices. At the individual level, we consider whether emotions2 are linked to better performance, and explore the relationship between loss-averse decision-making and emotional state. We analyze how emotions affect the behavior of traders with propensities to use different strategies. To do so, while the market is operating, we track the bidirectional relationship between specific emotions and overall valence on the one hand, and individual decisions and market price movements on the other hand. Emotional valence is a psychological term referring to a composite measure of the positivity of emotional state. Emotions associated with positive states such as joy are said to have positive valence, while those that are adverse such as fear or anger are said to be negative in valence. The overall valence of an individual’s emotional state reflects the relative strength of the positive and negative components. While our hypotheses concern overall valence and fear, we also consider whether other emotions correlate with market activity. Other than one recent study on individual decision-making (Nguyen and Noussair, 2014) and one of ultimatum games (van Leeuwen et al., 2016), this paper represents, to our knowledge, the first application of facial recognition technology in either economics or finance. We find a number of strong relationships between emotions, as measured in traders’ facial expressions, and market behavior. Positive emotion is associated with higher prices and larger bubbles. The more positive the emotional valence a group of traders exhibits before the market opens, the higher prices are in the subsequent market. Individuals in more positive emotional states are more likely to make purchases, but the effect is specific to those who trade, irrationally, on momentum. Fear on the part of traders before a market opens is associated with lower prices. At the individual level, fear is associated with sales, and again, the link is specific to momentum traders. A greater average level of fear in the market presages a price decline. Those who exhibit more neutrality during a crash earn greater profits. We also observe a strong correlation between fear and loss aversion, as registered in a loss aversion measurement task administered before the market opens. In general, the experiments confirm the intuition that there is a close connection between market behavior and emotional state. 2. Previous Experimental Literature on Emotions and Markets Moods have been linked to behavior in a number of well-known experimental paradigms, and some of these involve markets. For example, positive moods can influence product choices (Meloy, 2000) and bidding in random nth price auctions (Capra, Lanier, and Meer, 2010). Johnson and Tversky (1983) argue that a positive mood tends to make beliefs more optimistic in the sense that probabilities associated with positive events become transformed in a positive direction. This effect would push individuals to make less risk-averse choices when they are in a more positive emotional state. This suggests one mechanism whereby emotional state could influence market behavior. Asset markets involve the trading of a risky lottery and thus less risk-averse agents would tend to place higher value on the asset, and their activity would lead to higher demand and prices. Indeed, Breaban and Noussair (2015) find that more risk-averse cohorts of traders tend to generate lower prices in experimental asset markets. Fellner and Maciejekovsky (2007) find that risk aversion on the part of a group of traders is associated with lower trading volume. Bosman and Riedl (2003) find that negative mood increases bidding in first-price sealed bid auctions, which is consistent with exhibiting more risk averse behavior. Loewenstein et al. (2001), surveying a large body of research in psychology, argue that a direct link exists between decision making under risk and emotional state. The specific emotion of fear has been associated with risk aversion in a number of studies. Lerner and Keltner (2001) find that induction of fear leads to pessimistic risk assessments and anger to optimistic ones. Since pessimistic risk assessments lead to more risk-averse decisions with respect to objective risks, fear correlates positively with risk aversion. Kugler, Connolly, and Ordonez (2012) obtain similar results in a different impersonal lottery-based task. Nguyen and Noussair (2014) also find that fear in facial expressions is positively correlated with risk averse choices. We are aware of three previous studies that explore the role of emotion in generating bubbles in experimental asset markets. All three papers consider markets with the structure of Smith, Suchanek, and Williams (1988), as we do here. Andrade, Lin, and Odean (2016) induce mood exogenously with film clips before the market opens. Subjects watch video clips that are (a) pleasantly exciting, (b) calm, and (c) fearful. They find that the pleasantly exciting video clips are associated with larger bubbles than the other three treatments. The other three conditions are not different from each other in terms of average asset prices. Lahav and Meer (2010) conduct an experiment with two treatments, which they call the positive and neutral treatments. Like Andrade, Lin, and Odean (2016) they induce mood by showing film clips to subjects before the market opens. Positive affect was induced with routines by comedian Jerry Seinfeld, and in the neutral treatment, no clip was shown. They find that the positive treatment is characterized by greater bubbles and higher prices than the neutral treatment, though the neutral treatment nonetheless generated price bubbles. Hargreaves-Heap and Zizzo (2011) conduct an experiment in which emotions are tracked over the course of the session. They focus on anger, anxiety, excitement, and joy. They have four conditions. In all conditions, subjects participate in two asset markets. In two of the treatments, individuals rate, on a Likert scale from 1 to 7, how intensely they currently feel each of the four emotions. In one of these conditions, subjects can chat with each other, and in the other they cannot chat. Hargreaves-Heap and Zizzo report that eliciting emotions does not in itself have an effect on market prices, but they do find that the level of excitement reported is positively correlated with price level. They also find that buying assets is linked to excitement and selling assets is connected to anxiety. They do not find a correlation between emotional state and trading profits. 3. Theoretical Model In this section, we construct a theoretical model that serves as our source of hypotheses for the experiment. Our model draws on elements from behavioral finance and psychology. Specifically, it combines the assumptions of heterogeneous trader types of the model of asset bubbles and crashes of DeLong et al. (1990), with two prominent psychological theories concerning the relationship between emotions and decision-making under risk. These are the Affective Generalization Hypothesis (Johnson and Tversky, 1983), and the Appraisal Tendency Framework (ATF) (Lerner and Keltner, 2001; Han, Lerner, and Keltner, 2007). Both of these frameworks have strong empirical support in prior experimental work. The model of DeLong et al. (1990) is our starting point for a number of reasons. The first is that it generates bubbles and crashes, the most prominent qualitative feature in our markets. The second is that it distinguishes between three types of trader that are reasonable descriptions of how most traders behave in experimental markets. The heterogeneity of types in this model makes it possible to consider which specific types of trader are affected by changes in emotional state. The third is that the model has been successfully applied to data from experimental markets with a similar structure to ours by Haruvy and Noussair (2006) and Haruvy, Noussair, and Powell (2014). The model of DeLong et al. (1990) posits three interacting trader types: (1) Fundamental Value traders, (2) Rational Speculators, and (3) Momentum traders. The demand of fundamental value traders depends only on the current price and fundamental value. The rational speculator’s demand is a function of the current price and the individual’s beliefs about future prices. The momentum trader, who is irrational, follows previous trends. We augment the model by allowing emotional state to influence the demand of each type. To relate emotional valence to trader behavior, our model assumes the Affective Generalization Hypothesis (Johnson and Tversky, 1983), which concerns the role of overall affect in judgments of risk. The hypothesis asserts that unrelated experiences that provoke certain type of emotions induce changes in risk perception. In their original study, Johnson and Tversky (1983) show that an induced mood has a large and pervasive impact on perceived risk, with a strong generalization effect for both positive and negative emotions. Positive affect decreases assessments of the likelihood of unfavorable outcomes and negative affect increases their perceived likelihood. In other words, people tend to overestimate the frequency of unfavorable risky events when they are in a negative mood, even if their emotional state is completely independent of the events themselves, and they tend to underestimate the frequency of such events when in a positive mood. A number of studies have confirmed these initial findings, and shown that risk taking increases when one is in a more positive emotional state and decreases in negative emotional states (Isen and Geva, 1987; Arkes et al., 1988; Wright and Bower, 1992; Nygren et al., 1996; Capra, 2004; Fehr-Duda et al., 2011). This is true regardless of whether the emotional state is integral, related to present judgment and choices, or incidental, independent of the current decision situation (Forgas, 1995). The asset trading in our market represents a lottery because of uncertain future dividends and resale prices. Thus, the affective generalization hypothesis would suggest a positive correlation between emotional valence and beliefs about perceived expected future dividends and resale prices. To model the effect of the specific emotion of fear, we apply the Appraisal Tendency Framework (Lerner and Keltner, 2001), which builds on Cognitive Appraisal Theory (Lazarus, 1991). The ATF predicts certain relationships between fear and decision making under risk. Under this framework, emotions arise from appraisals (or evaluations) of observed events. These then lead to action tendencies, which are propensities to behave in particular ways. According to the ATF, a state of fear induces an appraisal of future events as unpredictable and under situational control. This translates into an action tendency toward the avoidance of future risk (Lerner and Keltner, 2001; Han, Lerner, and Keltner, 2007) by individuals experiencing incidental fear.3 For simplicity, in our model, we will assume that the negative component of valence consists entirely of fear, so that an increase in fear will operate similarly to a decrease in valence. This is consistent with the popular intuition that negative sentiment in a market is typically better described as fear rather than as sadness, anger, disgust, or other negative emotional states. Suppose that there is an asset, with a finite life of T periods, trading in a competitive market populated by n traders. At the end of each period, the asset pays a dividend that is independently drawn each period from a stationary distribution, which can take on a finite number M of possible values. After the last dividend is paid, the asset has no value. Let d1t,…,dMt be the set of possible dividends in period t. Because the distribution of possible dividends is stationary, we can write d1t,…,dMt as d1,…,dM. Let πm equal the probability that the realized dividend equals dm. Then ∑mπmdm is the expected dividend the asset pays in any one period. Because the dividends constitute the only intrinsic source of value for the asset, the fundamental value of the asset in period t is given by ft=∑tT∑m=1Mπm dmt. The time period t can be interpreted in several different ways. One natural way to do so, followed by the literature in experimental finance, is to take each trading round in the experiment as a period. Under this interpretation, the asset pays a dividend at the end of each period. The time period can also be interpreted as a smaller unit, say a 10-s or a 30-s interval. Because of the market mechanism we employ in the experiment, the continuous double auction process, these short intervals will typically include one or more trades, and prices and demand can be readily defined over short intervals. Similarly, the fundamental value and the distribution of dividends can also be defined over short blocks of time. Under this latter interpretation, a dividend is only paid in a subset of periods. In our model, the market activity in one period will affect emotional state in the next, and emotions in the current period affect current actions. One can think of this feedback loop as repeated over a few seconds or longer intervals. Consider now trader i, who is of the Fundamental Value trader type. If unaffected by emotional state, he buys and sells based on prices and fundamentals, making purchases if prices are lower than fundamentals and sales if prices exceed them. Thus, the demand of fundamental value trader i at time t is given by:   DFVit(pt,ft),  with ∂DFVit∂ft>0, ∂DFVit∂pt<0. Let us now suppose that the trader’s assessment of the fundamental value is colored by his emotional state. Specifically, we assume that the Affective Generalization Hypothesis is operative so that more positive valence induces more optimistic risk assessments, while negative valence leads to more pessimistic ones. Let wmdm,V be the belief of a fundamental value trader, with valence level V∈-1,1 about the probability that the dividend will be equal to dm in the current period. Higher values of the variables V indicate more positive valence. We interpret fear as the emotion generating negative valence, so that fear φ  =min⁡0,V. The Affective Generalization Hypothesis is that risk assessments become more optimistic when an individual is in an emotional state with more positive valence. Thus, we assume that i’s conjectured distribution of future dividends Fiwmdm,Vi first-order stochastically dominates Fi'wmdm,Vi', if Vi> Vi'. In other words,   Vi>Vi'=> ProbFi'≤x|dm,Vi'≥ProbFi≤x|dm,Vi, ∀ x. (2) Denote trader i’s belief about the expected future dividend stream, by fit˜Vi. Then, the demand at time t of a fundamental value trader who may be affected by her emotional state is given by:   DFVit(pt,fit˜(Vit)), with ∂DFVit∂pt <0, ∂DFVit∂ft˜>0, ∂DFVit∂Vit>0. (3) Consider now the determinants of the valence of the fundamental value trader, Vit. We assume that a trader’s emotional state at time t depends on activity in period t – 1. We assume that the valence for a fundamental value trader is increasing in her perceived value of her current financial position. We assume that there is a lag in the emotional response to market activity, so that valence in period t depends on activity in period t − 1. Specifically, letting cit and qit be the cash and quantity holdings of trader i at time t,   Vitci,t-1,Vit-1,qi,t-1= Vitci,t-1+∑t-1T∑mwmdm,Vit-1*qi,t-1 . (4) Valence, therefore, is a function of one’s current wealth, evaluated at fundamental prices, and taking into account any distortions in beliefs due to emotional state. We assume that greater cash, asset holdings, and expected future dividend payments increase valence, so that ∂Vit∂ci,t-1 >0, ∂Vit∂qi,t-1 >0, and ∂Vit∂fi,t-1˜  > 0. Low cash, low asset holdings, and pessimistic beliefs about future dividends lower emotional valence. Now consider trader j, who is a rational speculator. This trader seeks to purchase for resale. The demand of rational speculator j at time t, DRjt, is a function of the prices in the current period and the expected best future resale price net of the fundamental. The demand takes the form:    DRjtpt,sjt, with ∂DRjt∂pt<0, ∂DRjt∂sjt >0, (5) where sjt is the trader’s belief about the maximum possible future resale price net of fundamental value. For rational speculators, higher current prices lead to lower quantity demanded and the prospect of a higher resale price leads to greater quantity demanded. In the absence of any influence of emotions,   sjt=Emaxt+k∈t+1,…,Tpt+k-ft+k. (6) However, we allow valence to influence the trader’s risk assessment. For the rational speculator, we do so by permitting emotions to affect her belief about the peak price relative to fundamentals in the future. More positive valence leads to more optimistic beliefs, so that if Vjt>Vjt', then sjtVjt>sjtVjt'. This implies that the demand of a rational speculator who may be affected by her emotional state will increase with Vjt, holding all else constant. The valence of a Rational Speculator is determined by the perceived maximal liquidation value of his financial position given his expectations of the future peak price relative to fundamentals.   Vjtcj,t-1, sj,t-1,qj,t-1= Vjtcj,t-1+sj,t-1qj,t-1 , (7) with Vjtcj,t-1, sj,t-1,qj,t-1 increasing in each of its three arguments. Finally, the demand of momentum trader k at time t is given by:   DMk tpt-1-pt-2. (8) Momentum traders use the previous price trend as a basis for their demand. Following DeLong et al. (1990) and Haruvy and Noussair (2006), we model this as a dependence on the most recent price change, pt-1-pt-2. We assume that emotional valence and fear influence the behavior of momentum traders because they act as filters, which translate the data into a propensity to make purchases and sales, resulting in demand of the form:   DMk tVktpt-1-pt-2, (9) with ∂DMk t∂Vk>0, so that more positive valence correlates with higher demand. Thus, aggregate demand is:   Dt(pt,.)=∑iDFVi t(pt,fit˜(Vit)) + ∑jDRjt(pt,sjt(Vjt))+ ∑kDMk t(Vkt(pt−1−pt−2))  (10) with ∂Dt∂pt <0,∂Dt∂pt-1 -∂Dt∂pt-2 >0, and∂Dt∂Vlt >0 for any trader l.(11) The total quantity demanded is decreasing in current price and increasing in the magnitude of the recent trend. An increase in the emotional valence of all traders at time t from Vt to Vt' induces an increase in quantity demanded on the part of all three types and thus on aggregate quantity demanded at time t. An increase in the valence of any individual also increases her own demand at time t. Thus, an increase in the valence of an individual trader of any of the three types also induces an increase in market demand. Let Q be equal to the total number of units of asset available. Because the supply of asset is fixed, the market clearing price is given by:   pt*=D-1,t(Q), with∂pt∂Vlt >0. (12) An increase in demand at time t induces an increase in the market clearing price at time t. This implies that an increase in the valence of all traders, or of one trader l holding all else fixed, results in higher prices. Similarly, an increase in the fear of all traders, or of one trader holding all else fixed, leads to lower prices. The effects of changes in valence are illustrated in Figure 1. The upper panel in the figure graphs market demand as a function of market price, and shows that it is shifted outward by an increase in the positivity of valence and thus inward by an increase in fear on the part of any or all traders. The bottom three panels trace the change in demand and the market clearing price resulting from an increase in valence of one or more traders of each of the three types. We assume that an increase in fear operates in an opposite manner as an increase in valence for each type, since it comprises the negative component of a trader’s valence. Figure 1 View largeDownload slide Response of aggregate market demand and demand of each of the three trader types to changes in valence and in fear. This figure depicts changes in the demand for the three different trader types as well as the aggregate demand when there is an increase in valence or fear. In the upper panel, the effect of an increase in the net positivity of emotional valence or of fear is depicted. The lower panels illustrate the effect of an increase in the net positivity of emotional valence on the demand of each type of trader, Fundamental Value, Momentum, and Rational Speculator, respectively. More positive valence shifts demand to the right for each type of trader. Figure 1 View largeDownload slide Response of aggregate market demand and demand of each of the three trader types to changes in valence and in fear. This figure depicts changes in the demand for the three different trader types as well as the aggregate demand when there is an increase in valence or fear. In the upper panel, the effect of an increase in the net positivity of emotional valence or of fear is depicted. The lower panels illustrate the effect of an increase in the net positivity of emotional valence on the demand of each type of trader, Fundamental Value, Momentum, and Rational Speculator, respectively. More positive valence shifts demand to the right for each type of trader. Now consider the special case of initial demand at t = 0 before the market opens for period 1. The demand of Fundamental Value trader i is based on her initial expectation of the future dividend stream and her emotional state, fi0˜(Vi0). The demand of Rational Speculator j depends on sj0Vj0 and the Momentum trader’s demand reduces to a function of Vk0, with higher valence corresponding to greater demand. It follows that more positive valence on the part of any one or all traders before the market starts correlates with a higher initial market price. The average valence in the market at time t > 0 is a result of the state of the market at t – 1, and is a weighted average of the valence of each of the three types, where the weighting is the percentage of traders of each type in the market. It depends positively on the valence in the preceding period, the cash and units held by fundamental value traders and rational speculators, the future expectations of speculators, and the recent market price trend. The average valence in period t is given by:   Vt=[∑iVit(cit,Vit−1, qit) + ∑jVjt(cjt, Vjt−1,qjt) +∑k(Vkt(pt−1−pt−2))]/n, (13) where n is the total number of traders, and dVtdcit>0,dVtdqit>0,dVtdsjt>0,dVtdcjt>0,dVtdqjt>0,dVtdVit-1>0, dVtdVjt-1>0, and dVtd[pt-1-pt-2]>0. Valence evolves over time in the following manner. More positive expectations about the peak prices relative to fundamentals increase average valence, because of their effect on Rational Speculators. An increasing price trend over the last two periods also makes average valence more positive through its impact on Momentum traders. For Fundamental Value traders and Rational Speculators, increases in the cash and quantity of units that they hold increase their valence. All of these factors will tend to increase prices and encourage bubbles. Fundamental Value traders or Rational Speculators running low on cash or units, a decrease in future expected prices, or a more negative price trend, all reduce valence. A summary of the theoretical predictions is given in Table I. Table I Summary of the theoretical predictions Symbol (+) indicates that a positive relationship is predicted between the variable listed in each row and valence. The upper panel indicates changes induced by an increase in valence. The lower panel indicates the changes in valence induced by a change in the market variables. The effects of and on each of the three trader types are shown separately.   FV Traders  RS Traders  M Traders  Average aggregate    Effects of an increase in valence  Beliefs dividend distribution  +        Beliefs future peak price    +      Demand  +  +  +  +  Prices        +  Effects of market variables on valence  Cash holdings  +  +      Quantity held  +  +      Price trend      +    Beliefs dividend distribution  +        Beliefs future peak price    +        FV Traders  RS Traders  M Traders  Average aggregate    Effects of an increase in valence  Beliefs dividend distribution  +        Beliefs future peak price    +      Demand  +  +  +  +  Prices        +  Effects of market variables on valence  Cash holdings  +  +      Quantity held  +  +      Price trend      +    Beliefs dividend distribution  +        Beliefs future peak price    +      4. The Experiment 4.1 Experimental Design The structure of the market was based on the paradigm created and studied in Smith, Suchanek, and Williams (1988). The asset that was exchanged in the market had a finite lifespan of T periods. At the end of each period t∈{1,…,T}, each unit of the asset paid a dividend dt that was independently drawn from a distribution that was identical for all periods. In any period t the expected dividend E(dt) on a unit of the asset was equal to the expected value of the dividend distribution. Dividends were drawn independently in each period. Therefore, the expected future dividend stream at time t, E[∑tTdt], equaled the expected period dividend multiplied by the number of periods remaining in the life of the asset. In other words, E[∑tTdt] = (T–t+ 1)E(dt). Since dividends were the only source of intrinsic value for the asset, the fundamental value ft had a particularly simple structure. It was equal, at any time t, to the expected future dividend stream from time t onward. In other words, ft = (T – t + 1)E(dt). In our markets, the life of the asset was T = 15, and the dividend was dt ϵ{0, 8, 28, 60}, where each realization was equally likely, for all t. Thus, E(dt) = 24, and ft = 24(16 – t) = 384 – 24*t at time t. The dividend distribution had a standard deviation of 27 per period, which was greater than the expected dividend. Therefore, risk-averse traders could value the asset at considerably less than its fundamental value. In each period, each trader had the ability to trade units of the asset for cash with any other trader in an open market, provided that he always maintained non-negative cash and share balances. Transaction prices were determined in a continuous double-auction market (Smith, 1962). This type of market operates in the following manner. Each period, the market is open for a fixed time interval, which was 2 min in this experiment. At any time while the market is open, any trader can submit an offer to sell or to purchase a share. These offers are posted publicly on all traders’ computer screens. Also at any time, any trader can accept an offer that another trader has submitted. When a bid or ask is accepted by a trader, a transaction for one share takes place between the trader who posted the offer and the trader who accepted it, at the offered price. Thus, within a period, it was possible for different transactions to occur at different prices. An individual could trade as much as he wished provided he had sufficient cash and units of the asset to complete the trades. Each subject had an identical portfolio, consisting of an initial endowment of 5 units of asset, and 5,000 units of experimental currency, at the beginning of period 1. A subject’s final earnings in the market were equal to the cash he had at the end of the experiment, which corresponded to his initial cash, plus the value of dividends received, plus (minus) any profit (loss) from trading. The market was computerized and used the Ztree program developed at the University of Zurich (Fischbacher, 2007). Prior to the opening of the asset market, we administered the loss aversion measurement task used by Trautmann and Vlahu (2011), which is based on an earlier protocol of Fehr and Goette (2007). This task consisted of a series of six choices, presented in a price list format. Each choice offered the opportunity to play a gamble which paid 4.5 Euros with probability 0.5 and either −0.5, −1.5, −2.5, −3.5, −4.5, or −5.5 Euros with probability 0.5, with each choice appearing exactly once. Subjects were required to indicate whether or not they accepted to play each of the six gambles. The number of gambles one decided not to play is interpreted as a measure of her loss aversion. Subjects completed the loss aversion measurement task using pen and paper. They submitted all six of their decisions simultaneously when they turned in their completed sheet of paper to the experimenter. They were informed prior to beginning the task that only one of the decisions would count toward their earnings. After all decisions were turned in, a die was rolled. The outcome of the roll determined which gamble would count for all participants. If a subject had chosen not to play the relevant gamble, she received a payoff of zero for this part of the experiment. If a participant chose to play the selected gamble, a coin was flipped to determine whether she received 4.5 Euro or the negative payment specified in the gamble.4 A separate coin was flipped for each participant who chose to play the gamble. 4.2 The Facereader Software During the sessions, all subjects were videotaped and the videotapes were analyzed later with Noldus Facereader. The taping began at least 30 s before the opening of the market for the first period and recorded continuously until the session ended. This is the first study to employ face reading in experimental finance. In our opinion, face reading is especially well-suited to the study of emotions for several reasons. The first reason is that it classifies an individual’s physiological state along emotional dimensions in a quantitative manner. This allows us, for example, to claim that one stimulus provokes more disgust but less sadness than another, or that a particular decision is taken when an individual is surprised rather than angry. A second advantage is that it registers emotional measurement in a manner that is completely unobtrusive to the participant, and data acquisition would proceed unnoticed if the individual were not informed that it was occurring.5 The third reason is that the facial expressions corresponding to the six basic emotions appear to be universal (Ekman and Friesen, 1986). It has been claimed that these expressions accompanying these six emotions are common to all cultures and primates (Ekman, 2007), and that they are the same for blind and sighted individuals (Matsumoto and Willingham, 2009), which provides strong evidence that they are innate. This means that results of studies such as ours should be replicable in different population groups and cultures. Happiness is positive in valence, surprise is neutral, and the other four basic emotions are negative. Happiness and anger are approach emotions, which tend to lead an individual to move toward the situation that triggers the emotion. Sadness, disgust, and fear, are withdrawal emotions, meaning that an individual typically seeks to avoid the stimulus that induces these emotions.6 4.3 Structure of the Data Our dataset consists of thirteen sessions. The sessions were conducted at Tilburg University and all subjects were students at the university. Subjects were recruited via an online system. No subject participated in more than one session of the experiment. On average, the sessions lasted 1 h. Between six and eleven traders participated in each session, with an average of eight subjects per session. Participants’ earnings from the asset market were converted to Euro at a rate of 500 units of experimental currency to 1 Euro. This resulted in an average payment of 15.6 euros (including the loss aversion measurement task). The market data consist of submitted bids and asks, as well as transactions, which are acceptances of bids and asks. We have market data for fifteen periods in each session. We also have the emotions data from all thirteen sessions for the 30 s before the market opened in period 1, and for the crash period, defined as the period within a session in which the greatest price decrease occurred. However, for a purely technical reason, we only have complete real time data for the totality of five of the thirteen sessions. Due to an improvement introduced in the video quality in late 2013, which increased the speed that the analysis could be conducted, in these five sessions it was possible to analyze all videos for the entire duration of the experiment. Therefore, for all fifty subjects participating in these sessions was possible to match their trading activity with their emotional responses for the entire session. This allows for an analysis of the dynamics between emotions and asset market behavior. The data are organized by two different lengths of time interval. The first is in 10 s intervals, which means that a session contains a total of 198 intervals. This consists of twelve intervals within each of the fifteen 2-min market periods, one interval after each period when the results for the period are displayed, and 30 s prior to the opening of the market for the first period. The data are in panel data format in which fifty subjects each form a panel, and each 10 s interval is an observation. The reason for specifying blocks of 10 s is that it is somewhat greater than the typical time course of emotional reactions. Emotions arise as a consequence of some events and last for a few seconds. There is little evidence in the literature on emotion duration, but Sonnemans and Frijda (1994) find that it depends to a great extent on the intensity of the emotion. Scherer, Walbott, and Summerfield (1986) find that different emotions tend to have different duration and they classify sadness as the most lasting one, followed by joy, anger, and fear. We expect that a 10 s interval is enough to capture both emotional and behavioral reactions to specific market events as well as short enough to capture the reaction to current activity only. Each emotion variable (happiness, fear, anger, disgust, sadness, surprise, neutral, and valence) was averaged every 10 s beginning with the market opening, so that the 300 observations that Facereader software provides in 10 s (30 per seconds) were averaged for each subject. We will refer to the unit of time in this data set as an interval. The second time scale on which we organized the data was in terms of the trading period, as it is typical in experimental work. We construct a data set that has observations for each of the fifty subjects during fifteen periods. In this case the emotion variables were averaged over the 120 s of each period and subject. In our presentation of the data, the 10-s intervals are indexed by τ = 1,…, 198, and periods by t = 1,…,15. Table II panels a and b show the mean and standard deviation of emotions between subjects, for both periods and 10-s intervals. Table II Mean and standard deviation of emotions (a) n = 10 020. Mean is calculated over each 10-s time interval. Valence ranges from [−1, 1]. Each of the seven specific emotions varies from [0, 1]. (b) n = 750. Mean is calculated over each 10-s time interval. Valence ranges from [−1, 1]. Each of the seven specific emotions varies from [0, 1]. a. Mean and standard deviation of emotions across subjects and 10-s intervals     Neutrality  Happiness  Sadness  Anger  Fear  Disgust  Surprise  Valence  Mean  0.524  0.092  0.083  0.038  0.001  0.012  0.021  −0.029  Standard deviation  0.305  0.153  0.128  0.091  0.007  0.054  0.067  0.231    a. Mean and standard deviation of emotions across subjects and 10-s intervals     Neutrality  Happiness  Sadness  Anger  Fear  Disgust  Surprise  Valence  Mean  0.524  0.092  0.083  0.038  0.001  0.012  0.021  −0.029  Standard deviation  0.305  0.153  0.128  0.091  0.007  0.054  0.067  0.231    b. Mean and standard deviation of emotions across subjects and periods     Neutrality  Happiness  Sadness  Anger  Fear  Disgust  Surprise  Valence  Mean  0.649  0.127  0.119  0.056  0.002  0.017  0.027  −0.030  Standard deviation  0.213  0.135  0.131  0.098  0.006  0.046  0.051  0.151    b. Mean and standard deviation of emotions across subjects and periods     Neutrality  Happiness  Sadness  Anger  Fear  Disgust  Surprise  Valence  Mean  0.649  0.127  0.119  0.056  0.002  0.017  0.027  −0.030  Standard deviation  0.213  0.135  0.131  0.098  0.006  0.046  0.051  0.151    In our analysis, we classify each of the traders participating in our experiment by trader types that correspond to the model in Section 3. The three types are (1) Fundamental Value Traders, (2) Momentum Traders, or (3) Rational Speculators. The traders are categorized according to the following criteria. We define an individual’s behavior as consistent with the Fundamental Value Trader type in period t if either one of two conditions holds. The first condition is that, if pt > ft, then qit < qi,t−1, where pt is the average price in period t, ft is the fundamental value in period t, and qit is the number of units of asset that individual i holds at the end of period t. This means that if prices are above fundamentals, trader i is a net seller of units in period t. The second condition is that if pt < ft, then qit > qi,t−1. If prices are below fundamentals, trader i is a net buyer in period t. The fundamental value trader, then, acts as if she is using the fundamental value as a limit price. A trader’s behavior is consistent with the Momentum Trader type if either of two conditions hold. The first is that, if pt−1 < pt−2, then qit < qi,t−1. The second is that, if pt−1 > pt−2, then qit > qi,t−1. The momentum trader is a net purchaser in period t if there has been an increasing price trend in the last two periods, and sells off units if there has been a decreasing trend. A trader’s behavior is consistent with the Rational Speculator Trader type if her behavior in period t satisfies one of the following two conditions. The first is that, if pt+1 < pt, then qit < qi,t−1, and the second is that, if pt+1 > pt, then qit > qi,t−1. This type of agent anticipates the price in the next period in an unbiased manner. She makes positive net purchases if the price is about to increase between the current and the next period. She makes net sales if the price is about to decrease. To classify a subject as one of the trader types, we count the number of periods during which a person is consistent with each type, and then classify him as the type with which he is consistent for the greatest number of periods. If there is a tie between two types, we classify the trader as belonging to each type with proportion 0.5. If there is a tie between all three types, he is assigned each type with proportion 0.33. Our model allows emotions to influence traders’ strategies. However, when we classify individuals into types, we assume that valence is not influencing trading strategies. This constraint can create some discrepancies between the model and the classification in that, for example, a fundamental value trader in a positive emotional state may purchase the asset at a price greater than objective fundamental value, while if emotions do not influence the decision, the trader would always sell if prices are greater than fundamentals. Thus, for purposes of the classification, we are in effect assuming that individuals are in a neutral emotional state. Some restriction of this type is unavoidable. In the absence of being able to directly estimate the parameters linking valence and demand, the classification we have conducted strikes us at the best way of categorizing individuals based on the variables we can observe. The percentage of individuals classified as each type is rather stable across sessions within this study, and relatively uncorrelated with the level if mispricing in the market, indicating that the influence of history on the market is not enough to shift the classification dramatically. The percentage classified as each type is also close to those observed by Haruvy and Noussair (2006) and Haruvy, Noussair, and Powell (2014), the other two studies that have reported a similar classification analysis, indicating that we have a subject pool that is typical in the distribution of strategies that it uses. The percentages classified as each type is indicated in Online Appendix III. 5. Hypotheses Based on our model and previous work, we advance several hypotheses about the relationships between emotions and market behavior. This first hypothesis follows from our theoretical model and is also suggested by the previous studies of Lahav and Meer (2010); Andrade, Lin, and Odean (2016); and Hargreaves-Heap and Zizzo (2012), who induce emotions exogenously prior to the market open. We hypothesize that the more positive the emotions that traders exhibit before a market opens, the greater the average price level in the market. Thus, we hypothesize that positive emotion is positively related to subsequent price, and thus in all likelihood within our setting, to greater bubbles. Hypothesis 1a: More positive emotional valence on the part of the average trader before the market opens correlates with higher subsequent prices and a larger bubble. To test this hypothesis, we check whether there is a correlation between (a) the average emotional valence within a group of traders in the 30 s before their market opens for period one, and (b) the average price in period 1, as well as the average over the fifteen periods the market is open. We also consider whether, as predicted by our model, fear correlates with lower prices. Andrade, Lin, and Odean (2016) fail to detect such an effect, and their attempt to induce fear generates similar results to a market in which emotions were not induced. However, Hargreaves-Heap and Zizzo do find that anxiety, a closely related emotional state, is correlated with lower prices. Hypothesis 1b: Greater fear on the part of the average trader before the market opens is correlated with lower subsequent prices. Our model predicts that positive valence on the part of individuals will correlate with higher demand on the part of these individuals. Therefore, we hypothesize that, at the individual level, a positive emotional state at any time on the part of an individual will correlate with her making greater net purchases of asset immediately afterward. Similarly, according to our model, fearful individuals exhibit lower demand for the asset. Moreover, the psychological implications of fear, being a withdrawal emotion, are that individuals will tend to avoid the situation that produces this emotion. Therefore, we anticipate a correlation between an individual’s level of fear and his likelihood of making a sale. Hypothesis 2a: Individuals in a more positive emotional state are more likely to make purchases in the subsequent 10-second time interval. Hypothesis 2b: Individuals who exhibit more fear are more likely to make sales in the subsequent 10 second time interval. As described in our model in Section 3, a trader’s demand might be affected by her emotional valence through different channels, depending on the type of trader she is. In the model, a fundamental value trader will tend to make more purchases when her beliefs about the future dividend stream are influenced by a positive emotional state. A rational speculator will purchase more units when, in a positive emotional state, she will overestimate the expected best future resale price net of the fundamental. A momentum trader will tend to buy or sell more units as positive valence or fear colors his interpretation of the past price trend. Therefore, we hypothesize that regardless the type of trader we are considering, their purchases and sales will correlate with the valence and fear they experience, respectively. We thus evaluate hypotheses 2a and 2b for each of the three types separately, as well as for the pooled data for all individuals. Our last hypotheses, 3a and 3b, concern the determinants of emotions. The model asserts that Fundamental Value traders’ emotional state is influenced by their cash and asset holdings and their beliefs about the fundamental value distribution. Speculators’ emotional valence is affected by their holdings of cash and asset, and their expectations of future trends. The observable variables among these are cash and asset holdings, and we hypothesize that the greater these are, holding all else constant, the more positive the valence of Fundamental Value traders and Rational Speculators. Lower holdings are associated with greater fear for these two types. Hypothesis 3a: Increases in cash and asset holdings correlate with more positive subsequent valence for fundamental value traders and rational speculators. The model also proposes that Momentum traders are influenced by the most recent price trend. The more positive the trend is, the more positive the valence of momentum traders, and the more negative the trend is, the greater is their fear. Hypothesis 3b: The more positive the recent trend of asset prices, pt-1 – pt-2, the greater the valence of momentum traders at time t. 6. Results 6.1 Market Price Patterns The time series of transaction prices in each of the thirteen sessions are shown in Figure 2, along with the time path of the fundamental value. In the figure, the vertical axis is in terms of experimental currency, and the horizontal axis indicates the market period. The black line is the time series of fundamental value and each grey line indicates an individual session. As can be seen in the figure, there are large differences between sessions, but in most sessions the bubble and crash pattern is observed. Typically, prices remain above fundamental values for a considerable period of time, and then exhibit a rapid fall toward fundamental value.7 Figure 2 View largeDownload slide Average transaction price, all periods, all markets. Figure 2 View largeDownload slide Average transaction price, all periods, all markets. 6.2 Emotions and Market Activity We now evaluate the hypotheses advanced in Section 5 which concern how valence and fear correlate with market price movements, and with the individual decisions of different trader types. The first two hypotheses, 1a and 1b, are about the relationship between the initial emotional profile of a cohort of traders and the overall price pattern they generate, and are summarized as results 1 and 2. Result 1. A more positive emotional state on the part of a trader cohort before the market opens is positively correlated with subsequent market price level Support for Result 1: We take the average valence that Facereader measures over the 30-s interval before the market opens for each subject. We then average it for all subjects in a session. Then, we correlate this average with the average transaction price over the course of the session.8Figure 3 plots the average initial group valence against the average price level over the fifteen-period life of the asset. The figure shows a clear positive relationship between emotional state and price. The Spearman correlation coefficient between the valence before the market opens and the average price level in a session is ρ = 0.6190 (p < .01). The correlation between average valence before the market opens and the price level in period 1 is 0.576 (p < .05).9 □ Figure 3 View largeDownload slide The relationship between cohort-average emotional valence prior to market open and average price level in a session. Each data point is the average valence within a trader cohort before the market opens for period 1. Figure 3 View largeDownload slide The relationship between cohort-average emotional valence prior to market open and average price level in a session. Each data point is the average valence within a trader cohort before the market opens for period 1. Result 2. Average trader fear before the market opens is negatively correlated with the subsequent price level in the market Support for Result 2: The relationship between the average fear that a cohort expresses before the market opens and the price level over the subsequent market is very pronounced. Figure 4 relates the fear that Facereader registers for the average trader before the market opens to the average price in the session. The figure shows a strong negative relationship between the two variables. The correlation is highly significant (ρ = −0.8333, p = 0.01). The correlation between average fear before the market opens and the average price in period 1 is negative at −0.428, though not significant (p = 0.144). □ Figure 4 View largeDownload slide The relationship between cohort-average fear prior to market open and average price level in a session. Each data point is the average fear within a trader cohort before the market opens for period 1. Figure 4 View largeDownload slide The relationship between cohort-average fear prior to market open and average price level in a session. Each data point is the average fear within a trader cohort before the market opens for period 1. The initial level of each of the other emotions considered separately also correlates with the subsequent average price level of the session, though not significantly. The other three negative emotions, sadness, anger, and disgust, correlate negatively with price level at ρ = −0.381, −0.428, and −0.333, respectively, while happiness and neutrality correlate positively with price level at ρ = 0.476 and 0.357. Although none of these correlations are significant, they are consistent with higher (lower) prices being associated with positive (negative) emotional states. Although results 1 and 2 concerned the relationship between initial emotional state and average market behavior, and show that hypothesis 1a and 1b are supported, result 3 reports a relationship between emotional state and price movements while the market is in progress. Result 3: The average fear in a cohort in period t−1 is positively correlated with a price decrease in period t Support for Result 3: We construct a dummy variable to identify the average price movement across the fifteen periods. The dummy takes value 1 if pt−pt−1 < 0, and 0 otherwise. We then run a logit model with subject-level fixed effects with the dummy variable as the dependent variable.10 Emotions at time t−1 are the independent variables in this model. Table III shows that with more fear on the part of traders, prices in the market are more likely to decrease. On the contrary, neutrality, happiness, and anger are negatively correlated with price decreases.11 The pattern seems to be in line with Lerner and Keltner’s (2001) findings that fear is associated with greater risk aversion, and happiness and anger are associated with risk seeking behavior. Overall valence is not significantly correlated with subsequent price movements. □ Table III Negative price movements as a function of prior average emotions within a trader cohort (dependent variable = 1, if pt < pt−1, and 0 otherwise) The estimated equation is ytz = β0 Constant + β1 Feart−1 + β2 Neutralityt−1 + β3 Happinesst−1 + β4 Angert−1 + β5 Disgustt−1 + β6 Sadnesst−1. ytz = 1 if the price decreases from period t – 1 to t in session z. Each emotion is averaged over time and across all traders in period t−1 of session z. Emotions range from [0, 1]. *p < 0.1, **p < 0.05, and ***p < 0.01.   Logit  Random effects  Fixed effects  Feart−1  402.26  354.45**  98.08  Neutralityt−1  −5.34  −6.35**  −14.27***  Happinesst−1  −5.03  −6.14**  −15.19***  Angert−1  −4.46  −5.40*  −14.30***  Disgustt−1  −4.91  −6.17  −20.25***  Sadnesst−1  −2.16  −2.96  −10.68**  Constant  5.40  6.46**      Logit  Random effects  Fixed effects  Feart−1  402.26  354.45**  98.08  Neutralityt−1  −5.34  −6.35**  −14.27***  Happinesst−1  −5.03  −6.14**  −15.19***  Angert−1  −4.46  −5.40*  −14.30***  Disgustt−1  −4.91  −6.17  −20.25***  Sadnesst−1  −2.16  −2.96  −10.68**  Constant  5.40  6.46**    We have seen that there is a positive relation between initial valence and price levels. One possible mechanism for sustaining high price levels might therefore be that positive valence on the part of individuals is associated with higher demand for the asset. Similarly, the relationship between fear and lower prices suggests that fear prompts a willingness to sell. These two relationships are expressed as our hypotheses 2a and 2b and are predicted by our model. As reported in results 4 and 5, both hypotheses are supported, though the timing is different than predicted for fear, which has its effect contemporaneously, rather than with a 10-s lag. Result 4: Traders with more positive valence at 10s time interval τ−1 make more purchases at time τ Support for Result 4: In order to determine how emotions interact with individual trading activity we run a Poisson count regression with subject-fixed effects, where the dependent variable is the number of units a subject has bought or sold during each 10s interval. Each transaction is coded at the time at which a bid or ask has been accepted. We analyze the influence of prior emotional state, controlling for financial position and price level, on the number of purchases and sales. We find that subjects make more purchases in the current interval, the higher the valence they exhibited in the previous interval. The left portion of Table IV, which includes the lagged value of purchases in Model 1, shows that more positive emotional valence Granger-causes purchases. Another intuitive result from the estimation is that the larger the number of units in their inventory, the less likely subjects are to make more purchases. □ Table IV Individuals’ purchases and sales at time interval τ as a function of emotions at τ−1 (Poisson count regression with subject-fixed effects) Estimates come from Poisson count regressions with the specification ylτ = β1 valenceτ−1(or fearτ−1) + β2ci,τ−1 + β3qi,τ−1 + β4pτ−1 + β5yi,τ−1. ylτ is the number of units bought or sold by trader i in 10-s time interval τ. *p < 0.1, **p < 0.05, and ***p < 0.01.   Buyτ  Buyτ    Sellτ  Sellτ  Model 1  Model 2  Model 3  Model 4  valenceτ−1  0.237*  0.238*  fearτ  4.995***  4.861***  moneyτ−1  7.29e−06  4.95e−06  moneyτ−1  3.42e−06  3.87e−06  unitsτ−1  −0.021**  −0.015  unitsτ−1  0.048***  0.044***  p levelτ−1  −0.00007  −0.00008  P levelτ−1  0.00012  −0.00012  Buyτ−1  0.355***    Sellτ−1  0.363***      Prob>chi2 = 0.000  Prob>chi2 = 0.0586    Prob>chi2 = 0.000  Prob>chi2 = 0.000    9,770 obs  9,770 obs    9,970 obs  9,970 obs    49 groups  49 groups    50 groups  50 groups    Buyτ  Buyτ    Sellτ  Sellτ  Model 1  Model 2  Model 3  Model 4  valenceτ−1  0.237*  0.238*  fearτ  4.995***  4.861***  moneyτ−1  7.29e−06  4.95e−06  moneyτ−1  3.42e−06  3.87e−06  unitsτ−1  −0.021**  −0.015  unitsτ−1  0.048***  0.044***  p levelτ−1  −0.00007  −0.00008  P levelτ−1  0.00012  −0.00012  Buyτ−1  0.355***    Sellτ−1  0.363***      Prob>chi2 = 0.000  Prob>chi2 = 0.0586    Prob>chi2 = 0.000  Prob>chi2 = 0.000    9,770 obs  9,770 obs    9,970 obs  9,970 obs    49 groups  49 groups    50 groups  50 groups  Result 5: Traders who are more fearful at time τ sell more units at time τ Support for Result 5: Controlling for the units and cash they have and how high average prices are compared with fundamentals, we do not find a significant effect of the lagged value of fear on current sales. This could be due to the fact that fear has a more immediate effect than some other emotions. To check this, we consider the contemporaneous level of individual fear as a correlate of sales, in Models 3 and 4 in Table IV. We find that fear is indeed related to larger reductions of the number of assets in inventory in a given period. It appears that positive emotional state has a relatively slow impact on purchases, while fear is contemporaneously correlated with sales. □ We now study the relationship between valence and purchases, as well as between fear and sales, for each trader type separately. Recall that we hypothesized that all trader types are swayed by emotion when making purchases and sales. Result 6 provides support for this assertion for Momentum traders only. Result 6: Momentum traders purchase more when they are in a more positive emotional state and sell more when they exhibit fear. There is no correlation between emotions and the number of purchases and sales of fundamental value traders or rational speculators Support for Result 6: As described in Section 3, we classify individuals according to their trading strategies, into Fundamental Value traders, Momentum traders, and Rational Speculators. We observe that 40% of our participants are Fundamental Value traders, 34% are Momentum traders, and 26% are Rational Speculators. We then analyze how emotions affect their trading behavior. Table V shows that the Momentum types, who are relatively unsophisticated and behave irrationally, both buy and sell based on their emotional state. They earn less money than the other two types. Momentum traders earn on average 6,233 experimental currency, while Fundamental Value traders and Rational Speculators earn 7,263 and 7,565, respectively (see also Haruvy and Noussair (2006) and Haruvy, Noussair, and Powell (2014). The behavior of Momentum traders accounts for the positive relationship between valence and subsequent purchases, as well as between fear and subsequent sales, documented in Table V. The coefficients are not significant for Fundamental Value and Rational Speculator types. □ Table V Individuals’ purchases and sales at time interval τ as a function of emotions at τ −1 for each trader type separately (Poisson count regression with subject-fixed effects) Estimates come from Poisson count regressions with the specification ylτ = β1 valenceτ−1 (or fearτ−1) + β2ci,τ−1 + β3qi,τ−1 + β4p τ−1 + β5yi,τ−1. ylτ is the number of units bought or sold by trader i in 10s time interval τ. *p < 0.1, **p < 0.05, and ***p < 0.01. Buyτ  Fundamental Value Trader  Momentum Trader  Rational Speculator Trader  Valenceτ−1  0.280  0.521**  −0.025  Moneyτ−1  0.00003  −0.00004  −0.00007**  Unitsτ−1  0.015  −0.119***  −0.068***  Price levelτ−1  −0.0004*  0.0007***  3.15e−06  Buyτ−1  0.435***  0.141*  0.370***    Obs: 3,762  Obs: 3,396  Obs: 2,612    Groups: 19  Groups: 17  Groups: 13    Prob>F = 0.0000  Prob>F = 0.0000  Prob>F = 0.0000    Buyτ  Fundamental Value Trader  Momentum Trader  Rational Speculator Trader  Valenceτ−1  0.280  0.521**  −0.025  Moneyτ−1  0.00003  −0.00004  −0.00007**  Unitsτ−1  0.015  −0.119***  −0.068***  Price levelτ−1  −0.0004*  0.0007***  3.15e−06  Buyτ−1  0.435***  0.141*  0.370***    Obs: 3,762  Obs: 3,396  Obs: 2,612    Groups: 19  Groups: 17  Groups: 13    Prob>F = 0.0000  Prob>F = 0.0000  Prob>F = 0.0000    Sellτ   Fundamental Value Trader  Momentum Trader  Rational Speculator Trader  Fearτ  4.601  4.641**  9.492  Moneyτ−1  −0.00005**  −0.00006*  0.00001***  Unitsτ−1  0.069***  0.027  0.077***  Price levelτ−1  0.0003  −0.0006**  0.0001  Sellτ−1  0.354***  0.294***  0.458***    Obs: 3,963  Obs: 3,396  Obs: 2,612    Groups: 20  Groups: 17  Groups: 13    Prob>F = 0.0000  Prob>F = 0.0000  Prob>F = 0.0000    Sellτ   Fundamental Value Trader  Momentum Trader  Rational Speculator Trader  Fearτ  4.601  4.641**  9.492  Moneyτ−1  −0.00005**  −0.00006*  0.00001***  Unitsτ−1  0.069***  0.027  0.077***  Price levelτ−1  0.0003  −0.0006**  0.0001  Sellτ−1  0.354***  0.294***  0.458***    Obs: 3,963  Obs: 3,396  Obs: 2,612    Groups: 20  Groups: 17  Groups: 13    Prob>F = 0.0000  Prob>F = 0.0000  Prob>F = 0.0000    The relationship between strong emotions and poor trading decisions seems to be in line with previous research that connects weaker emotions to higher earnings in asset markets. For example, Lo et al. (2005) and Lo and Repin (2002) find that emotional individuals achieve lower earnings as day traders. On the other hand, Coates (2012) documents how emotions are closely linked to effective stock trading, and a number of authors have argued that emotional responses can generally be beneficial for decision making (Damasio, 1994). However, our reading of the balance of the evidence from financial markets suggests that strong emotional responses would be correlated with unprofitable decisions. Of our three trader types, the Momentum traders are the least rational, and more influenced by their emotions. In previous studies it has been noted that relatively unsophisticated traders tend to accept offers made by other traders rather than submitting offers themselves (Plott et al., 1998). This suggests that it may be the Momentum traders who are accepting other traders’ offers to sell at high prices during the bubble. To consider this, we compute the total number of purchases relative to the total number of bids for each type of trader. A higher ratio indicates that subjects are less active submitting offers to the market; that is to say, their trades are being concluded by accepting other participants’ offers. For Rational Speculators the ratio is 0.49, for Fundamental Value traders it goes up to 0.60, and for Momentum traders the ratio is 0.72, which is consistent with the findings of Plott et al. (1998). The ratio of sales relative to asks is not significantly different between trader types: 0.45, 0.48, and 0.42, respectively. This means that it is the Momentum traders who tend to accept offers and make purchases at high prices in our overpriced markets, behavior that is typically suboptimal. We now consider the impact of market activity on emotions. As indicated in hypothesis 3a and 3b, our model assumes that the valence of both fundamental value traders and rational speculators is influenced by the amount of cash and the quantity of units they hold. The emotional state of momentum traders is not affected by their current holdings of cash or asset. Rather, the valence of momentum traders is affected by the recent price trend pt−1– pt−2. The estimates of valence as a function of current holdings of cash and asset, as well as of recent price trend, conducted for each trader type separately, are given in Table VI. The estimates are the basis of our next result. Table VI Valence as a function of current cash and asset holdings, as well as recent price trend, estimated separately for each trader type The dependent variable is the average valence of trader l over 10-s interval τ. The estimation is conducted separately for each of the three trader types. The independent variables are “money” ciτ and “value of units” pt * qiτ in the preceding 10-s time interval, and the price trend over the last two periods. *p < 0.1, **p < 0.05, and ***p < 0.01.   Valenceτ Fundamental Traders  Valenceτ Momentum Traders  Valenceτ Rational Speculators  const  −0.108***  −0.027  −0.082***  moneyτ−1  7.18e−06***  −1.56e−07  9.12e−06  Value of unitsτ−1  0.006***  −0.0009  0.0032*  pt−1−pt−2  −0.00026*  0.00034*  −0.0004**    Obs: 3,942  Obs: 3,379  Obs: 2,599    Groups: 20  Groups: 17  Groups: 13    Valenceτ Fundamental Traders  Valenceτ Momentum Traders  Valenceτ Rational Speculators  const  −0.108***  −0.027  −0.082***  moneyτ−1  7.18e−06***  −1.56e−07  9.12e−06  Value of unitsτ−1  0.006***  −0.0009  0.0032*  pt−1−pt−2  −0.00026*  0.00034*  −0.0004**    Obs: 3,942  Obs: 3,379  Obs: 2,599    Groups: 20  Groups: 17  Groups: 13  Result 7: Cash and asset holdings are positively correlated with emotional valence for Fundamental Value Traders and Rational Speculators, but not for Momentum Traders. The recent price trend is positively correlated with emotional valence for Momentum Traders, but not for the other two types. Hypotheses 3a and 3b are supported Support for Result 7: The table shows that hypotheses 3a and 3b are supported in the data. The variables Quantity Held (qit) and Cash Balance (cit) are significantly positive determinants of valence for both Fundamental Value traders and Rational Speculators, but not for Momentum Traders. The recent price trend is significantly positively related to valence for Momentum Traders, but takes negative and significant coefficients for the other two types. □ 6.3 Other Emotional Correlates of Behavior 6.3.a. Neutrality and the submission of offers The decision to submit offers to buy or to sell might depend on the emotional state of a trader. It seems reasonable to believe that if traders base their trading decisions at least partly on emotions, more neutrality on the part of traders would correlate with their being less active in the market. That is to say, more emotional agents would submit more offers to sell and to buy. In order to determine whether this is the case, we control for the overall financial position of an individual and the market prices and we investigate the relationship between emotions and submitting bids and asks. We estimate a Poisson count model where the dependent variable is the number of bids/asks/total number of orders that a subject has made in an interval of 10 s. Table VII shows that, at the individual level, more neutrality is associated with less initiation of orders, especially fewer bids. This seems to indicate that individuals who experience stronger emotions are more active in the market. Including the lagged value of the dependent variable in each model we obtain the result that neutrality Granger-causes the number of bids as well as the total number of bids and asks that a trader submits to the market. Table VII Number of bids and asks submitted by an individual at time τ as a function of her neutrality and market variables at τ – 1 Logit model with subject-fixed effects. Dependent variables are the number of units an individual offers to purchase (bids), the number of units an individual offers to sell (asks). Each unit of time is a 10-s intervals. *p < 0.1, **p < 0.05, and ***p < 0.01.   Bidsτ  Asksτ  Bids + Asksτ  Neutralityτ−1  −0.247**  −0.031  −0.126*  Moneyτ−1  −0.00008***  0.00005***  9.81e−06  Unitsτ−1  −0.058***  0.054***  0.020***  Price levelτ−1  −0.0002  0.0001*  −0.0001  Bidsτ−1  0.128***      Asksτ−1    0.097***    Bids+Asksτ−1      0.103***    Obs: 9,770  Obs: 9,971  Obs: 9,971    Groups: 49  Groups: 50  Groups: 50    Prob>F = 0.000  Prob>F = 0.000  Prob>F = 0.000    Bidsτ  Asksτ  Bids + Asksτ  Neutralityτ−1  −0.247**  −0.031  −0.126*  Moneyτ−1  −0.00008***  0.00005***  9.81e−06  Unitsτ−1  −0.058***  0.054***  0.020***  Price levelτ−1  −0.0002  0.0001*  −0.0001  Bidsτ−1  0.128***      Asksτ−1    0.097***    Bids+Asksτ−1      0.103***    Obs: 9,770  Obs: 9,971  Obs: 9,971    Groups: 49  Groups: 50  Groups: 50    Prob>F = 0.000  Prob>F = 0.000  Prob>F = 0.000  6.3.b. High prices and fear Table VIII reports estimates from a fixed effects regression, which shows that greater holdings of cash and units are related to a more positive emotional state, while higher price level is negatively correlated with valence. In the specification, the price level is defined as the difference between the average transaction price and the fundamental value of the asset at any time. The estimates confirm that given the emotional state in interval τ −1, more money and units are associated with higher valence in interval τ, and therefore valence is Granger-caused by these two variables. However, Table VIII also shows that price level takes a negative coefficient. Since the prices are nearly always above fundamentals, the price level variable here is nearly equivalent to a measure of mispricing. Our results seem to indicate that mispricing correlates negatively with valence, all else equal. This is presumably because some traders realize that the current price trajectory and thus the market value of their assets is not sustainable, prompting more negative emotions. As we have seen, traders with more positive valence make more purchases. Although our data are correlational, we could argue that the emotional process underlying the formation of a bubble might be that positive emotional state enhances purchases, but as prices increase and traders find themselves in an overpriced market, their emotions become less positive. Table VIII Emotional valence of individuals as a function of their financial position (subject-fixed effects) The dependent variable is the average valence of trader l over 10-s interval τ. The independent variables are “money” ci,τ−1, current quantity of “units” qi,τ−1, and the price pτ−1, during the preceding 10-s time interval. *p < 0.1, **p < 0.05, and ***p < 0.01.   Valenceτ  Valenceτ  moneyτ−1  2.51e−06*  4.01e−06**  unitsτ−1  0.0013*  0.0022***  p levelτ−1  −0.000044***  −0.000087***  valenceτ−1  0.480***    const.  −0.028**  −0.046***    Obs: 9,927  Obs: 9,970    Groups: 50  Groups: 50    Prob>F = 0.000  Prob>F = 0.000    Valenceτ  Valenceτ  moneyτ−1  2.51e−06*  4.01e−06**  unitsτ−1  0.0013*  0.0022***  p levelτ−1  −0.000044***  −0.000087***  valenceτ−1  0.480***    const.  −0.028**  −0.046***    Obs: 9,927  Obs: 9,970    Groups: 50  Groups: 50    Prob>F = 0.000  Prob>F = 0.000  Table IX shows the correlation between price level and valence, as well as with the individual emotions. The table shows that Spearman correlation test between price level and valence is negative and significant at 1% (ρ = −0.09). The correlation between price level, 0.395, and fear is significant at p < 0.01. Table IX Correlation between current price level and current emotional state of trader cohort Emotions are averaged for all traders in a given session. The unit of time is the market period t. *p < 0.1, **p < 0.05, and ***p < 0.01.   Fear  Valence  Happiness  Anger  Surprise  Disgust  Sadness  Neutrality  Price level  0.390***  −0.090***  −0.007  −0.070***  −0.267***  −0.033***  0.074***  −0.363***    Fear  Valence  Happiness  Anger  Surprise  Disgust  Sadness  Neutrality  Price level  0.390***  −0.090***  −0.007  −0.070***  −0.267***  −0.033***  0.074***  −0.363***  6.3.c. Crashes Of special interest are market crashes. These can be very large and are generally unanticipated by participants (Smith, Suchanek, and Williams, 1988; Haruvy, Lahav, and Noussair, 2007). The relationship between a crash episode and the dynamics of emotion is illustrated in Figure 5. The data are from sessions 8–13, markets with stark crashes, where the strength of the average level of several emotions that members of the session cohort exhibit over the periods just before, during, and after the crash, is plotted.12 These emotions are anger, fear, happiness, and surprise. They are normalized at the level observed in the period immediately preceding the crash. The data show a clear pattern. Sadness and anger exhibit modest increases during a crash as traders’ paper wealth declines. However, fear and surprise exhibit sharp increases, as uncertainty increases. By the time the crash ends, surprise has fallen sharply, and fear has stabilized at high levels. However, sadness and anger continue to increase, as traders apparently realize the extent of the losses that the crash has imposed. The figure illustrates the existence of a multi-faceted emotional reaction to a key market event and the ability of Facereader to coherently characterize this reaction. Figure 5 View largeDownload slide Normalized emotions in the period immediately before, during, and immediately after a crash. Crash period is defined as the period with the greatest price decrease in each session. “Before” and “after” indicate the periods immediately prior and subsequent to the crash period. Emotions are averaged among all individuals in a cohort of traders. Averaged emotions are normalized to a value of 1 in the period before the crash. Figure 5 View largeDownload slide Normalized emotions in the period immediately before, during, and immediately after a crash. Crash period is defined as the period with the greatest price decrease in each session. “Before” and “after” indicate the periods immediately prior and subsequent to the crash period. Emotions are averaged among all individuals in a cohort of traders. Averaged emotions are normalized to a value of 1 in the period before the crash. At the individual level, we analyze the relationship between the overall strength of emotions during a crash and trader profits. Lo and Repin (2005) and Lo et al. (2002) find that those who exhibit less volatility in their emotional state in the face of fluctuations in the market have better trading performance. In our experiment, the analogy to this pattern would be that the level of neutrality in one’s facial expression during a crash is correlated with greater trading profits. Figure 6 plots the relationship between the level of neutrality individuals exhibit during a crash period, and the final earnings an individual accrues over the entire fifteen-period market. The figure suggests that more neutrality during a crash is indeed correlated with better performance. Figure 6 View largeDownload slide Relationship between neutrality during crash period and total individual session earnings. Each data point corresponds to an individual participant. Neutrality is averaged across the entire crash period. Figure 6 View largeDownload slide Relationship between neutrality during crash period and total individual session earnings. Each data point corresponds to an individual participant. Neutrality is averaged across the entire crash period. The correlation, at the level of the individual, between her average neutrality during the crash period and her final earnings is 0.205 (p = 0.16). Neutrality correlates negatively with units held at the end of the crash period at ρ = −0.27 (p = 0.064). The other emotions do not correlate with the number of units held, and thus with the amount of unrealized capital losses, during a crash. The results are similar if the units held at the beginning of the crash period are considered (very few units are exchanged during a crash because of very low demand). 6.3.d. Loss aversion and emotions The last result describes a strong correlation between loss aversion and fear. The loss aversion protocol that was administered at the beginning of the sessions, and the measurement of the emotional profile of individuals before the market opens, permit an analysis of the correlation between loss aversion and the emotional state of participants at the individual level that is independent of any experience on the market. Our results show that those who make more loss-averse decisions exhibit more fear in their facial expressions, and have a more negative overall emotional state. There is no correlation between loss aversion and any other of the six basic emotions or with neutrality. Table X contains the correlations between the number of gambles declined in the loss aversion task and the average consistency of facial expressions with each of the six emotions that Facereader registers in the 30 s before the market opens. A greater number of gambles declined indicate greater loss aversion. The table shows that the correlation between fear and loss aversion, 0.3427, is positive and significant at the p < 0.05 level. The correlation between loss aversion and valence is negative (ρ = −0.3012, p < 0.05). In contrast, none of the correlations with other emotions are significant at even the 10% level. Table X The correlation between loss aversion and emotional state before experiment begins Number of observations: 55. Emotion data are from the 30 s time interval before the market opens for period 1. *p < 0.1, **p < 0.05, and ***p < 0.01.   Fear  Valence  Happiness  Anger  Surprise  Disgust  Sadness  Neutrality  Loss aversion  0.342***  −0.301**  −0.045  −0.068  −0.085  0.209  0.109  −0.198    (0.018)  (0.025)  (0.759)  (0.649)  (0.569)  (0.157)  (0.463)  (0.180)    Fear  Valence  Happiness  Anger  Surprise  Disgust  Sadness  Neutrality  Loss aversion  0.342***  −0.301**  −0.045  −0.068  −0.085  0.209  0.109  −0.198    (0.018)  (0.025)  (0.759)  (0.649)  (0.569)  (0.157)  (0.463)  (0.180)  7. Conclusion In this paper, we study the connection between emotions and asset market behavior. We find a number of patterns that conform to commonly expressed intuition about the link between emotions and asset prices. When traders are in a more positive emotional state at the time the market opens, subsequent asset prices are higher. When they exhibit more fear, subsequent prices are lower. Momentum traders’ purchases are correlated with positive emotional state and their sales are correlated with the fear they experience. Those who keep a relatively neutral emotional state during a crash earn greater profits. Greater fear on the part of the average participant is a harbinger of a price decrease. Thus, the data suggest a scientific basis for some popular notions regarding emotions and market behavior. Had we not found these correlations, it would suggest that the popular notions were biased or misguided, and our research agenda would have turned to the question of where such biases might originate. A number of factors have been shown to influence the incidence and magnitude of bubbles in the laboratory. These include the institutions of exchange; the time path of fundamentals; and the risk aversion, loss aversion, and cognitive ability of traders. The results reported here show that another factor can be added to the list; the emotional state of traders. Our findings build on similar results that have recently been reported (Lahav and Meer, 2010; Hargreaves-Heap and Zizzo, 2011; Andrade, Lin, and Odean, 2016), extending them to show that emotions and market behavior display interdependencies over time intervals as short as a few seconds. Overall, it is becoming clear that asset price bubbles in experimental markets are a complex phenomenon, subject to many determining influences, including emotional factors. We find a strong correlation between fear and loss aversion. Such a connection is, in our view, quite intuitive, though to our knowledge such a relationship has not been previously documented. Those who anticipate that they will have a more negative response to a financial loss exhibit more fear when placed in a situation in which losses are possible. In Section 3, we propose a mechanism whereby emotions influence market activity. Our model predicts correlations between emotional valence and fear on the one hand, and individual behavior and market outcomes on the other hand. The model provides for different ways that emotions can influence traders, based on the type of trading strategy that they tend to employ. The trading strategies are taken from the model of DeLong et al. (1990). We observe that emotions correlate with trading decisions only for Momentum Traders, who are irrational. The more rational Fundamental Value and Rational Speculator types do not exhibit a correlation between emotional state and trading activity. We also observed that, as predicted by our model, Fundamental Value traders and Rational Speculators are in a more positive emotional state when their holdings of cash and asset are greater, and that Momentum traders are in a more positive emotional state, the more positive the recent market price trend. This study is the first application of Facereading to experimental finance. This methodology had yielded what are, in our view, coherent results. Our perspective is that the strength of our results contributes to the validation of the methodology. In particular, we believe that Facereading has considerable potential for the study of markets. In starker experimental settings than the one studied here, the emotional response to specific events, such as to a price quote one has received, or to a specific transaction one has made or observed, can be isolated and studied. In particular, in future work, face reading can be used to study face-to-face market transactions. In such situations, facial cues are important sources of information about the intentions and emotional states of other parties to a potential transaction. In these settings, individuals may try to manipulate their facial expression as part of their strategy to obtain more favorable terms. Face reading technology is highly conducive to the study of such behavior. The finding that emotions correlate with market behavior has interesting policy implications. It shows that there is concrete basis for the idea that central banks, governments, firms, and the media, must take into account the effects of their announcements and releases of data on the emotional state of market participants and how this might in turn affect market prices. Such effects would arise in addition to the influence of the new information on fundamentals of the economy that might in turn affect asset price. While the effects of information on fundamentals can be well-understood with established techniques in finance and economics, studying the emotional component requires new tools. In our view, the methods described here constitute a step forward in this direction. Our results also demonstrate that a trader’s disposition to react emotionally to market events correlates with his trading performance. The traders whose behavior was irrational were more likely to be influenced by emotions, suggesting that fluctuations in emotional state are associated with poor trading decisions. This has implications for firms that employ traders. It shows that screening for emotional profiles might be valuable for hiring decisions. Furthermore, when making personal financial decisions, one would do well to consider one’s own emotionality. An emotional investor may be better off delegating her trading decisions to other, less emotional, decision makers. Supplementary Material Supplementary data are available at Review of Finance online. Footnotes 1 Our notion of positive emotional state is a short-term, as distinct from a longer-term, more stable state of well-being. Bernanke (2012) clearly articulates this distinction with regard to happiness. Happiness is a “short-term state of awareness that depends on a person’s perceptions of one’s immediate reality, as well as on immediate external circumstances and outcomes. By ‘life satisfaction’ I mean a longer-term state of contentment and well-being that results from a person’s experiences over time.” 2 By emotion, we refer to short-term affective states. This is a distinct, though related, notion to that of mood. See Capra (2004) for a discussion. While moods are affective states of relatively low-intensity, diffuse, enduring, and typically without a salient antecedent cause, emotions are more intense and short lived, and they usually have a proximate cause. 3 In the theory of emotion, a distinction between incidental and integral emotions is typically made. Incidental emotions are feelings that are normatively unrelated to the decision task at hand. Integral emotions, on the contrary, are generated by the activity current being undertaken or experienced. In this study, the initial emotional state before the market starts presumably primarily reflects incidental emotions. Emotions that we observe after the market activity has started are likely a different mix of integral and incidental emotions, with more weight on the integral. Both incidental and integral emotions fall within the purview of the Appraisal Tendency Framework (Cavanaugh et al., 2007). 4 Some subjects experienced monetary losses in this part of the experiment. However, they were informed that there would be subsequent activities in the session in which they could expect to earn money on average. No subject ended the session with negative final earnings, because income in the market phase of the experiment in all cases more than fully offset losses incurred in the loss aversion measurement task. 5 Subjects were told beforehand that they were being videotaped but not that their facial expressions were to be analyzed. They were informed that the videotapes would be viewed only by the two researchers conducting the studies. Although the study received the appropriate Institutional Review Board approval, in future projects we intend to debrief subjects after the experiment about the specific procedures to be followed regarding the analysis of the videotapes. 6 Online Appendix IV contains a detailed explanatory note on how the Facereader software operates. 7 The facial expression data exhibit several general characteristics. The first is that the valence is typically negative. This likely means that participation in the experiment yields disutility for participants compared with other activities. There is great volatility in emotional state even over short time intervals. This may reflect the large number and heterogeneity of events that one experiences in a period. There is no discernible decline in the overall strength of emotion over time, over the roughly 35-min period during which the asset market is in progress. Anger tends to be greater at the outset, possibly reflecting the fact that individuals who are concentrating tend to look like they are angry (see Zaman and Shrimpton-Smith, 2006), but within a few minutes it stabilizes. Valence reflects this pattern because it integrates this early anger, typically being very negative at the very beginning of a session but stabilizing at a moderately negative level for the rest of the session. 8 The same results would obtain if we used the average price difference from fundamentals pt – ft. This difference is referred to as the Bias in a market by Haruvy and Noussair (2006). 9 We test whether the heterogeneity of initial emotional states correlates with greater volume of trade over the session. However, the correlation between the variance of valence among participants before a session begins, and the volume of trade over the entire session, is 0.12, and is not significant at conventional levels. 10 We test the fixed effects versus random effects specifications for this estimation equation using the test of overidentifying restrictions and we reject the hypothesis that RE is consistent (p = 0.043). However, if we compare the estimates in Table III in the same manner as Rodriguez and Elo (2002), it seems to be the case that with random effects it is mainly the fear coefficient that differs by specification, while for some of the other coefficients the change is not as pronounced. This indicates that estimates from both specifications have some robustness. Rodriguez and Elo (2002) argue that such robustness is typically associated with the consistency of estimates. 11 It is interesting to mention here that emotion levels among traders correlate with the subsequent direction of price movements, but not their magnitude. Additional analysis shows that none of the emotions appear to have predictive power for how much prices will decrease. More precisely we constructed a variable that measured the magnitude of the price decrease as max{pt−pt−1, 0} and used it as a dependent variable in a regression with the lagged values of emotions as independent variables. None of the coefficients were significant. 12 The crash period is defined as the period with the largest decrease in prices compared with the previous period. References Andrade E., Lin S., Odean T. ( 2016) Bubbling with excitement: en experiment, Review of Finance  20, 447– 466. Google Scholar CrossRef Search ADS   Arkes H. R., Faust D., Guilmette T. J., Hart K. ( 1988) Eliminating the hindsight bias, Journal of Applied Psychology  73, 305– 307. Google Scholar CrossRef Search ADS   Bernanke B. ( 2012) Excerpt from Speech at the 32nd General Conference of the International Association of Income and Wealth on 6.8.2012, Cambridge, MA, USA. Bollen J., Mao H., Zeng X. ( 2011) Twitter mood predicts the stock market, Journal of Computational Science  2, 1– 8. Google Scholar CrossRef Search ADS   Bosman R., Riedl A. ( 2003) Emotions and shocks in a first-price sealed bid auction. Working paper, Maastricht University. Breaban A., Noussair C. N. ( 2015) Trader characteristics and fundamental value trajectories in an asset market experiment, Journal of Behavioral and Experimental Finance  8, 1– 17. Google Scholar CrossRef Search ADS   Buffett W. ( 2008) Buy American. I Am. The New York Times, October 16. Caginalp G., Porter D., Smith V. ( 1998) Initial cash/asset ratio and asset prices: an experimental study, Proceedings of the National Academy of Science of United States of America  95, 756– 761. Google Scholar CrossRef Search ADS   Capra M. ( 2004) Mood driven behavior in strategic interactions, American Economic Review  94, 387–372. Capra M., Lanier K., Meer S. ( 2010) The effects of induced mood on behavior in random nth price auctions, Journal of Economic Behavior and Organization  7, 223– 234. Google Scholar CrossRef Search ADS   Cavanaugh L. A., Bettman J. R., Luce M. F., Payne J. W. ( 2007) Appraising the appraisal-tendency framework, Journal of Consumer Psychology  17, 169– 173. Google Scholar CrossRef Search ADS   Coates J. ( 2012) The Hour between Dog and Wolf: Risk Taking, Gut Feelings and the Biology of Boom and Bust , Penguin Press, New York, NY, USA. Damasio A. ( 1994) Descartes’ Error: Emotion, Reason, and the Human Brain , Penguin Books, New York, NY, USA. DeLong, J. B., Shleifer, A., Summers, L. H., and Waldmann R. J. (1990) Positive feedback investment strategies and destabilizing rational speculation, Journal of Finance 45, 379–395. Edmans E., Garcia D., Norli O. ( 2007) Sports sentiment and stock returns, Journal of Finance  62, 1967– 1998. Google Scholar CrossRef Search ADS   Ekman P. ( 2007) Emotions Revealed, Recognizing Faces and Feelings to Improve Communication and Emotional Life , Henry Holt and Company, New York. Ekman P., Friesen W. V. ( 1986) A new pan-cultural expression of emotion, Motivation and Emotion  10, 159– 168. Google Scholar CrossRef Search ADS   Elster J. ( 1998) Emotions and economic theory, Journal of Economic Literature  36, 47– 74. Fehr, E. and Goette L. (2007) Do workers work more if wages are high? Evidence from a randomized field experiment, American Economic Review 97(1), 298–317. Fehr-Duda H., Epper T., Bruhin A., Schubert R. ( 2011) Risk and rationality: the effects of mood and decision rules on probability weighting, Journal of Economic Behavior and Organization  78, 14– 24. Google Scholar CrossRef Search ADS   Fellner G., Maciejekovsky B. ( 2007) Risk attitude and market behavior: evidence from experimental asset markets, Journal of Economic Psychology  28, 338– 350. Google Scholar CrossRef Search ADS   Fischbacher U. ( 2007) z-Tree: Zurich toolbox for ready-made economic experiments, Experimental Economics  10, 171– 178. Google Scholar CrossRef Search ADS   Forgas J. P. ( 1995) Mood and judgment: the affect infusion model (AIM), Psychological Bulletin  117, 39– 66. Google Scholar CrossRef Search ADS PubMed  Galbraith J. K. ( 1984) A Short History of Financial Euphoria , Penguin Books, UK. Gilbert E., Karahalios K. ( 2010) Widespread worry and the stock market, 4th International AAAI Conference on Weblogs and Social Media. Giusti G., Jiang J., Xu Y. ( 2016) Interest on cash, fundamental value process and bubble formation: an experimental study, Journal of Behavioral and Experimental Finance  11, 44– 51. Google Scholar CrossRef Search ADS   Han S., Lerner J. S., Keltner D. ( 2007) Feelings and consumer decision making: the appraisal-tendency framework, Journal of Consumer Psychology  17, 158– 168. Google Scholar CrossRef Search ADS   Hargreaves-Heap S., Zizzo D. ( 2011) Emotions and chat in a financial markets experiment. Working paper, University of East Anglia. Haruvy E., Lahav Y., Noussair C. N. ( 2007) Traders’ expectations in asset markets: experimental evidence, American Economic Review  97, 1901– 1920. Google Scholar CrossRef Search ADS   Haruvy E., Noussair C. N. ( 2006) The effect of short selling on bubbles and crashes in experimental spot asset markets, Journal of Finance  61, 1119– 1157. Google Scholar CrossRef Search ADS   Haruvy E., Noussair C. N., Powell O. ( 2014) The impact of asset repurchases and issues in an experimental market, Review of Finance  18, 681– 713. Google Scholar CrossRef Search ADS   Hirshleifer D., Shumway T. ( 2003) Good day sunshine: stock returns and the weather, Journal of Finance  3, 1009– 1032. Google Scholar CrossRef Search ADS   Isen A. M., Geva N. ( 1987) The influence of positive affect on acceptable level of risk: the person with a large canoe has a large worry, Organizational Behavior and Human Decision Processes  39, 145– 154. Google Scholar CrossRef Search ADS   Johnson E. J., Tversky A. ( 1983) Affect, generalization, and the perception of risk, Journal of Personality and Social Psychology  45, 20– 31. Google Scholar CrossRef Search ADS   Kamstra M. J., Kramer L. A., Levi M. D. ( 2003) Winter blues: a sad stock market cycle, American Economic Review  93, 324– 343. Google Scholar CrossRef Search ADS   Kirchler M., Huber J., Stöckl T. ( 2012) Thar she bursts—reducing confusion reduces bubbles, American Economic Review  102, 865– 883. Google Scholar CrossRef Search ADS   Kugler T., Connolly T., Ordonez L. D. ( 2012) Emotion, decision, and risk: betting on gambles versus betting on people, Journal of Behavioral Decision Making  25, 123– 145. Google Scholar CrossRef Search ADS   Lahav Y., Meer S. ( 2010) The effect of induced mood on prices in asset prices: experimental evidence. Working paper, Ben-Gurion University of the Negev. Lazarus R. S. ( 1991) Emotion and Adaptation , Oxford University Press, New York. Lei V., Vesely F. ( 2009) Market efficiency: evidence from a no-bubble asset market experiment, Pacific Economic Review  14, 246– 258. Google Scholar CrossRef Search ADS   Lerner J. S., Keltner D. ( 2001) Fear, anger and risk, Journal of Personality and Social Psychology  81, 146– 159. Google Scholar CrossRef Search ADS PubMed  Lewinski P., de Uyl T., Butler C. ( 2014) Automated facial coding: validation of basic emotions and FACS AUs in facereader, Journal of Neuroscience, Psychology and Economics  7, 227– 236. Google Scholar CrossRef Search ADS   Lo A., Repin D., Steenbarger B. ( 2005) Fear and greed in financial markets: a clinical study of day traders, American Economic Review  95, 352– 359. Google Scholar CrossRef Search ADS   Lo A., Repin D. ( 2002) The psychophysiology of real-time financial risk processing, Journal of Cognitive Neuroscience  14, 323– 339. Google Scholar CrossRef Search ADS PubMed  Loewenstein G., Weber E., Hsee C., Welch N. ( 2001) Risk as feelings, Psychological Bulletin  127, 267– 286. Google Scholar CrossRef Search ADS PubMed  Lugovskyy V., Puzzello D., Tucker S. ( 2011) An experimental study of bubble formation in asset markets using the tatonnement trading institution. Working paper 07/2011, Canterbury University. Matsumoto D., Willingham B. ( 2009) Spontaneous facial expressions of emotion of blind individuals, Journal of Personality and Social Psychology  96, 1– 10. Google Scholar CrossRef Search ADS PubMed  Meloy M. ( 2000) Mood‐driven distortion of product information, Journal of Consumer Research  27, 345– 359. Google Scholar CrossRef Search ADS   Nguyen Y., Noussair C. N. ( 2014) Risk aversion and fear, Pacific Economic Review  19, 296– 312. Google Scholar CrossRef Search ADS   Noussair C. N., Robin S., Ruffieux B. ( 2001) Price bubbles in laboratory asset markets with constant fundamental values, Experimental Economics  4, 87– 105. Google Scholar CrossRef Search ADS   Noussair C. N., Powell O. ( 2010) Peaks and valleys: price discovery in experimental asset markets with non-monotonic fundamentals, Journal of Economic Studies  37, 152– 180. Google Scholar CrossRef Search ADS   Nygren T. E., Isen A. M., Taylor P. J., Dulin J. ( 1996) The influence of positive affect on the decision rule in risk situations: focus on outcome (and especially avoidance of loss) rather than probability, Organizational Behavior and Human Decision Processes  66, 59– 72. Google Scholar CrossRef Search ADS   Plott C. R., Menshikov I., Menshikova O., Myagkov M. ( 1998) From non market attitudes to market behavior: laboratory market experiments in Moscow, and the Hvatat Property of Human Behavior. Social Science Working Paper, 987, California Institute of Technology, Pasadena, California. Rodriguez G., Elo I. ( 2002) Intra-class correlation in random-effects models for binary data, The Stata Journal  3, 32– 46. Scherer K. R., Walbott H. G., Summerfield A. B. ( 1986) Experiencing Emotions: A Cross-Cultural Study , Cambridge University Press, Cambridge, UK. Smith V. ( 1962) An experimental study of competitive market behavior, Journal of Political Economy  70, 111– 137. Google Scholar CrossRef Search ADS   Smith V., Suchanek G., Williams A. ( 1988) Bubbles, crashes, and endogenous expectations in experimental spot asset markets, Econometrica  56, 1119– 1151. Google Scholar CrossRef Search ADS   Sonnemans J., Frijda N. H. ( 1994) The structure of subjective emotional intensity, Cognition and Emotion  8, 329– 350. Google Scholar CrossRef Search ADS   Terzis V., Moridis C., Economides A. ( 2010) Measuring instant emotions during a self-assessment test: the use of FaceReader, Proceedings of Measuring Behavior, 192–195. Trautmann S., Vlahu R. ( 2011) Strategic loan defaults and coordination: an experimental analysis. Working paper, Dutch National Bank. Van Boening M., Williams A., LaMaster S. ( 1993) Price bubbles and crashes in experimental call markets, Economics Letters  41, 179– 185. Google Scholar CrossRef Search ADS   Van Leeuwen B., Noussair C. N., Offerman T., Suetens S., van de Ven J., van Veelen M. ( 2016) Predictably angry: facial cues provide a credible signal of destructive behavior, Management Science , forthcoming. Wright W. F., Bower G. H. ( 1992) Mood effects on subjective probability assessment, Organizational Behavior and Human Decision Processes  52, 276– 291. Google Scholar CrossRef Search ADS   Zaman B., Shrimpton-Smith T. ( 2006) The FaceReader: measuring instant fun of use, Proceedings of the fourth Nordic Conference on Human–Computer Interaction, NordiCHI, Oslo, Norway, pp. 457– 460. © The Authors 2017. Published by Oxford University Press on behalf of the European Finance Association. All rights reserved. For Permissions, please email: journals.permissions@oup.com http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Review of Finance Oxford University Press

Emotional State and Market Behavior

Loading next page...
 
/lp/ou_press/emotional-state-and-market-behavior-dxFds60wWU
Publisher
Oxford University Press
Copyright
© The Authors 2017. Published by Oxford University Press on behalf of the European Finance Association. All rights reserved. For Permissions, please email: journals.permissions@oup.com
ISSN
1572-3097
eISSN
1573-692X
D.O.I.
10.1093/rof/rfx022
Publisher site
See Article on Publisher Site

Abstract

Abstract We consider the relationship between trader emotions and asset market behavior. We create experimental asset markets with the structure first studied by Smith, Suchanek, and Williams (1988), which is known to generate price bubbles and crashes. To track traders’ emotions in real time, we analyze participants’ facial expressions with facereading software before and while the market is operating. We find that a positive emotional state correlates with purchases and overpricing. Fear correlates with selling, low prices, and price decreases. The experiment confirms the intuition that emotions and market dynamics are closely related. 1. Introduction The connection between asset market price movements and emotions has been widely accepted in popular press and commentary. The supposed existence of fear and exuberance as influences on prices is reaffirmed with great frequency in such quarters. Positive emotion is generally associated with booms and high price levels. Alan Greenspan, while chairman of the Federal Reserve, famously remarked that the American stock market exhibited an “irrational exuberance” when it experienced a rapid run up in 1996. The remark betrayed a belief on his part that the increase had, in part, an origin in positive emotions of traders.1Galbraith (1984) describes stock market price bubbles as “speculative euphoria”. On the other hand, fear is associated with anticipated price variability and cited as a force leading to selloffs and price declines. Market volatility indices such as the CBOE’s VIX, an index of option prices, are referred to colloquially as “fear” indices. The legendary investor Warren Buffett (2008) writes, “A simple rule dictates my buying: be fearful when others are greedy and be greedy when others are fearful”, associating the presence of fear in the market with opportunities to make profitable purchases. There is data supporting the contention that traders’ moods can lead to price movements at the market level. Hirshleifer and Shumway (2003) find that good weather is correlated with higher stock returns. They presume that the mechanism whereby this effect operates is through the positive effect that weather has on mood. Kamstra, Kramer, and Levi (2003) observe that returns are relatively low in the dark seasons of fall and winter and appeal to a similar intuition to explain their results. Sports scores seem to matter for financial returns (Edmans, Garcia, and Norli, 2007), with home team losses translating to lower prices. Bollen, Mao, and Zeng (2011) find that Twitter mood predicts subsequent stock market movements. Gilbert and Karahalios (2010) find that the level of anxiety of posts on the blog site Live Journal predicts price declines. In all of this literature, more positive (negative) emotional states are associated with higher (lower) prices. In this paper, we focus on the connection between trader emotions and extreme pricing episodes: asset price bubbles and crashes. We use an experimental approach, which exploits the fact that bubbles and crashes can be reliably created and studied in the laboratory. The bubble and crash pattern was first observed in the laboratory with a paradigm introduced by Smith, Suchanek, and Williams (1988). Subsequent authors have replicated and established the robustness of this price pattern, and the Smith, Suchanek, and Williams (1988) design has become the dominant experimental paradigm for studying bubbles and crashes. We adhere to this design in the work reported here, and it is described in Section 4. Bubbles can be eliminated in this setting when participants are inexperienced with the market environment, but it typically requires either a very strong framing that deemphasizes the importance of speculative possibilities or a considerable degree of specialized instruction (Lei and Vesely, 2009; Kirchler, Huber, and Stöckl, 2012). The magnitude of bubbles is sensitive to environmental parameters such as the amount of liquidity available (Caginalp, Porter, and Smith, 1998), institutional factors such as the ability to sell short and the trading process (Van Boening, Williams, and LaMaster, 1993; Haruvy and Noussair, 2006; Lugovskyy, Puzzello, and Tucker, 2011), and the time path of fundamentals (Noussair, Robin, and Ruffieux, 2001; Noussair and Powell, 2010; Kirchler, Huber, and Stöckl, 2012; Breaban and Noussair, 2015; Giusti, Jiang, and Xu, 2016). Nonetheless, there is considerable variation within all conditions that remains unexplained. That is, some sessions generate larger bubbles than others despite identical economic structure. We consider here whether variation in the emotional state of participants between different cohorts can account for some of this heterogeneity. In our experiment, we use facereading software to track the emotional state of all traders, as captured in their facial expressions. The software provides measures of happiness, surprise, anger, disgust, sadness, fear, neutrality, and overall emotional valence. According to Elster (1998), emotions can be differentiated from other mental states on the basis of six features: cognitive antecedents, intentional objects, arousal, valence, action tendencies, and physiological expressions. The work we report here focuses on the last feature, the physiological, as manifested in facial expressions. We consider several issues. First, at the market level, we study how emotional factors correlate with the magnitude of bubbles. We test the hypotheses that a positive emotional state on the part of traders before a market opens correlates with higher prices, and that greater fear correlates with lower prices. At the individual level, we consider whether emotions2 are linked to better performance, and explore the relationship between loss-averse decision-making and emotional state. We analyze how emotions affect the behavior of traders with propensities to use different strategies. To do so, while the market is operating, we track the bidirectional relationship between specific emotions and overall valence on the one hand, and individual decisions and market price movements on the other hand. Emotional valence is a psychological term referring to a composite measure of the positivity of emotional state. Emotions associated with positive states such as joy are said to have positive valence, while those that are adverse such as fear or anger are said to be negative in valence. The overall valence of an individual’s emotional state reflects the relative strength of the positive and negative components. While our hypotheses concern overall valence and fear, we also consider whether other emotions correlate with market activity. Other than one recent study on individual decision-making (Nguyen and Noussair, 2014) and one of ultimatum games (van Leeuwen et al., 2016), this paper represents, to our knowledge, the first application of facial recognition technology in either economics or finance. We find a number of strong relationships between emotions, as measured in traders’ facial expressions, and market behavior. Positive emotion is associated with higher prices and larger bubbles. The more positive the emotional valence a group of traders exhibits before the market opens, the higher prices are in the subsequent market. Individuals in more positive emotional states are more likely to make purchases, but the effect is specific to those who trade, irrationally, on momentum. Fear on the part of traders before a market opens is associated with lower prices. At the individual level, fear is associated with sales, and again, the link is specific to momentum traders. A greater average level of fear in the market presages a price decline. Those who exhibit more neutrality during a crash earn greater profits. We also observe a strong correlation between fear and loss aversion, as registered in a loss aversion measurement task administered before the market opens. In general, the experiments confirm the intuition that there is a close connection between market behavior and emotional state. 2. Previous Experimental Literature on Emotions and Markets Moods have been linked to behavior in a number of well-known experimental paradigms, and some of these involve markets. For example, positive moods can influence product choices (Meloy, 2000) and bidding in random nth price auctions (Capra, Lanier, and Meer, 2010). Johnson and Tversky (1983) argue that a positive mood tends to make beliefs more optimistic in the sense that probabilities associated with positive events become transformed in a positive direction. This effect would push individuals to make less risk-averse choices when they are in a more positive emotional state. This suggests one mechanism whereby emotional state could influence market behavior. Asset markets involve the trading of a risky lottery and thus less risk-averse agents would tend to place higher value on the asset, and their activity would lead to higher demand and prices. Indeed, Breaban and Noussair (2015) find that more risk-averse cohorts of traders tend to generate lower prices in experimental asset markets. Fellner and Maciejekovsky (2007) find that risk aversion on the part of a group of traders is associated with lower trading volume. Bosman and Riedl (2003) find that negative mood increases bidding in first-price sealed bid auctions, which is consistent with exhibiting more risk averse behavior. Loewenstein et al. (2001), surveying a large body of research in psychology, argue that a direct link exists between decision making under risk and emotional state. The specific emotion of fear has been associated with risk aversion in a number of studies. Lerner and Keltner (2001) find that induction of fear leads to pessimistic risk assessments and anger to optimistic ones. Since pessimistic risk assessments lead to more risk-averse decisions with respect to objective risks, fear correlates positively with risk aversion. Kugler, Connolly, and Ordonez (2012) obtain similar results in a different impersonal lottery-based task. Nguyen and Noussair (2014) also find that fear in facial expressions is positively correlated with risk averse choices. We are aware of three previous studies that explore the role of emotion in generating bubbles in experimental asset markets. All three papers consider markets with the structure of Smith, Suchanek, and Williams (1988), as we do here. Andrade, Lin, and Odean (2016) induce mood exogenously with film clips before the market opens. Subjects watch video clips that are (a) pleasantly exciting, (b) calm, and (c) fearful. They find that the pleasantly exciting video clips are associated with larger bubbles than the other three treatments. The other three conditions are not different from each other in terms of average asset prices. Lahav and Meer (2010) conduct an experiment with two treatments, which they call the positive and neutral treatments. Like Andrade, Lin, and Odean (2016) they induce mood by showing film clips to subjects before the market opens. Positive affect was induced with routines by comedian Jerry Seinfeld, and in the neutral treatment, no clip was shown. They find that the positive treatment is characterized by greater bubbles and higher prices than the neutral treatment, though the neutral treatment nonetheless generated price bubbles. Hargreaves-Heap and Zizzo (2011) conduct an experiment in which emotions are tracked over the course of the session. They focus on anger, anxiety, excitement, and joy. They have four conditions. In all conditions, subjects participate in two asset markets. In two of the treatments, individuals rate, on a Likert scale from 1 to 7, how intensely they currently feel each of the four emotions. In one of these conditions, subjects can chat with each other, and in the other they cannot chat. Hargreaves-Heap and Zizzo report that eliciting emotions does not in itself have an effect on market prices, but they do find that the level of excitement reported is positively correlated with price level. They also find that buying assets is linked to excitement and selling assets is connected to anxiety. They do not find a correlation between emotional state and trading profits. 3. Theoretical Model In this section, we construct a theoretical model that serves as our source of hypotheses for the experiment. Our model draws on elements from behavioral finance and psychology. Specifically, it combines the assumptions of heterogeneous trader types of the model of asset bubbles and crashes of DeLong et al. (1990), with two prominent psychological theories concerning the relationship between emotions and decision-making under risk. These are the Affective Generalization Hypothesis (Johnson and Tversky, 1983), and the Appraisal Tendency Framework (ATF) (Lerner and Keltner, 2001; Han, Lerner, and Keltner, 2007). Both of these frameworks have strong empirical support in prior experimental work. The model of DeLong et al. (1990) is our starting point for a number of reasons. The first is that it generates bubbles and crashes, the most prominent qualitative feature in our markets. The second is that it distinguishes between three types of trader that are reasonable descriptions of how most traders behave in experimental markets. The heterogeneity of types in this model makes it possible to consider which specific types of trader are affected by changes in emotional state. The third is that the model has been successfully applied to data from experimental markets with a similar structure to ours by Haruvy and Noussair (2006) and Haruvy, Noussair, and Powell (2014). The model of DeLong et al. (1990) posits three interacting trader types: (1) Fundamental Value traders, (2) Rational Speculators, and (3) Momentum traders. The demand of fundamental value traders depends only on the current price and fundamental value. The rational speculator’s demand is a function of the current price and the individual’s beliefs about future prices. The momentum trader, who is irrational, follows previous trends. We augment the model by allowing emotional state to influence the demand of each type. To relate emotional valence to trader behavior, our model assumes the Affective Generalization Hypothesis (Johnson and Tversky, 1983), which concerns the role of overall affect in judgments of risk. The hypothesis asserts that unrelated experiences that provoke certain type of emotions induce changes in risk perception. In their original study, Johnson and Tversky (1983) show that an induced mood has a large and pervasive impact on perceived risk, with a strong generalization effect for both positive and negative emotions. Positive affect decreases assessments of the likelihood of unfavorable outcomes and negative affect increases their perceived likelihood. In other words, people tend to overestimate the frequency of unfavorable risky events when they are in a negative mood, even if their emotional state is completely independent of the events themselves, and they tend to underestimate the frequency of such events when in a positive mood. A number of studies have confirmed these initial findings, and shown that risk taking increases when one is in a more positive emotional state and decreases in negative emotional states (Isen and Geva, 1987; Arkes et al., 1988; Wright and Bower, 1992; Nygren et al., 1996; Capra, 2004; Fehr-Duda et al., 2011). This is true regardless of whether the emotional state is integral, related to present judgment and choices, or incidental, independent of the current decision situation (Forgas, 1995). The asset trading in our market represents a lottery because of uncertain future dividends and resale prices. Thus, the affective generalization hypothesis would suggest a positive correlation between emotional valence and beliefs about perceived expected future dividends and resale prices. To model the effect of the specific emotion of fear, we apply the Appraisal Tendency Framework (Lerner and Keltner, 2001), which builds on Cognitive Appraisal Theory (Lazarus, 1991). The ATF predicts certain relationships between fear and decision making under risk. Under this framework, emotions arise from appraisals (or evaluations) of observed events. These then lead to action tendencies, which are propensities to behave in particular ways. According to the ATF, a state of fear induces an appraisal of future events as unpredictable and under situational control. This translates into an action tendency toward the avoidance of future risk (Lerner and Keltner, 2001; Han, Lerner, and Keltner, 2007) by individuals experiencing incidental fear.3 For simplicity, in our model, we will assume that the negative component of valence consists entirely of fear, so that an increase in fear will operate similarly to a decrease in valence. This is consistent with the popular intuition that negative sentiment in a market is typically better described as fear rather than as sadness, anger, disgust, or other negative emotional states. Suppose that there is an asset, with a finite life of T periods, trading in a competitive market populated by n traders. At the end of each period, the asset pays a dividend that is independently drawn each period from a stationary distribution, which can take on a finite number M of possible values. After the last dividend is paid, the asset has no value. Let d1t,…,dMt be the set of possible dividends in period t. Because the distribution of possible dividends is stationary, we can write d1t,…,dMt as d1,…,dM. Let πm equal the probability that the realized dividend equals dm. Then ∑mπmdm is the expected dividend the asset pays in any one period. Because the dividends constitute the only intrinsic source of value for the asset, the fundamental value of the asset in period t is given by ft=∑tT∑m=1Mπm dmt. The time period t can be interpreted in several different ways. One natural way to do so, followed by the literature in experimental finance, is to take each trading round in the experiment as a period. Under this interpretation, the asset pays a dividend at the end of each period. The time period can also be interpreted as a smaller unit, say a 10-s or a 30-s interval. Because of the market mechanism we employ in the experiment, the continuous double auction process, these short intervals will typically include one or more trades, and prices and demand can be readily defined over short intervals. Similarly, the fundamental value and the distribution of dividends can also be defined over short blocks of time. Under this latter interpretation, a dividend is only paid in a subset of periods. In our model, the market activity in one period will affect emotional state in the next, and emotions in the current period affect current actions. One can think of this feedback loop as repeated over a few seconds or longer intervals. Consider now trader i, who is of the Fundamental Value trader type. If unaffected by emotional state, he buys and sells based on prices and fundamentals, making purchases if prices are lower than fundamentals and sales if prices exceed them. Thus, the demand of fundamental value trader i at time t is given by:   DFVit(pt,ft),  with ∂DFVit∂ft>0, ∂DFVit∂pt<0. Let us now suppose that the trader’s assessment of the fundamental value is colored by his emotional state. Specifically, we assume that the Affective Generalization Hypothesis is operative so that more positive valence induces more optimistic risk assessments, while negative valence leads to more pessimistic ones. Let wmdm,V be the belief of a fundamental value trader, with valence level V∈-1,1 about the probability that the dividend will be equal to dm in the current period. Higher values of the variables V indicate more positive valence. We interpret fear as the emotion generating negative valence, so that fear φ  =min⁡0,V. The Affective Generalization Hypothesis is that risk assessments become more optimistic when an individual is in an emotional state with more positive valence. Thus, we assume that i’s conjectured distribution of future dividends Fiwmdm,Vi first-order stochastically dominates Fi'wmdm,Vi', if Vi> Vi'. In other words,   Vi>Vi'=> ProbFi'≤x|dm,Vi'≥ProbFi≤x|dm,Vi, ∀ x. (2) Denote trader i’s belief about the expected future dividend stream, by fit˜Vi. Then, the demand at time t of a fundamental value trader who may be affected by her emotional state is given by:   DFVit(pt,fit˜(Vit)), with ∂DFVit∂pt <0, ∂DFVit∂ft˜>0, ∂DFVit∂Vit>0. (3) Consider now the determinants of the valence of the fundamental value trader, Vit. We assume that a trader’s emotional state at time t depends on activity in period t – 1. We assume that the valence for a fundamental value trader is increasing in her perceived value of her current financial position. We assume that there is a lag in the emotional response to market activity, so that valence in period t depends on activity in period t − 1. Specifically, letting cit and qit be the cash and quantity holdings of trader i at time t,   Vitci,t-1,Vit-1,qi,t-1= Vitci,t-1+∑t-1T∑mwmdm,Vit-1*qi,t-1 . (4) Valence, therefore, is a function of one’s current wealth, evaluated at fundamental prices, and taking into account any distortions in beliefs due to emotional state. We assume that greater cash, asset holdings, and expected future dividend payments increase valence, so that ∂Vit∂ci,t-1 >0, ∂Vit∂qi,t-1 >0, and ∂Vit∂fi,t-1˜  > 0. Low cash, low asset holdings, and pessimistic beliefs about future dividends lower emotional valence. Now consider trader j, who is a rational speculator. This trader seeks to purchase for resale. The demand of rational speculator j at time t, DRjt, is a function of the prices in the current period and the expected best future resale price net of the fundamental. The demand takes the form:    DRjtpt,sjt, with ∂DRjt∂pt<0, ∂DRjt∂sjt >0, (5) where sjt is the trader’s belief about the maximum possible future resale price net of fundamental value. For rational speculators, higher current prices lead to lower quantity demanded and the prospect of a higher resale price leads to greater quantity demanded. In the absence of any influence of emotions,   sjt=Emaxt+k∈t+1,…,Tpt+k-ft+k. (6) However, we allow valence to influence the trader’s risk assessment. For the rational speculator, we do so by permitting emotions to affect her belief about the peak price relative to fundamentals in the future. More positive valence leads to more optimistic beliefs, so that if Vjt>Vjt', then sjtVjt>sjtVjt'. This implies that the demand of a rational speculator who may be affected by her emotional state will increase with Vjt, holding all else constant. The valence of a Rational Speculator is determined by the perceived maximal liquidation value of his financial position given his expectations of the future peak price relative to fundamentals.   Vjtcj,t-1, sj,t-1,qj,t-1= Vjtcj,t-1+sj,t-1qj,t-1 , (7) with Vjtcj,t-1, sj,t-1,qj,t-1 increasing in each of its three arguments. Finally, the demand of momentum trader k at time t is given by:   DMk tpt-1-pt-2. (8) Momentum traders use the previous price trend as a basis for their demand. Following DeLong et al. (1990) and Haruvy and Noussair (2006), we model this as a dependence on the most recent price change, pt-1-pt-2. We assume that emotional valence and fear influence the behavior of momentum traders because they act as filters, which translate the data into a propensity to make purchases and sales, resulting in demand of the form:   DMk tVktpt-1-pt-2, (9) with ∂DMk t∂Vk>0, so that more positive valence correlates with higher demand. Thus, aggregate demand is:   Dt(pt,.)=∑iDFVi t(pt,fit˜(Vit)) + ∑jDRjt(pt,sjt(Vjt))+ ∑kDMk t(Vkt(pt−1−pt−2))  (10) with ∂Dt∂pt <0,∂Dt∂pt-1 -∂Dt∂pt-2 >0, and∂Dt∂Vlt >0 for any trader l.(11) The total quantity demanded is decreasing in current price and increasing in the magnitude of the recent trend. An increase in the emotional valence of all traders at time t from Vt to Vt' induces an increase in quantity demanded on the part of all three types and thus on aggregate quantity demanded at time t. An increase in the valence of any individual also increases her own demand at time t. Thus, an increase in the valence of an individual trader of any of the three types also induces an increase in market demand. Let Q be equal to the total number of units of asset available. Because the supply of asset is fixed, the market clearing price is given by:   pt*=D-1,t(Q), with∂pt∂Vlt >0. (12) An increase in demand at time t induces an increase in the market clearing price at time t. This implies that an increase in the valence of all traders, or of one trader l holding all else fixed, results in higher prices. Similarly, an increase in the fear of all traders, or of one trader holding all else fixed, leads to lower prices. The effects of changes in valence are illustrated in Figure 1. The upper panel in the figure graphs market demand as a function of market price, and shows that it is shifted outward by an increase in the positivity of valence and thus inward by an increase in fear on the part of any or all traders. The bottom three panels trace the change in demand and the market clearing price resulting from an increase in valence of one or more traders of each of the three types. We assume that an increase in fear operates in an opposite manner as an increase in valence for each type, since it comprises the negative component of a trader’s valence. Figure 1 View largeDownload slide Response of aggregate market demand and demand of each of the three trader types to changes in valence and in fear. This figure depicts changes in the demand for the three different trader types as well as the aggregate demand when there is an increase in valence or fear. In the upper panel, the effect of an increase in the net positivity of emotional valence or of fear is depicted. The lower panels illustrate the effect of an increase in the net positivity of emotional valence on the demand of each type of trader, Fundamental Value, Momentum, and Rational Speculator, respectively. More positive valence shifts demand to the right for each type of trader. Figure 1 View largeDownload slide Response of aggregate market demand and demand of each of the three trader types to changes in valence and in fear. This figure depicts changes in the demand for the three different trader types as well as the aggregate demand when there is an increase in valence or fear. In the upper panel, the effect of an increase in the net positivity of emotional valence or of fear is depicted. The lower panels illustrate the effect of an increase in the net positivity of emotional valence on the demand of each type of trader, Fundamental Value, Momentum, and Rational Speculator, respectively. More positive valence shifts demand to the right for each type of trader. Now consider the special case of initial demand at t = 0 before the market opens for period 1. The demand of Fundamental Value trader i is based on her initial expectation of the future dividend stream and her emotional state, fi0˜(Vi0). The demand of Rational Speculator j depends on sj0Vj0 and the Momentum trader’s demand reduces to a function of Vk0, with higher valence corresponding to greater demand. It follows that more positive valence on the part of any one or all traders before the market starts correlates with a higher initial market price. The average valence in the market at time t > 0 is a result of the state of the market at t – 1, and is a weighted average of the valence of each of the three types, where the weighting is the percentage of traders of each type in the market. It depends positively on the valence in the preceding period, the cash and units held by fundamental value traders and rational speculators, the future expectations of speculators, and the recent market price trend. The average valence in period t is given by:   Vt=[∑iVit(cit,Vit−1, qit) + ∑jVjt(cjt, Vjt−1,qjt) +∑k(Vkt(pt−1−pt−2))]/n, (13) where n is the total number of traders, and dVtdcit>0,dVtdqit>0,dVtdsjt>0,dVtdcjt>0,dVtdqjt>0,dVtdVit-1>0, dVtdVjt-1>0, and dVtd[pt-1-pt-2]>0. Valence evolves over time in the following manner. More positive expectations about the peak prices relative to fundamentals increase average valence, because of their effect on Rational Speculators. An increasing price trend over the last two periods also makes average valence more positive through its impact on Momentum traders. For Fundamental Value traders and Rational Speculators, increases in the cash and quantity of units that they hold increase their valence. All of these factors will tend to increase prices and encourage bubbles. Fundamental Value traders or Rational Speculators running low on cash or units, a decrease in future expected prices, or a more negative price trend, all reduce valence. A summary of the theoretical predictions is given in Table I. Table I Summary of the theoretical predictions Symbol (+) indicates that a positive relationship is predicted between the variable listed in each row and valence. The upper panel indicates changes induced by an increase in valence. The lower panel indicates the changes in valence induced by a change in the market variables. The effects of and on each of the three trader types are shown separately.   FV Traders  RS Traders  M Traders  Average aggregate    Effects of an increase in valence  Beliefs dividend distribution  +        Beliefs future peak price    +      Demand  +  +  +  +  Prices        +  Effects of market variables on valence  Cash holdings  +  +      Quantity held  +  +      Price trend      +    Beliefs dividend distribution  +        Beliefs future peak price    +        FV Traders  RS Traders  M Traders  Average aggregate    Effects of an increase in valence  Beliefs dividend distribution  +        Beliefs future peak price    +      Demand  +  +  +  +  Prices        +  Effects of market variables on valence  Cash holdings  +  +      Quantity held  +  +      Price trend      +    Beliefs dividend distribution  +        Beliefs future peak price    +      4. The Experiment 4.1 Experimental Design The structure of the market was based on the paradigm created and studied in Smith, Suchanek, and Williams (1988). The asset that was exchanged in the market had a finite lifespan of T periods. At the end of each period t∈{1,…,T}, each unit of the asset paid a dividend dt that was independently drawn from a distribution that was identical for all periods. In any period t the expected dividend E(dt) on a unit of the asset was equal to the expected value of the dividend distribution. Dividends were drawn independently in each period. Therefore, the expected future dividend stream at time t, E[∑tTdt], equaled the expected period dividend multiplied by the number of periods remaining in the life of the asset. In other words, E[∑tTdt] = (T–t+ 1)E(dt). Since dividends were the only source of intrinsic value for the asset, the fundamental value ft had a particularly simple structure. It was equal, at any time t, to the expected future dividend stream from time t onward. In other words, ft = (T – t + 1)E(dt). In our markets, the life of the asset was T = 15, and the dividend was dt ϵ{0, 8, 28, 60}, where each realization was equally likely, for all t. Thus, E(dt) = 24, and ft = 24(16 – t) = 384 – 24*t at time t. The dividend distribution had a standard deviation of 27 per period, which was greater than the expected dividend. Therefore, risk-averse traders could value the asset at considerably less than its fundamental value. In each period, each trader had the ability to trade units of the asset for cash with any other trader in an open market, provided that he always maintained non-negative cash and share balances. Transaction prices were determined in a continuous double-auction market (Smith, 1962). This type of market operates in the following manner. Each period, the market is open for a fixed time interval, which was 2 min in this experiment. At any time while the market is open, any trader can submit an offer to sell or to purchase a share. These offers are posted publicly on all traders’ computer screens. Also at any time, any trader can accept an offer that another trader has submitted. When a bid or ask is accepted by a trader, a transaction for one share takes place between the trader who posted the offer and the trader who accepted it, at the offered price. Thus, within a period, it was possible for different transactions to occur at different prices. An individual could trade as much as he wished provided he had sufficient cash and units of the asset to complete the trades. Each subject had an identical portfolio, consisting of an initial endowment of 5 units of asset, and 5,000 units of experimental currency, at the beginning of period 1. A subject’s final earnings in the market were equal to the cash he had at the end of the experiment, which corresponded to his initial cash, plus the value of dividends received, plus (minus) any profit (loss) from trading. The market was computerized and used the Ztree program developed at the University of Zurich (Fischbacher, 2007). Prior to the opening of the asset market, we administered the loss aversion measurement task used by Trautmann and Vlahu (2011), which is based on an earlier protocol of Fehr and Goette (2007). This task consisted of a series of six choices, presented in a price list format. Each choice offered the opportunity to play a gamble which paid 4.5 Euros with probability 0.5 and either −0.5, −1.5, −2.5, −3.5, −4.5, or −5.5 Euros with probability 0.5, with each choice appearing exactly once. Subjects were required to indicate whether or not they accepted to play each of the six gambles. The number of gambles one decided not to play is interpreted as a measure of her loss aversion. Subjects completed the loss aversion measurement task using pen and paper. They submitted all six of their decisions simultaneously when they turned in their completed sheet of paper to the experimenter. They were informed prior to beginning the task that only one of the decisions would count toward their earnings. After all decisions were turned in, a die was rolled. The outcome of the roll determined which gamble would count for all participants. If a subject had chosen not to play the relevant gamble, she received a payoff of zero for this part of the experiment. If a participant chose to play the selected gamble, a coin was flipped to determine whether she received 4.5 Euro or the negative payment specified in the gamble.4 A separate coin was flipped for each participant who chose to play the gamble. 4.2 The Facereader Software During the sessions, all subjects were videotaped and the videotapes were analyzed later with Noldus Facereader. The taping began at least 30 s before the opening of the market for the first period and recorded continuously until the session ended. This is the first study to employ face reading in experimental finance. In our opinion, face reading is especially well-suited to the study of emotions for several reasons. The first reason is that it classifies an individual’s physiological state along emotional dimensions in a quantitative manner. This allows us, for example, to claim that one stimulus provokes more disgust but less sadness than another, or that a particular decision is taken when an individual is surprised rather than angry. A second advantage is that it registers emotional measurement in a manner that is completely unobtrusive to the participant, and data acquisition would proceed unnoticed if the individual were not informed that it was occurring.5 The third reason is that the facial expressions corresponding to the six basic emotions appear to be universal (Ekman and Friesen, 1986). It has been claimed that these expressions accompanying these six emotions are common to all cultures and primates (Ekman, 2007), and that they are the same for blind and sighted individuals (Matsumoto and Willingham, 2009), which provides strong evidence that they are innate. This means that results of studies such as ours should be replicable in different population groups and cultures. Happiness is positive in valence, surprise is neutral, and the other four basic emotions are negative. Happiness and anger are approach emotions, which tend to lead an individual to move toward the situation that triggers the emotion. Sadness, disgust, and fear, are withdrawal emotions, meaning that an individual typically seeks to avoid the stimulus that induces these emotions.6 4.3 Structure of the Data Our dataset consists of thirteen sessions. The sessions were conducted at Tilburg University and all subjects were students at the university. Subjects were recruited via an online system. No subject participated in more than one session of the experiment. On average, the sessions lasted 1 h. Between six and eleven traders participated in each session, with an average of eight subjects per session. Participants’ earnings from the asset market were converted to Euro at a rate of 500 units of experimental currency to 1 Euro. This resulted in an average payment of 15.6 euros (including the loss aversion measurement task). The market data consist of submitted bids and asks, as well as transactions, which are acceptances of bids and asks. We have market data for fifteen periods in each session. We also have the emotions data from all thirteen sessions for the 30 s before the market opened in period 1, and for the crash period, defined as the period within a session in which the greatest price decrease occurred. However, for a purely technical reason, we only have complete real time data for the totality of five of the thirteen sessions. Due to an improvement introduced in the video quality in late 2013, which increased the speed that the analysis could be conducted, in these five sessions it was possible to analyze all videos for the entire duration of the experiment. Therefore, for all fifty subjects participating in these sessions was possible to match their trading activity with their emotional responses for the entire session. This allows for an analysis of the dynamics between emotions and asset market behavior. The data are organized by two different lengths of time interval. The first is in 10 s intervals, which means that a session contains a total of 198 intervals. This consists of twelve intervals within each of the fifteen 2-min market periods, one interval after each period when the results for the period are displayed, and 30 s prior to the opening of the market for the first period. The data are in panel data format in which fifty subjects each form a panel, and each 10 s interval is an observation. The reason for specifying blocks of 10 s is that it is somewhat greater than the typical time course of emotional reactions. Emotions arise as a consequence of some events and last for a few seconds. There is little evidence in the literature on emotion duration, but Sonnemans and Frijda (1994) find that it depends to a great extent on the intensity of the emotion. Scherer, Walbott, and Summerfield (1986) find that different emotions tend to have different duration and they classify sadness as the most lasting one, followed by joy, anger, and fear. We expect that a 10 s interval is enough to capture both emotional and behavioral reactions to specific market events as well as short enough to capture the reaction to current activity only. Each emotion variable (happiness, fear, anger, disgust, sadness, surprise, neutral, and valence) was averaged every 10 s beginning with the market opening, so that the 300 observations that Facereader software provides in 10 s (30 per seconds) were averaged for each subject. We will refer to the unit of time in this data set as an interval. The second time scale on which we organized the data was in terms of the trading period, as it is typical in experimental work. We construct a data set that has observations for each of the fifty subjects during fifteen periods. In this case the emotion variables were averaged over the 120 s of each period and subject. In our presentation of the data, the 10-s intervals are indexed by τ = 1,…, 198, and periods by t = 1,…,15. Table II panels a and b show the mean and standard deviation of emotions between subjects, for both periods and 10-s intervals. Table II Mean and standard deviation of emotions (a) n = 10 020. Mean is calculated over each 10-s time interval. Valence ranges from [−1, 1]. Each of the seven specific emotions varies from [0, 1]. (b) n = 750. Mean is calculated over each 10-s time interval. Valence ranges from [−1, 1]. Each of the seven specific emotions varies from [0, 1]. a. Mean and standard deviation of emotions across subjects and 10-s intervals     Neutrality  Happiness  Sadness  Anger  Fear  Disgust  Surprise  Valence  Mean  0.524  0.092  0.083  0.038  0.001  0.012  0.021  −0.029  Standard deviation  0.305  0.153  0.128  0.091  0.007  0.054  0.067  0.231    a. Mean and standard deviation of emotions across subjects and 10-s intervals     Neutrality  Happiness  Sadness  Anger  Fear  Disgust  Surprise  Valence  Mean  0.524  0.092  0.083  0.038  0.001  0.012  0.021  −0.029  Standard deviation  0.305  0.153  0.128  0.091  0.007  0.054  0.067  0.231    b. Mean and standard deviation of emotions across subjects and periods     Neutrality  Happiness  Sadness  Anger  Fear  Disgust  Surprise  Valence  Mean  0.649  0.127  0.119  0.056  0.002  0.017  0.027  −0.030  Standard deviation  0.213  0.135  0.131  0.098  0.006  0.046  0.051  0.151    b. Mean and standard deviation of emotions across subjects and periods     Neutrality  Happiness  Sadness  Anger  Fear  Disgust  Surprise  Valence  Mean  0.649  0.127  0.119  0.056  0.002  0.017  0.027  −0.030  Standard deviation  0.213  0.135  0.131  0.098  0.006  0.046  0.051  0.151    In our analysis, we classify each of the traders participating in our experiment by trader types that correspond to the model in Section 3. The three types are (1) Fundamental Value Traders, (2) Momentum Traders, or (3) Rational Speculators. The traders are categorized according to the following criteria. We define an individual’s behavior as consistent with the Fundamental Value Trader type in period t if either one of two conditions holds. The first condition is that, if pt > ft, then qit < qi,t−1, where pt is the average price in period t, ft is the fundamental value in period t, and qit is the number of units of asset that individual i holds at the end of period t. This means that if prices are above fundamentals, trader i is a net seller of units in period t. The second condition is that if pt < ft, then qit > qi,t−1. If prices are below fundamentals, trader i is a net buyer in period t. The fundamental value trader, then, acts as if she is using the fundamental value as a limit price. A trader’s behavior is consistent with the Momentum Trader type if either of two conditions hold. The first is that, if pt−1 < pt−2, then qit < qi,t−1. The second is that, if pt−1 > pt−2, then qit > qi,t−1. The momentum trader is a net purchaser in period t if there has been an increasing price trend in the last two periods, and sells off units if there has been a decreasing trend. A trader’s behavior is consistent with the Rational Speculator Trader type if her behavior in period t satisfies one of the following two conditions. The first is that, if pt+1 < pt, then qit < qi,t−1, and the second is that, if pt+1 > pt, then qit > qi,t−1. This type of agent anticipates the price in the next period in an unbiased manner. She makes positive net purchases if the price is about to increase between the current and the next period. She makes net sales if the price is about to decrease. To classify a subject as one of the trader types, we count the number of periods during which a person is consistent with each type, and then classify him as the type with which he is consistent for the greatest number of periods. If there is a tie between two types, we classify the trader as belonging to each type with proportion 0.5. If there is a tie between all three types, he is assigned each type with proportion 0.33. Our model allows emotions to influence traders’ strategies. However, when we classify individuals into types, we assume that valence is not influencing trading strategies. This constraint can create some discrepancies between the model and the classification in that, for example, a fundamental value trader in a positive emotional state may purchase the asset at a price greater than objective fundamental value, while if emotions do not influence the decision, the trader would always sell if prices are greater than fundamentals. Thus, for purposes of the classification, we are in effect assuming that individuals are in a neutral emotional state. Some restriction of this type is unavoidable. In the absence of being able to directly estimate the parameters linking valence and demand, the classification we have conducted strikes us at the best way of categorizing individuals based on the variables we can observe. The percentage of individuals classified as each type is rather stable across sessions within this study, and relatively uncorrelated with the level if mispricing in the market, indicating that the influence of history on the market is not enough to shift the classification dramatically. The percentage classified as each type is also close to those observed by Haruvy and Noussair (2006) and Haruvy, Noussair, and Powell (2014), the other two studies that have reported a similar classification analysis, indicating that we have a subject pool that is typical in the distribution of strategies that it uses. The percentages classified as each type is indicated in Online Appendix III. 5. Hypotheses Based on our model and previous work, we advance several hypotheses about the relationships between emotions and market behavior. This first hypothesis follows from our theoretical model and is also suggested by the previous studies of Lahav and Meer (2010); Andrade, Lin, and Odean (2016); and Hargreaves-Heap and Zizzo (2012), who induce emotions exogenously prior to the market open. We hypothesize that the more positive the emotions that traders exhibit before a market opens, the greater the average price level in the market. Thus, we hypothesize that positive emotion is positively related to subsequent price, and thus in all likelihood within our setting, to greater bubbles. Hypothesis 1a: More positive emotional valence on the part of the average trader before the market opens correlates with higher subsequent prices and a larger bubble. To test this hypothesis, we check whether there is a correlation between (a) the average emotional valence within a group of traders in the 30 s before their market opens for period one, and (b) the average price in period 1, as well as the average over the fifteen periods the market is open. We also consider whether, as predicted by our model, fear correlates with lower prices. Andrade, Lin, and Odean (2016) fail to detect such an effect, and their attempt to induce fear generates similar results to a market in which emotions were not induced. However, Hargreaves-Heap and Zizzo do find that anxiety, a closely related emotional state, is correlated with lower prices. Hypothesis 1b: Greater fear on the part of the average trader before the market opens is correlated with lower subsequent prices. Our model predicts that positive valence on the part of individuals will correlate with higher demand on the part of these individuals. Therefore, we hypothesize that, at the individual level, a positive emotional state at any time on the part of an individual will correlate with her making greater net purchases of asset immediately afterward. Similarly, according to our model, fearful individuals exhibit lower demand for the asset. Moreover, the psychological implications of fear, being a withdrawal emotion, are that individuals will tend to avoid the situation that produces this emotion. Therefore, we anticipate a correlation between an individual’s level of fear and his likelihood of making a sale. Hypothesis 2a: Individuals in a more positive emotional state are more likely to make purchases in the subsequent 10-second time interval. Hypothesis 2b: Individuals who exhibit more fear are more likely to make sales in the subsequent 10 second time interval. As described in our model in Section 3, a trader’s demand might be affected by her emotional valence through different channels, depending on the type of trader she is. In the model, a fundamental value trader will tend to make more purchases when her beliefs about the future dividend stream are influenced by a positive emotional state. A rational speculator will purchase more units when, in a positive emotional state, she will overestimate the expected best future resale price net of the fundamental. A momentum trader will tend to buy or sell more units as positive valence or fear colors his interpretation of the past price trend. Therefore, we hypothesize that regardless the type of trader we are considering, their purchases and sales will correlate with the valence and fear they experience, respectively. We thus evaluate hypotheses 2a and 2b for each of the three types separately, as well as for the pooled data for all individuals. Our last hypotheses, 3a and 3b, concern the determinants of emotions. The model asserts that Fundamental Value traders’ emotional state is influenced by their cash and asset holdings and their beliefs about the fundamental value distribution. Speculators’ emotional valence is affected by their holdings of cash and asset, and their expectations of future trends. The observable variables among these are cash and asset holdings, and we hypothesize that the greater these are, holding all else constant, the more positive the valence of Fundamental Value traders and Rational Speculators. Lower holdings are associated with greater fear for these two types. Hypothesis 3a: Increases in cash and asset holdings correlate with more positive subsequent valence for fundamental value traders and rational speculators. The model also proposes that Momentum traders are influenced by the most recent price trend. The more positive the trend is, the more positive the valence of momentum traders, and the more negative the trend is, the greater is their fear. Hypothesis 3b: The more positive the recent trend of asset prices, pt-1 – pt-2, the greater the valence of momentum traders at time t. 6. Results 6.1 Market Price Patterns The time series of transaction prices in each of the thirteen sessions are shown in Figure 2, along with the time path of the fundamental value. In the figure, the vertical axis is in terms of experimental currency, and the horizontal axis indicates the market period. The black line is the time series of fundamental value and each grey line indicates an individual session. As can be seen in the figure, there are large differences between sessions, but in most sessions the bubble and crash pattern is observed. Typically, prices remain above fundamental values for a considerable period of time, and then exhibit a rapid fall toward fundamental value.7 Figure 2 View largeDownload slide Average transaction price, all periods, all markets. Figure 2 View largeDownload slide Average transaction price, all periods, all markets. 6.2 Emotions and Market Activity We now evaluate the hypotheses advanced in Section 5 which concern how valence and fear correlate with market price movements, and with the individual decisions of different trader types. The first two hypotheses, 1a and 1b, are about the relationship between the initial emotional profile of a cohort of traders and the overall price pattern they generate, and are summarized as results 1 and 2. Result 1. A more positive emotional state on the part of a trader cohort before the market opens is positively correlated with subsequent market price level Support for Result 1: We take the average valence that Facereader measures over the 30-s interval before the market opens for each subject. We then average it for all subjects in a session. Then, we correlate this average with the average transaction price over the course of the session.8Figure 3 plots the average initial group valence against the average price level over the fifteen-period life of the asset. The figure shows a clear positive relationship between emotional state and price. The Spearman correlation coefficient between the valence before the market opens and the average price level in a session is ρ = 0.6190 (p < .01). The correlation between average valence before the market opens and the price level in period 1 is 0.576 (p < .05).9 □ Figure 3 View largeDownload slide The relationship between cohort-average emotional valence prior to market open and average price level in a session. Each data point is the average valence within a trader cohort before the market opens for period 1. Figure 3 View largeDownload slide The relationship between cohort-average emotional valence prior to market open and average price level in a session. Each data point is the average valence within a trader cohort before the market opens for period 1. Result 2. Average trader fear before the market opens is negatively correlated with the subsequent price level in the market Support for Result 2: The relationship between the average fear that a cohort expresses before the market opens and the price level over the subsequent market is very pronounced. Figure 4 relates the fear that Facereader registers for the average trader before the market opens to the average price in the session. The figure shows a strong negative relationship between the two variables. The correlation is highly significant (ρ = −0.8333, p = 0.01). The correlation between average fear before the market opens and the average price in period 1 is negative at −0.428, though not significant (p = 0.144). □ Figure 4 View largeDownload slide The relationship between cohort-average fear prior to market open and average price level in a session. Each data point is the average fear within a trader cohort before the market opens for period 1. Figure 4 View largeDownload slide The relationship between cohort-average fear prior to market open and average price level in a session. Each data point is the average fear within a trader cohort before the market opens for period 1. The initial level of each of the other emotions considered separately also correlates with the subsequent average price level of the session, though not significantly. The other three negative emotions, sadness, anger, and disgust, correlate negatively with price level at ρ = −0.381, −0.428, and −0.333, respectively, while happiness and neutrality correlate positively with price level at ρ = 0.476 and 0.357. Although none of these correlations are significant, they are consistent with higher (lower) prices being associated with positive (negative) emotional states. Although results 1 and 2 concerned the relationship between initial emotional state and average market behavior, and show that hypothesis 1a and 1b are supported, result 3 reports a relationship between emotional state and price movements while the market is in progress. Result 3: The average fear in a cohort in period t−1 is positively correlated with a price decrease in period t Support for Result 3: We construct a dummy variable to identify the average price movement across the fifteen periods. The dummy takes value 1 if pt−pt−1 < 0, and 0 otherwise. We then run a logit model with subject-level fixed effects with the dummy variable as the dependent variable.10 Emotions at time t−1 are the independent variables in this model. Table III shows that with more fear on the part of traders, prices in the market are more likely to decrease. On the contrary, neutrality, happiness, and anger are negatively correlated with price decreases.11 The pattern seems to be in line with Lerner and Keltner’s (2001) findings that fear is associated with greater risk aversion, and happiness and anger are associated with risk seeking behavior. Overall valence is not significantly correlated with subsequent price movements. □ Table III Negative price movements as a function of prior average emotions within a trader cohort (dependent variable = 1, if pt < pt−1, and 0 otherwise) The estimated equation is ytz = β0 Constant + β1 Feart−1 + β2 Neutralityt−1 + β3 Happinesst−1 + β4 Angert−1 + β5 Disgustt−1 + β6 Sadnesst−1. ytz = 1 if the price decreases from period t – 1 to t in session z. Each emotion is averaged over time and across all traders in period t−1 of session z. Emotions range from [0, 1]. *p < 0.1, **p < 0.05, and ***p < 0.01.   Logit  Random effects  Fixed effects  Feart−1  402.26  354.45**  98.08  Neutralityt−1  −5.34  −6.35**  −14.27***  Happinesst−1  −5.03  −6.14**  −15.19***  Angert−1  −4.46  −5.40*  −14.30***  Disgustt−1  −4.91  −6.17  −20.25***  Sadnesst−1  −2.16  −2.96  −10.68**  Constant  5.40  6.46**      Logit  Random effects  Fixed effects  Feart−1  402.26  354.45**  98.08  Neutralityt−1  −5.34  −6.35**  −14.27***  Happinesst−1  −5.03  −6.14**  −15.19***  Angert−1  −4.46  −5.40*  −14.30***  Disgustt−1  −4.91  −6.17  −20.25***  Sadnesst−1  −2.16  −2.96  −10.68**  Constant  5.40  6.46**    We have seen that there is a positive relation between initial valence and price levels. One possible mechanism for sustaining high price levels might therefore be that positive valence on the part of individuals is associated with higher demand for the asset. Similarly, the relationship between fear and lower prices suggests that fear prompts a willingness to sell. These two relationships are expressed as our hypotheses 2a and 2b and are predicted by our model. As reported in results 4 and 5, both hypotheses are supported, though the timing is different than predicted for fear, which has its effect contemporaneously, rather than with a 10-s lag. Result 4: Traders with more positive valence at 10s time interval τ−1 make more purchases at time τ Support for Result 4: In order to determine how emotions interact with individual trading activity we run a Poisson count regression with subject-fixed effects, where the dependent variable is the number of units a subject has bought or sold during each 10s interval. Each transaction is coded at the time at which a bid or ask has been accepted. We analyze the influence of prior emotional state, controlling for financial position and price level, on the number of purchases and sales. We find that subjects make more purchases in the current interval, the higher the valence they exhibited in the previous interval. The left portion of Table IV, which includes the lagged value of purchases in Model 1, shows that more positive emotional valence Granger-causes purchases. Another intuitive result from the estimation is that the larger the number of units in their inventory, the less likely subjects are to make more purchases. □ Table IV Individuals’ purchases and sales at time interval τ as a function of emotions at τ−1 (Poisson count regression with subject-fixed effects) Estimates come from Poisson count regressions with the specification ylτ = β1 valenceτ−1(or fearτ−1) + β2ci,τ−1 + β3qi,τ−1 + β4pτ−1 + β5yi,τ−1. ylτ is the number of units bought or sold by trader i in 10-s time interval τ. *p < 0.1, **p < 0.05, and ***p < 0.01.   Buyτ  Buyτ    Sellτ  Sellτ  Model 1  Model 2  Model 3  Model 4  valenceτ−1  0.237*  0.238*  fearτ  4.995***  4.861***  moneyτ−1  7.29e−06  4.95e−06  moneyτ−1  3.42e−06  3.87e−06  unitsτ−1  −0.021**  −0.015  unitsτ−1  0.048***  0.044***  p levelτ−1  −0.00007  −0.00008  P levelτ−1  0.00012  −0.00012  Buyτ−1  0.355***    Sellτ−1  0.363***      Prob>chi2 = 0.000  Prob>chi2 = 0.0586    Prob>chi2 = 0.000  Prob>chi2 = 0.000    9,770 obs  9,770 obs    9,970 obs  9,970 obs    49 groups  49 groups    50 groups  50 groups    Buyτ  Buyτ    Sellτ  Sellτ  Model 1  Model 2  Model 3  Model 4  valenceτ−1  0.237*  0.238*  fearτ  4.995***  4.861***  moneyτ−1  7.29e−06  4.95e−06  moneyτ−1  3.42e−06  3.87e−06  unitsτ−1  −0.021**  −0.015  unitsτ−1  0.048***  0.044***  p levelτ−1  −0.00007  −0.00008  P levelτ−1  0.00012  −0.00012  Buyτ−1  0.355***    Sellτ−1  0.363***      Prob>chi2 = 0.000  Prob>chi2 = 0.0586    Prob>chi2 = 0.000  Prob>chi2 = 0.000    9,770 obs  9,770 obs    9,970 obs  9,970 obs    49 groups  49 groups    50 groups  50 groups  Result 5: Traders who are more fearful at time τ sell more units at time τ Support for Result 5: Controlling for the units and cash they have and how high average prices are compared with fundamentals, we do not find a significant effect of the lagged value of fear on current sales. This could be due to the fact that fear has a more immediate effect than some other emotions. To check this, we consider the contemporaneous level of individual fear as a correlate of sales, in Models 3 and 4 in Table IV. We find that fear is indeed related to larger reductions of the number of assets in inventory in a given period. It appears that positive emotional state has a relatively slow impact on purchases, while fear is contemporaneously correlated with sales. □ We now study the relationship between valence and purchases, as well as between fear and sales, for each trader type separately. Recall that we hypothesized that all trader types are swayed by emotion when making purchases and sales. Result 6 provides support for this assertion for Momentum traders only. Result 6: Momentum traders purchase more when they are in a more positive emotional state and sell more when they exhibit fear. There is no correlation between emotions and the number of purchases and sales of fundamental value traders or rational speculators Support for Result 6: As described in Section 3, we classify individuals according to their trading strategies, into Fundamental Value traders, Momentum traders, and Rational Speculators. We observe that 40% of our participants are Fundamental Value traders, 34% are Momentum traders, and 26% are Rational Speculators. We then analyze how emotions affect their trading behavior. Table V shows that the Momentum types, who are relatively unsophisticated and behave irrationally, both buy and sell based on their emotional state. They earn less money than the other two types. Momentum traders earn on average 6,233 experimental currency, while Fundamental Value traders and Rational Speculators earn 7,263 and 7,565, respectively (see also Haruvy and Noussair (2006) and Haruvy, Noussair, and Powell (2014). The behavior of Momentum traders accounts for the positive relationship between valence and subsequent purchases, as well as between fear and subsequent sales, documented in Table V. The coefficients are not significant for Fundamental Value and Rational Speculator types. □ Table V Individuals’ purchases and sales at time interval τ as a function of emotions at τ −1 for each trader type separately (Poisson count regression with subject-fixed effects) Estimates come from Poisson count regressions with the specification ylτ = β1 valenceτ−1 (or fearτ−1) + β2ci,τ−1 + β3qi,τ−1 + β4p τ−1 + β5yi,τ−1. ylτ is the number of units bought or sold by trader i in 10s time interval τ. *p < 0.1, **p < 0.05, and ***p < 0.01. Buyτ  Fundamental Value Trader  Momentum Trader  Rational Speculator Trader  Valenceτ−1  0.280  0.521**  −0.025  Moneyτ−1  0.00003  −0.00004  −0.00007**  Unitsτ−1  0.015  −0.119***  −0.068***  Price levelτ−1  −0.0004*  0.0007***  3.15e−06  Buyτ−1  0.435***  0.141*  0.370***    Obs: 3,762  Obs: 3,396  Obs: 2,612    Groups: 19  Groups: 17  Groups: 13    Prob>F = 0.0000  Prob>F = 0.0000  Prob>F = 0.0000    Buyτ  Fundamental Value Trader  Momentum Trader  Rational Speculator Trader  Valenceτ−1  0.280  0.521**  −0.025  Moneyτ−1  0.00003  −0.00004  −0.00007**  Unitsτ−1  0.015  −0.119***  −0.068***  Price levelτ−1  −0.0004*  0.0007***  3.15e−06  Buyτ−1  0.435***  0.141*  0.370***    Obs: 3,762  Obs: 3,396  Obs: 2,612    Groups: 19  Groups: 17  Groups: 13    Prob>F = 0.0000  Prob>F = 0.0000  Prob>F = 0.0000    Sellτ   Fundamental Value Trader  Momentum Trader  Rational Speculator Trader  Fearτ  4.601  4.641**  9.492  Moneyτ−1  −0.00005**  −0.00006*  0.00001***  Unitsτ−1  0.069***  0.027  0.077***  Price levelτ−1  0.0003  −0.0006**  0.0001  Sellτ−1  0.354***  0.294***  0.458***    Obs: 3,963  Obs: 3,396  Obs: 2,612    Groups: 20  Groups: 17  Groups: 13    Prob>F = 0.0000  Prob>F = 0.0000  Prob>F = 0.0000    Sellτ   Fundamental Value Trader  Momentum Trader  Rational Speculator Trader  Fearτ  4.601  4.641**  9.492  Moneyτ−1  −0.00005**  −0.00006*  0.00001***  Unitsτ−1  0.069***  0.027  0.077***  Price levelτ−1  0.0003  −0.0006**  0.0001  Sellτ−1  0.354***  0.294***  0.458***    Obs: 3,963  Obs: 3,396  Obs: 2,612    Groups: 20  Groups: 17  Groups: 13    Prob>F = 0.0000  Prob>F = 0.0000  Prob>F = 0.0000    The relationship between strong emotions and poor trading decisions seems to be in line with previous research that connects weaker emotions to higher earnings in asset markets. For example, Lo et al. (2005) and Lo and Repin (2002) find that emotional individuals achieve lower earnings as day traders. On the other hand, Coates (2012) documents how emotions are closely linked to effective stock trading, and a number of authors have argued that emotional responses can generally be beneficial for decision making (Damasio, 1994). However, our reading of the balance of the evidence from financial markets suggests that strong emotional responses would be correlated with unprofitable decisions. Of our three trader types, the Momentum traders are the least rational, and more influenced by their emotions. In previous studies it has been noted that relatively unsophisticated traders tend to accept offers made by other traders rather than submitting offers themselves (Plott et al., 1998). This suggests that it may be the Momentum traders who are accepting other traders’ offers to sell at high prices during the bubble. To consider this, we compute the total number of purchases relative to the total number of bids for each type of trader. A higher ratio indicates that subjects are less active submitting offers to the market; that is to say, their trades are being concluded by accepting other participants’ offers. For Rational Speculators the ratio is 0.49, for Fundamental Value traders it goes up to 0.60, and for Momentum traders the ratio is 0.72, which is consistent with the findings of Plott et al. (1998). The ratio of sales relative to asks is not significantly different between trader types: 0.45, 0.48, and 0.42, respectively. This means that it is the Momentum traders who tend to accept offers and make purchases at high prices in our overpriced markets, behavior that is typically suboptimal. We now consider the impact of market activity on emotions. As indicated in hypothesis 3a and 3b, our model assumes that the valence of both fundamental value traders and rational speculators is influenced by the amount of cash and the quantity of units they hold. The emotional state of momentum traders is not affected by their current holdings of cash or asset. Rather, the valence of momentum traders is affected by the recent price trend pt−1– pt−2. The estimates of valence as a function of current holdings of cash and asset, as well as of recent price trend, conducted for each trader type separately, are given in Table VI. The estimates are the basis of our next result. Table VI Valence as a function of current cash and asset holdings, as well as recent price trend, estimated separately for each trader type The dependent variable is the average valence of trader l over 10-s interval τ. The estimation is conducted separately for each of the three trader types. The independent variables are “money” ciτ and “value of units” pt * qiτ in the preceding 10-s time interval, and the price trend over the last two periods. *p < 0.1, **p < 0.05, and ***p < 0.01.   Valenceτ Fundamental Traders  Valenceτ Momentum Traders  Valenceτ Rational Speculators  const  −0.108***  −0.027  −0.082***  moneyτ−1  7.18e−06***  −1.56e−07  9.12e−06  Value of unitsτ−1  0.006***  −0.0009  0.0032*  pt−1−pt−2  −0.00026*  0.00034*  −0.0004**    Obs: 3,942  Obs: 3,379  Obs: 2,599    Groups: 20  Groups: 17  Groups: 13    Valenceτ Fundamental Traders  Valenceτ Momentum Traders  Valenceτ Rational Speculators  const  −0.108***  −0.027  −0.082***  moneyτ−1  7.18e−06***  −1.56e−07  9.12e−06  Value of unitsτ−1  0.006***  −0.0009  0.0032*  pt−1−pt−2  −0.00026*  0.00034*  −0.0004**    Obs: 3,942  Obs: 3,379  Obs: 2,599    Groups: 20  Groups: 17  Groups: 13  Result 7: Cash and asset holdings are positively correlated with emotional valence for Fundamental Value Traders and Rational Speculators, but not for Momentum Traders. The recent price trend is positively correlated with emotional valence for Momentum Traders, but not for the other two types. Hypotheses 3a and 3b are supported Support for Result 7: The table shows that hypotheses 3a and 3b are supported in the data. The variables Quantity Held (qit) and Cash Balance (cit) are significantly positive determinants of valence for both Fundamental Value traders and Rational Speculators, but not for Momentum Traders. The recent price trend is significantly positively related to valence for Momentum Traders, but takes negative and significant coefficients for the other two types. □ 6.3 Other Emotional Correlates of Behavior 6.3.a. Neutrality and the submission of offers The decision to submit offers to buy or to sell might depend on the emotional state of a trader. It seems reasonable to believe that if traders base their trading decisions at least partly on emotions, more neutrality on the part of traders would correlate with their being less active in the market. That is to say, more emotional agents would submit more offers to sell and to buy. In order to determine whether this is the case, we control for the overall financial position of an individual and the market prices and we investigate the relationship between emotions and submitting bids and asks. We estimate a Poisson count model where the dependent variable is the number of bids/asks/total number of orders that a subject has made in an interval of 10 s. Table VII shows that, at the individual level, more neutrality is associated with less initiation of orders, especially fewer bids. This seems to indicate that individuals who experience stronger emotions are more active in the market. Including the lagged value of the dependent variable in each model we obtain the result that neutrality Granger-causes the number of bids as well as the total number of bids and asks that a trader submits to the market. Table VII Number of bids and asks submitted by an individual at time τ as a function of her neutrality and market variables at τ – 1 Logit model with subject-fixed effects. Dependent variables are the number of units an individual offers to purchase (bids), the number of units an individual offers to sell (asks). Each unit of time is a 10-s intervals. *p < 0.1, **p < 0.05, and ***p < 0.01.   Bidsτ  Asksτ  Bids + Asksτ  Neutralityτ−1  −0.247**  −0.031  −0.126*  Moneyτ−1  −0.00008***  0.00005***  9.81e−06  Unitsτ−1  −0.058***  0.054***  0.020***  Price levelτ−1  −0.0002  0.0001*  −0.0001  Bidsτ−1  0.128***      Asksτ−1    0.097***    Bids+Asksτ−1      0.103***    Obs: 9,770  Obs: 9,971  Obs: 9,971    Groups: 49  Groups: 50  Groups: 50    Prob>F = 0.000  Prob>F = 0.000  Prob>F = 0.000    Bidsτ  Asksτ  Bids + Asksτ  Neutralityτ−1  −0.247**  −0.031  −0.126*  Moneyτ−1  −0.00008***  0.00005***  9.81e−06  Unitsτ−1  −0.058***  0.054***  0.020***  Price levelτ−1  −0.0002  0.0001*  −0.0001  Bidsτ−1  0.128***      Asksτ−1    0.097***    Bids+Asksτ−1      0.103***    Obs: 9,770  Obs: 9,971  Obs: 9,971    Groups: 49  Groups: 50  Groups: 50    Prob>F = 0.000  Prob>F = 0.000  Prob>F = 0.000  6.3.b. High prices and fear Table VIII reports estimates from a fixed effects regression, which shows that greater holdings of cash and units are related to a more positive emotional state, while higher price level is negatively correlated with valence. In the specification, the price level is defined as the difference between the average transaction price and the fundamental value of the asset at any time. The estimates confirm that given the emotional state in interval τ −1, more money and units are associated with higher valence in interval τ, and therefore valence is Granger-caused by these two variables. However, Table VIII also shows that price level takes a negative coefficient. Since the prices are nearly always above fundamentals, the price level variable here is nearly equivalent to a measure of mispricing. Our results seem to indicate that mispricing correlates negatively with valence, all else equal. This is presumably because some traders realize that the current price trajectory and thus the market value of their assets is not sustainable, prompting more negative emotions. As we have seen, traders with more positive valence make more purchases. Although our data are correlational, we could argue that the emotional process underlying the formation of a bubble might be that positive emotional state enhances purchases, but as prices increase and traders find themselves in an overpriced market, their emotions become less positive. Table VIII Emotional valence of individuals as a function of their financial position (subject-fixed effects) The dependent variable is the average valence of trader l over 10-s interval τ. The independent variables are “money” ci,τ−1, current quantity of “units” qi,τ−1, and the price pτ−1, during the preceding 10-s time interval. *p < 0.1, **p < 0.05, and ***p < 0.01.   Valenceτ  Valenceτ  moneyτ−1  2.51e−06*  4.01e−06**  unitsτ−1  0.0013*  0.0022***  p levelτ−1  −0.000044***  −0.000087***  valenceτ−1  0.480***    const.  −0.028**  −0.046***    Obs: 9,927  Obs: 9,970    Groups: 50  Groups: 50    Prob>F = 0.000  Prob>F = 0.000    Valenceτ  Valenceτ  moneyτ−1  2.51e−06*  4.01e−06**  unitsτ−1  0.0013*  0.0022***  p levelτ−1  −0.000044***  −0.000087***  valenceτ−1  0.480***    const.  −0.028**  −0.046***    Obs: 9,927  Obs: 9,970    Groups: 50  Groups: 50    Prob>F = 0.000  Prob>F = 0.000  Table IX shows the correlation between price level and valence, as well as with the individual emotions. The table shows that Spearman correlation test between price level and valence is negative and significant at 1% (ρ = −0.09). The correlation between price level, 0.395, and fear is significant at p < 0.01. Table IX Correlation between current price level and current emotional state of trader cohort Emotions are averaged for all traders in a given session. The unit of time is the market period t. *p < 0.1, **p < 0.05, and ***p < 0.01.   Fear  Valence  Happiness  Anger  Surprise  Disgust  Sadness  Neutrality  Price level  0.390***  −0.090***  −0.007  −0.070***  −0.267***  −0.033***  0.074***  −0.363***    Fear  Valence  Happiness  Anger  Surprise  Disgust  Sadness  Neutrality  Price level  0.390***  −0.090***  −0.007  −0.070***  −0.267***  −0.033***  0.074***  −0.363***  6.3.c. Crashes Of special interest are market crashes. These can be very large and are generally unanticipated by participants (Smith, Suchanek, and Williams, 1988; Haruvy, Lahav, and Noussair, 2007). The relationship between a crash episode and the dynamics of emotion is illustrated in Figure 5. The data are from sessions 8–13, markets with stark crashes, where the strength of the average level of several emotions that members of the session cohort exhibit over the periods just before, during, and after the crash, is plotted.12 These emotions are anger, fear, happiness, and surprise. They are normalized at the level observed in the period immediately preceding the crash. The data show a clear pattern. Sadness and anger exhibit modest increases during a crash as traders’ paper wealth declines. However, fear and surprise exhibit sharp increases, as uncertainty increases. By the time the crash ends, surprise has fallen sharply, and fear has stabilized at high levels. However, sadness and anger continue to increase, as traders apparently realize the extent of the losses that the crash has imposed. The figure illustrates the existence of a multi-faceted emotional reaction to a key market event and the ability of Facereader to coherently characterize this reaction. Figure 5 View largeDownload slide Normalized emotions in the period immediately before, during, and immediately after a crash. Crash period is defined as the period with the greatest price decrease in each session. “Before” and “after” indicate the periods immediately prior and subsequent to the crash period. Emotions are averaged among all individuals in a cohort of traders. Averaged emotions are normalized to a value of 1 in the period before the crash. Figure 5 View largeDownload slide Normalized emotions in the period immediately before, during, and immediately after a crash. Crash period is defined as the period with the greatest price decrease in each session. “Before” and “after” indicate the periods immediately prior and subsequent to the crash period. Emotions are averaged among all individuals in a cohort of traders. Averaged emotions are normalized to a value of 1 in the period before the crash. At the individual level, we analyze the relationship between the overall strength of emotions during a crash and trader profits. Lo and Repin (2005) and Lo et al. (2002) find that those who exhibit less volatility in their emotional state in the face of fluctuations in the market have better trading performance. In our experiment, the analogy to this pattern would be that the level of neutrality in one’s facial expression during a crash is correlated with greater trading profits. Figure 6 plots the relationship between the level of neutrality individuals exhibit during a crash period, and the final earnings an individual accrues over the entire fifteen-period market. The figure suggests that more neutrality during a crash is indeed correlated with better performance. Figure 6 View largeDownload slide Relationship between neutrality during crash period and total individual session earnings. Each data point corresponds to an individual participant. Neutrality is averaged across the entire crash period. Figure 6 View largeDownload slide Relationship between neutrality during crash period and total individual session earnings. Each data point corresponds to an individual participant. Neutrality is averaged across the entire crash period. The correlation, at the level of the individual, between her average neutrality during the crash period and her final earnings is 0.205 (p = 0.16). Neutrality correlates negatively with units held at the end of the crash period at ρ = −0.27 (p = 0.064). The other emotions do not correlate with the number of units held, and thus with the amount of unrealized capital losses, during a crash. The results are similar if the units held at the beginning of the crash period are considered (very few units are exchanged during a crash because of very low demand). 6.3.d. Loss aversion and emotions The last result describes a strong correlation between loss aversion and fear. The loss aversion protocol that was administered at the beginning of the sessions, and the measurement of the emotional profile of individuals before the market opens, permit an analysis of the correlation between loss aversion and the emotional state of participants at the individual level that is independent of any experience on the market. Our results show that those who make more loss-averse decisions exhibit more fear in their facial expressions, and have a more negative overall emotional state. There is no correlation between loss aversion and any other of the six basic emotions or with neutrality. Table X contains the correlations between the number of gambles declined in the loss aversion task and the average consistency of facial expressions with each of the six emotions that Facereader registers in the 30 s before the market opens. A greater number of gambles declined indicate greater loss aversion. The table shows that the correlation between fear and loss aversion, 0.3427, is positive and significant at the p < 0.05 level. The correlation between loss aversion and valence is negative (ρ = −0.3012, p < 0.05). In contrast, none of the correlations with other emotions are significant at even the 10% level. Table X The correlation between loss aversion and emotional state before experiment begins Number of observations: 55. Emotion data are from the 30 s time interval before the market opens for period 1. *p < 0.1, **p < 0.05, and ***p < 0.01.   Fear  Valence  Happiness  Anger  Surprise  Disgust  Sadness  Neutrality  Loss aversion  0.342***  −0.301**  −0.045  −0.068  −0.085  0.209  0.109  −0.198    (0.018)  (0.025)  (0.759)  (0.649)  (0.569)  (0.157)  (0.463)  (0.180)    Fear  Valence  Happiness  Anger  Surprise  Disgust  Sadness  Neutrality  Loss aversion  0.342***  −0.301**  −0.045  −0.068  −0.085  0.209  0.109  −0.198    (0.018)  (0.025)  (0.759)  (0.649)  (0.569)  (0.157)  (0.463)  (0.180)  7. Conclusion In this paper, we study the connection between emotions and asset market behavior. We find a number of patterns that conform to commonly expressed intuition about the link between emotions and asset prices. When traders are in a more positive emotional state at the time the market opens, subsequent asset prices are higher. When they exhibit more fear, subsequent prices are lower. Momentum traders’ purchases are correlated with positive emotional state and their sales are correlated with the fear they experience. Those who keep a relatively neutral emotional state during a crash earn greater profits. Greater fear on the part of the average participant is a harbinger of a price decrease. Thus, the data suggest a scientific basis for some popular notions regarding emotions and market behavior. Had we not found these correlations, it would suggest that the popular notions were biased or misguided, and our research agenda would have turned to the question of where such biases might originate. A number of factors have been shown to influence the incidence and magnitude of bubbles in the laboratory. These include the institutions of exchange; the time path of fundamentals; and the risk aversion, loss aversion, and cognitive ability of traders. The results reported here show that another factor can be added to the list; the emotional state of traders. Our findings build on similar results that have recently been reported (Lahav and Meer, 2010; Hargreaves-Heap and Zizzo, 2011; Andrade, Lin, and Odean, 2016), extending them to show that emotions and market behavior display interdependencies over time intervals as short as a few seconds. Overall, it is becoming clear that asset price bubbles in experimental markets are a complex phenomenon, subject to many determining influences, including emotional factors. We find a strong correlation between fear and loss aversion. Such a connection is, in our view, quite intuitive, though to our knowledge such a relationship has not been previously documented. Those who anticipate that they will have a more negative response to a financial loss exhibit more fear when placed in a situation in which losses are possible. In Section 3, we propose a mechanism whereby emotions influence market activity. Our model predicts correlations between emotional valence and fear on the one hand, and individual behavior and market outcomes on the other hand. The model provides for different ways that emotions can influence traders, based on the type of trading strategy that they tend to employ. The trading strategies are taken from the model of DeLong et al. (1990). We observe that emotions correlate with trading decisions only for Momentum Traders, who are irrational. The more rational Fundamental Value and Rational Speculator types do not exhibit a correlation between emotional state and trading activity. We also observed that, as predicted by our model, Fundamental Value traders and Rational Speculators are in a more positive emotional state when their holdings of cash and asset are greater, and that Momentum traders are in a more positive emotional state, the more positive the recent market price trend. This study is the first application of Facereading to experimental finance. This methodology had yielded what are, in our view, coherent results. Our perspective is that the strength of our results contributes to the validation of the methodology. In particular, we believe that Facereading has considerable potential for the study of markets. In starker experimental settings than the one studied here, the emotional response to specific events, such as to a price quote one has received, or to a specific transaction one has made or observed, can be isolated and studied. In particular, in future work, face reading can be used to study face-to-face market transactions. In such situations, facial cues are important sources of information about the intentions and emotional states of other parties to a potential transaction. In these settings, individuals may try to manipulate their facial expression as part of their strategy to obtain more favorable terms. Face reading technology is highly conducive to the study of such behavior. The finding that emotions correlate with market behavior has interesting policy implications. It shows that there is concrete basis for the idea that central banks, governments, firms, and the media, must take into account the effects of their announcements and releases of data on the emotional state of market participants and how this might in turn affect market prices. Such effects would arise in addition to the influence of the new information on fundamentals of the economy that might in turn affect asset price. While the effects of information on fundamentals can be well-understood with established techniques in finance and economics, studying the emotional component requires new tools. In our view, the methods described here constitute a step forward in this direction. Our results also demonstrate that a trader’s disposition to react emotionally to market events correlates with his trading performance. The traders whose behavior was irrational were more likely to be influenced by emotions, suggesting that fluctuations in emotional state are associated with poor trading decisions. This has implications for firms that employ traders. It shows that screening for emotional profiles might be valuable for hiring decisions. Furthermore, when making personal financial decisions, one would do well to consider one’s own emotionality. An emotional investor may be better off delegating her trading decisions to other, less emotional, decision makers. Supplementary Material Supplementary data are available at Review of Finance online. Footnotes 1 Our notion of positive emotional state is a short-term, as distinct from a longer-term, more stable state of well-being. Bernanke (2012) clearly articulates this distinction with regard to happiness. Happiness is a “short-term state of awareness that depends on a person’s perceptions of one’s immediate reality, as well as on immediate external circumstances and outcomes. By ‘life satisfaction’ I mean a longer-term state of contentment and well-being that results from a person’s experiences over time.” 2 By emotion, we refer to short-term affective states. This is a distinct, though related, notion to that of mood. See Capra (2004) for a discussion. While moods are affective states of relatively low-intensity, diffuse, enduring, and typically without a salient antecedent cause, emotions are more intense and short lived, and they usually have a proximate cause. 3 In the theory of emotion, a distinction between incidental and integral emotions is typically made. Incidental emotions are feelings that are normatively unrelated to the decision task at hand. Integral emotions, on the contrary, are generated by the activity current being undertaken or experienced. In this study, the initial emotional state before the market starts presumably primarily reflects incidental emotions. Emotions that we observe after the market activity has started are likely a different mix of integral and incidental emotions, with more weight on the integral. Both incidental and integral emotions fall within the purview of the Appraisal Tendency Framework (Cavanaugh et al., 2007). 4 Some subjects experienced monetary losses in this part of the experiment. However, they were informed that there would be subsequent activities in the session in which they could expect to earn money on average. No subject ended the session with negative final earnings, because income in the market phase of the experiment in all cases more than fully offset losses incurred in the loss aversion measurement task. 5 Subjects were told beforehand that they were being videotaped but not that their facial expressions were to be analyzed. They were informed that the videotapes would be viewed only by the two researchers conducting the studies. Although the study received the appropriate Institutional Review Board approval, in future projects we intend to debrief subjects after the experiment about the specific procedures to be followed regarding the analysis of the videotapes. 6 Online Appendix IV contains a detailed explanatory note on how the Facereader software operates. 7 The facial expression data exhibit several general characteristics. The first is that the valence is typically negative. This likely means that participation in the experiment yields disutility for participants compared with other activities. There is great volatility in emotional state even over short time intervals. This may reflect the large number and heterogeneity of events that one experiences in a period. There is no discernible decline in the overall strength of emotion over time, over the roughly 35-min period during which the asset market is in progress. Anger tends to be greater at the outset, possibly reflecting the fact that individuals who are concentrating tend to look like they are angry (see Zaman and Shrimpton-Smith, 2006), but within a few minutes it stabilizes. Valence reflects this pattern because it integrates this early anger, typically being very negative at the very beginning of a session but stabilizing at a moderately negative level for the rest of the session. 8 The same results would obtain if we used the average price difference from fundamentals pt – ft. This difference is referred to as the Bias in a market by Haruvy and Noussair (2006). 9 We test whether the heterogeneity of initial emotional states correlates with greater volume of trade over the session. However, the correlation between the variance of valence among participants before a session begins, and the volume of trade over the entire session, is 0.12, and is not significant at conventional levels. 10 We test the fixed effects versus random effects specifications for this estimation equation using the test of overidentifying restrictions and we reject the hypothesis that RE is consistent (p = 0.043). However, if we compare the estimates in Table III in the same manner as Rodriguez and Elo (2002), it seems to be the case that with random effects it is mainly the fear coefficient that differs by specification, while for some of the other coefficients the change is not as pronounced. This indicates that estimates from both specifications have some robustness. Rodriguez and Elo (2002) argue that such robustness is typically associated with the consistency of estimates. 11 It is interesting to mention here that emotion levels among traders correlate with the subsequent direction of price movements, but not their magnitude. Additional analysis shows that none of the emotions appear to have predictive power for how much prices will decrease. More precisely we constructed a variable that measured the magnitude of the price decrease as max{pt−pt−1, 0} and used it as a dependent variable in a regression with the lagged values of emotions as independent variables. None of the coefficients were significant. 12 The crash period is defined as the period with the largest decrease in prices compared with the previous period. References Andrade E., Lin S., Odean T. ( 2016) Bubbling with excitement: en experiment, Review of Finance  20, 447– 466. Google Scholar CrossRef Search ADS   Arkes H. R., Faust D., Guilmette T. J., Hart K. ( 1988) Eliminating the hindsight bias, Journal of Applied Psychology  73, 305– 307. Google Scholar CrossRef Search ADS   Bernanke B. ( 2012) Excerpt from Speech at the 32nd General Conference of the International Association of Income and Wealth on 6.8.2012, Cambridge, MA, USA. Bollen J., Mao H., Zeng X. ( 2011) Twitter mood predicts the stock market, Journal of Computational Science  2, 1– 8. Google Scholar CrossRef Search ADS   Bosman R., Riedl A. ( 2003) Emotions and shocks in a first-price sealed bid auction. Working paper, Maastricht University. Breaban A., Noussair C. N. ( 2015) Trader characteristics and fundamental value trajectories in an asset market experiment, Journal of Behavioral and Experimental Finance  8, 1– 17. Google Scholar CrossRef Search ADS   Buffett W. ( 2008) Buy American. I Am. The New York Times, October 16. Caginalp G., Porter D., Smith V. ( 1998) Initial cash/asset ratio and asset prices: an experimental study, Proceedings of the National Academy of Science of United States of America  95, 756– 761. Google Scholar CrossRef Search ADS   Capra M. ( 2004) Mood driven behavior in strategic interactions, American Economic Review  94, 387–372. Capra M., Lanier K., Meer S. ( 2010) The effects of induced mood on behavior in random nth price auctions, Journal of Economic Behavior and Organization  7, 223– 234. Google Scholar CrossRef Search ADS   Cavanaugh L. A., Bettman J. R., Luce M. F., Payne J. W. ( 2007) Appraising the appraisal-tendency framework, Journal of Consumer Psychology  17, 169– 173. Google Scholar CrossRef Search ADS   Coates J. ( 2012) The Hour between Dog and Wolf: Risk Taking, Gut Feelings and the Biology of Boom and Bust , Penguin Press, New York, NY, USA. Damasio A. ( 1994) Descartes’ Error: Emotion, Reason, and the Human Brain , Penguin Books, New York, NY, USA. DeLong, J. B., Shleifer, A., Summers, L. H., and Waldmann R. J. (1990) Positive feedback investment strategies and destabilizing rational speculation, Journal of Finance 45, 379–395. Edmans E., Garcia D., Norli O. ( 2007) Sports sentiment and stock returns, Journal of Finance  62, 1967– 1998. Google Scholar CrossRef Search ADS   Ekman P. ( 2007) Emotions Revealed, Recognizing Faces and Feelings to Improve Communication and Emotional Life , Henry Holt and Company, New York. Ekman P., Friesen W. V. ( 1986) A new pan-cultural expression of emotion, Motivation and Emotion  10, 159– 168. Google Scholar CrossRef Search ADS   Elster J. ( 1998) Emotions and economic theory, Journal of Economic Literature  36, 47– 74. Fehr, E. and Goette L. (2007) Do workers work more if wages are high? Evidence from a randomized field experiment, American Economic Review 97(1), 298–317. Fehr-Duda H., Epper T., Bruhin A., Schubert R. ( 2011) Risk and rationality: the effects of mood and decision rules on probability weighting, Journal of Economic Behavior and Organization  78, 14– 24. Google Scholar CrossRef Search ADS   Fellner G., Maciejekovsky B. ( 2007) Risk attitude and market behavior: evidence from experimental asset markets, Journal of Economic Psychology  28, 338– 350. Google Scholar CrossRef Search ADS   Fischbacher U. ( 2007) z-Tree: Zurich toolbox for ready-made economic experiments, Experimental Economics  10, 171– 178. Google Scholar CrossRef Search ADS   Forgas J. P. ( 1995) Mood and judgment: the affect infusion model (AIM), Psychological Bulletin  117, 39– 66. Google Scholar CrossRef Search ADS PubMed  Galbraith J. K. ( 1984) A Short History of Financial Euphoria , Penguin Books, UK. Gilbert E., Karahalios K. ( 2010) Widespread worry and the stock market, 4th International AAAI Conference on Weblogs and Social Media. Giusti G., Jiang J., Xu Y. ( 2016) Interest on cash, fundamental value process and bubble formation: an experimental study, Journal of Behavioral and Experimental Finance  11, 44– 51. Google Scholar CrossRef Search ADS   Han S., Lerner J. S., Keltner D. ( 2007) Feelings and consumer decision making: the appraisal-tendency framework, Journal of Consumer Psychology  17, 158– 168. Google Scholar CrossRef Search ADS   Hargreaves-Heap S., Zizzo D. ( 2011) Emotions and chat in a financial markets experiment. Working paper, University of East Anglia. Haruvy E., Lahav Y., Noussair C. N. ( 2007) Traders’ expectations in asset markets: experimental evidence, American Economic Review  97, 1901– 1920. Google Scholar CrossRef Search ADS   Haruvy E., Noussair C. N. ( 2006) The effect of short selling on bubbles and crashes in experimental spot asset markets, Journal of Finance  61, 1119– 1157. Google Scholar CrossRef Search ADS   Haruvy E., Noussair C. N., Powell O. ( 2014) The impact of asset repurchases and issues in an experimental market, Review of Finance  18, 681– 713. Google Scholar CrossRef Search ADS   Hirshleifer D., Shumway T. ( 2003) Good day sunshine: stock returns and the weather, Journal of Finance  3, 1009– 1032. Google Scholar CrossRef Search ADS   Isen A. M., Geva N. ( 1987) The influence of positive affect on acceptable level of risk: the person with a large canoe has a large worry, Organizational Behavior and Human Decision Processes  39, 145– 154. Google Scholar CrossRef Search ADS   Johnson E. J., Tversky A. ( 1983) Affect, generalization, and the perception of risk, Journal of Personality and Social Psychology  45, 20– 31. Google Scholar CrossRef Search ADS   Kamstra M. J., Kramer L. A., Levi M. D. ( 2003) Winter blues: a sad stock market cycle, American Economic Review  93, 324– 343. Google Scholar CrossRef Search ADS   Kirchler M., Huber J., Stöckl T. ( 2012) Thar she bursts—reducing confusion reduces bubbles, American Economic Review  102, 865– 883. Google Scholar CrossRef Search ADS   Kugler T., Connolly T., Ordonez L. D. ( 2012) Emotion, decision, and risk: betting on gambles versus betting on people, Journal of Behavioral Decision Making  25, 123– 145. Google Scholar CrossRef Search ADS   Lahav Y., Meer S. ( 2010) The effect of induced mood on prices in asset prices: experimental evidence. Working paper, Ben-Gurion University of the Negev. Lazarus R. S. ( 1991) Emotion and Adaptation , Oxford University Press, New York. Lei V., Vesely F. ( 2009) Market efficiency: evidence from a no-bubble asset market experiment, Pacific Economic Review  14, 246– 258. Google Scholar CrossRef Search ADS   Lerner J. S., Keltner D. ( 2001) Fear, anger and risk, Journal of Personality and Social Psychology  81, 146– 159. Google Scholar CrossRef Search ADS PubMed  Lewinski P., de Uyl T., Butler C. ( 2014) Automated facial coding: validation of basic emotions and FACS AUs in facereader, Journal of Neuroscience, Psychology and Economics  7, 227– 236. Google Scholar CrossRef Search ADS   Lo A., Repin D., Steenbarger B. ( 2005) Fear and greed in financial markets: a clinical study of day traders, American Economic Review  95, 352– 359. Google Scholar CrossRef Search ADS   Lo A., Repin D. ( 2002) The psychophysiology of real-time financial risk processing, Journal of Cognitive Neuroscience  14, 323– 339. Google Scholar CrossRef Search ADS PubMed  Loewenstein G., Weber E., Hsee C., Welch N. ( 2001) Risk as feelings, Psychological Bulletin  127, 267– 286. Google Scholar CrossRef Search ADS PubMed  Lugovskyy V., Puzzello D., Tucker S. ( 2011) An experimental study of bubble formation in asset markets using the tatonnement trading institution. Working paper 07/2011, Canterbury University. Matsumoto D., Willingham B. ( 2009) Spontaneous facial expressions of emotion of blind individuals, Journal of Personality and Social Psychology  96, 1– 10. Google Scholar CrossRef Search ADS PubMed  Meloy M. ( 2000) Mood‐driven distortion of product information, Journal of Consumer Research  27, 345– 359. Google Scholar CrossRef Search ADS   Nguyen Y., Noussair C. N. ( 2014) Risk aversion and fear, Pacific Economic Review  19, 296– 312. Google Scholar CrossRef Search ADS   Noussair C. N., Robin S., Ruffieux B. ( 2001) Price bubbles in laboratory asset markets with constant fundamental values, Experimental Economics  4, 87– 105. Google Scholar CrossRef Search ADS   Noussair C. N., Powell O. ( 2010) Peaks and valleys: price discovery in experimental asset markets with non-monotonic fundamentals, Journal of Economic Studies  37, 152– 180. Google Scholar CrossRef Search ADS   Nygren T. E., Isen A. M., Taylor P. J., Dulin J. ( 1996) The influence of positive affect on the decision rule in risk situations: focus on outcome (and especially avoidance of loss) rather than probability, Organizational Behavior and Human Decision Processes  66, 59– 72. Google Scholar CrossRef Search ADS   Plott C. R., Menshikov I., Menshikova O., Myagkov M. ( 1998) From non market attitudes to market behavior: laboratory market experiments in Moscow, and the Hvatat Property of Human Behavior. Social Science Working Paper, 987, California Institute of Technology, Pasadena, California. Rodriguez G., Elo I. ( 2002) Intra-class correlation in random-effects models for binary data, The Stata Journal  3, 32– 46. Scherer K. R., Walbott H. G., Summerfield A. B. ( 1986) Experiencing Emotions: A Cross-Cultural Study , Cambridge University Press, Cambridge, UK. Smith V. ( 1962) An experimental study of competitive market behavior, Journal of Political Economy  70, 111– 137. Google Scholar CrossRef Search ADS   Smith V., Suchanek G., Williams A. ( 1988) Bubbles, crashes, and endogenous expectations in experimental spot asset markets, Econometrica  56, 1119– 1151. Google Scholar CrossRef Search ADS   Sonnemans J., Frijda N. H. ( 1994) The structure of subjective emotional intensity, Cognition and Emotion  8, 329– 350. Google Scholar CrossRef Search ADS   Terzis V., Moridis C., Economides A. ( 2010) Measuring instant emotions during a self-assessment test: the use of FaceReader, Proceedings of Measuring Behavior, 192–195. Trautmann S., Vlahu R. ( 2011) Strategic loan defaults and coordination: an experimental analysis. Working paper, Dutch National Bank. Van Boening M., Williams A., LaMaster S. ( 1993) Price bubbles and crashes in experimental call markets, Economics Letters  41, 179– 185. Google Scholar CrossRef Search ADS   Van Leeuwen B., Noussair C. N., Offerman T., Suetens S., van de Ven J., van Veelen M. ( 2016) Predictably angry: facial cues provide a credible signal of destructive behavior, Management Science , forthcoming. Wright W. F., Bower G. H. ( 1992) Mood effects on subjective probability assessment, Organizational Behavior and Human Decision Processes  52, 276– 291. Google Scholar CrossRef Search ADS   Zaman B., Shrimpton-Smith T. ( 2006) The FaceReader: measuring instant fun of use, Proceedings of the fourth Nordic Conference on Human–Computer Interaction, NordiCHI, Oslo, Norway, pp. 457– 460. © The Authors 2017. Published by Oxford University Press on behalf of the European Finance Association. All rights reserved. For Permissions, please email: journals.permissions@oup.com

Journal

Review of FinanceOxford University Press

Published: Feb 1, 2018

There are no references for this article.

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
discover and read the research
that matters to you.

Enjoy affordable access to
over 12 million articles from more than
10,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Unlimited reading

Read as many articles as you need. Full articles with original layout, charts and figures. Read online, from anywhere.

Stay up to date

Keep up with your field with Personalized Recommendations and Follow Journals to get automatic updates.

Organize your research

It’s easy to organize your research with our built-in tools.

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.

See the journals in your area

Monthly Plan

  • Read unlimited articles
  • Personalized recommendations
  • No expiration
  • Print 20 pages per month
  • 20% off on PDF purchases
  • Organize your research
  • Get updates on your journals and topic searches

$49/month

Start Free Trial

14-day Free Trial

Best Deal — 39% off

Annual Plan

  • All the features of the Professional Plan, but for 39% off!
  • Billed annually
  • No expiration
  • For the normal price of 10 articles elsewhere, you get one full year of unlimited access to articles.

$588

$360/year

billed annually
Start Free Trial

14-day Free Trial