Exploring Mispricing in the Term Structure of CDS Spreads

Exploring Mispricing in the Term Structure of CDS Spreads Abstract Based on a reduced-form model of credit risk, we explore mispricing in the credit default swaps (CDS) spreads of North American companies and its economic content. Specifically, we develop a trading strategy using the model to trade out of sample market-neutral portfolios across the term structure of CDS contracts. Our empirical results show that the trading strategy exhibits abnormally large returns, confirming the existence and persistence of a mispricing. The aggregate returns of the trading strategy are positively related to the square of market-wide credit and liquidity risks, indicating that the mispricing is more pronounced when the market is more volatile. When implemented on the Markit data, the strategy shows significant economic value even after controlling for realistic transaction costs. 1. Introduction The credit derivatives markets have experienced tremendous growth in the last decade. According to the Bank for International Settlements (BIS), the notional value of outstanding credit derivatives peaked at the end of 2007 with $58 trillion, then dropped sharply in 2008, and gradually stabilized at around $26 trillion in recent years. The single-name credit default swaps (CDS) are believed to be the most liquid and popular product, as they account for more than two thirds of all outstanding credit derivatives. Though some exotic credit derivatives, such as subprime collateralized debt obligations (CDO) caused tremendous problems in the financial crisis, the vanilla CDS contracts play important economic roles. The newly proposed regulations, such as the establishment of a central clearing house for CDS, would help to reduce systemic risk and improve transparency in the CDS market. Therefore, the CDS contracts are likely to remain the preferred vehicle for investing, speculating, and managing single name credit risk.1 The rapid growth of the CDS market makes it possible to speculate on the relative pricing of the credit risk of a company across a wide range of maturities. Although 5-year CDS have historically been the most liquid contracts, nowadays a complete credit curve (CDS spreads over different maturities) is available for many companies. As a result, it is possible to buy and sell protections on a given firm at different maturities. An interesting question to both academics and practitioners arises as to whether the credit risk of a firm is consistently priced across maturities. From an academic perspective, an important issue is whether existing credit risk models, either structural or reduced-form, can capture the rich term structure behaviors of credit spreads. From a practical perspective, one challenging issue is whether one can design trading strategies to exploit potential mispricings along the credit curve. In two recent WSJ articles, Burne (2015, 2016) reports “Wall Street’s biggest banks have agreed to a tentative settlement over allegations that they colluded to influence the market for credit derivatives. […] The lawsuit brought by a group of investors accused the banks, the International Swaps and Derivatives Association and data provider Markit Group Ltd. of colluding to block competing providers, including exchanges, from entering the market for derivatives called credit-default swaps.” The likely settlement over these allegations casts doubt on the price discovery function and efficiency of the CDS market. Instead of directly studying the liquidity and competitiveness of the CDS market, our article provides a test of the hypothesis that the CDS market is efficient given its current liquidity and competitiveness. As pointed out by Phillips and Smith (1980) and Hogan et al. (2004), the existence of abnormal returns after taking into consideration of proper trading costs rejects market efficiency. We provide strong evidence of the existence of mispricing in our empirical section below. Using a reduced-form model of credit risk, we explore mispricing in the term structure of CDS spreads for a large number of North American companies. Specifically, we consider 500 firms with almost continuous daily observations of CDS spreads with maturities of 1, 2, 3, 5, 7, 10, 15, 20, and 30 years between January 2005 and October 2011. We estimate an affine model of credit risk for each company based on its term structure of CDS spreads and identify “mis-valued” CDS contracts relative to the model. Based on the estimated model parameters, we construct a portfolio of CDS contracts that are both delta- and gamma-neutral to the potential changes in credit spreads. Then we long (short) the portfolio if it is under (over) valued relative to our model and unwind the portfolio a week later. We conduct both in-sample and out-of-sample analysis. In the in-sample analysis, we estimate model parameters, construct trading portfolios, and calculate trading profits using all of the data. In the out-of-sample analysis, we estimate model parameters using the first half of the sample, and we construct trading portfolios and calculate trading profits using the second half of the sample. In both in-sample and out-of-sample exercises, the strategy generates significantly positive excess returns. These results, especially the out-of-sample tests, provide strong evidence of CDS mispricing. To explore the economic content of the mispricing, we average the out-of-sample returns across all firms and look at the relation between aggregate returns and various risk factors. We find that the aggregate returns are significantly and positively correlated with the square of market-wide credit and liquidity risks. This also means that the mispricing is more pronounced when the market is more volatile. Our robustness tests show the mispricing persists not only during the financial crisis period but in normal times. We do find evidence that the mispricing is more significant during the financial crisis period. Although not our main focus, we also conduct a performance analysis accounting for realistic transaction costs. The purpose of the analysis is not to offer a complete and practical guide to the implementation of a statistical arbitrage strategy, but merely to shed some light on its economic significance. This might be potentially useful: given the positive relation between aggregate returns and nonlinear risk factors, our trading strategy might provide quantitative CDS investors a means to obtain a highly profitable risk position. In this analysis, we find that our trading strategy is quite profitable: for half of the firms, the annualized Sharpe ratio (SR) is well above one in out-of-sample tests.The transaction costs included are as realistic as possible. For example, the included bid-ask spreads applying to the term structure of the CDS contracts are based on the evidence documented in recent CDS empirical studies (see Biswas, Nikolova, and Stahel, 2014; Arakelyan and Serrano, 2012). We also find that the trading strategy has significant economic value in terms of the maximum performance fee due to Fleming, Kirby, and Ostdiek (2001) with respect to popular exchange-traded funds (ETF), and the computed statistical arbitrage opportunities are statistically significant in terms of the formal test developed by Hogan et al. (2004). Our article contributes to three areas in the literature. First, our study relates to the informational efficiency of CDS markets. There are only a few studies looking at the efficiency of the CDS market. Norden and Weber (2004), Zhang (2009), and Zhang and Zhang (2013) study the informational efficiency of the CDS market by analyzing the response of CDS to rating announcements, a variety of credit events, and earnings surprises, and they find evidence supportive of an efficient market. Recently, Du and Zhu (2015) study the design of CDS auctions and find that the current design leads to biased and inefficient prices. Although these studies have shed some light on the efficiency issue, and provide some interesting insights, they do not confront the core issue. As pointed out by Stulz (2010), the ultimate way to test the CDS market is to look at “the inefficiency brought about by limits of arbitrage—because otherwise arbitrageurs could exploit any mispricing of risk.” We take the direct route by developing a trading strategy to test the pricing efficiency of the CDS market. Second, our article also contributes to the growing literature of persistent mispricing. For example, Mitchell, Pedersen, and Pulvino (2007) and Duffie (2010) discuss the role that slow-moving capital may play in allowing arbitrage opportunities to exist for extended periods of time. Fleckenstein, Longstaff, and Lustig (2014) show the persistent mispricing driven by slow-moving capital exists in the TIPS-Treasury market. Meanwhile, Brunnermeier and Pedersen (2009) show that funding availability to intermediaries in financial markets is a potential explanation for deviations of security prices from the no-arbitrage conditions. Mayordomo, Peña, and Romo (2014) study the relative pricing between CDS and Asset Swap Packages (ASPs) and find persistent mispricing; however, they only consider the relative pricing between CDS and ASPs. Our study documents persistent mispricing in the CDS market that might be due to market manipulations (see, e.g., Allen, Litov, and Mei, 2006; Aggarwal and Wu, 2006). Our third contribution lies in the credit risk literature that has arisen in the past decade. There are many empirical studies on CDS involving the modeling of the entire credit curve given the increased availability of CDS spreads for a wide range of maturities. A few studies that are closely related to ours are Zhang (2008), Pan and Singleton (2008), and Chen, Cheng, and Wu (2013), who estimate default risk models using the entire credit curve of CDS spreads. One contribution of our article is that we are among the first to explore mispricing in the term structure of CDS spreads and its economic content. The rest of this article is organized as follows. In Section 2 we motivate our study. In Section 3, we develop a one-factor affine term structure model for CDS and discuss the econometric methods for estimating the model. Section 4 discusses the data and empirical results from the estimation. In Section 5 we discuss our trading strategy. We show the evidence of mispricing, explore and interpret its economic content in Section 6. Section 7 shows the net fee performance and economic significance of the trading strategy. Section 8 concludes. Appendices and Online Appendices contain technical details and supplementary results. 2. Motivation The CDS market boomed after the Fed permitted banks to use CDSs to reduce capital reserves in 1996 (Levine, 2012). The growth of this market was exponentially fast from 2005 to 2007. The overall CDS market reached a notional value of $58 trillion in 2007. Due to this explosive growth, market participants extensively use CDS for risk management and speculation (Gibson, 2007). This CDS usage might be partially due to the promotion of the CDS market by both the academic (e.g., Longstaff, Mithal, and Neis, 2005; Blanco, Brennan, and Marsh, 2005) and regulatory communities (e.g., CDS implied default probabilities are being considered to replace credit ratings in US financial regulations, Jarrow (2012)). This rapid rise of the CDS market (unintentionally) gave rise to a common impression that the market is liquid, competitive,2 and efficient. This impression is debatable. van Deventer (2015) notes that 72.48% of the trading volume in single name CDS consists of only dealer to dealer trades.3 A problem with a dealer dominated market is that it is nearly costless for dealers to inflate gross trading volume by trading among themselves. However, this article does not directly examine the liquidity of the CDS market since there are already extensive studies in the literature looking at this issue. Instead, we take the current liquidity of this market as given and focus on market efficiency. It is reported in recent WSJ articles (Burne, 2015, 2016) that twelve big banks and two industry groups4 have tentatively agreed to pay $1.87 billion to settle allegations that they conspired to rig the CDS market. The accusations are that there was collusion among the banks and industry groups to block competing providers, e.g., exchanges from entering the CDS market. The simplest way of colluding is to manipulate the prices of CDS contracts. Indeed, the current CDS market with a majority of the players being institutional traders with large market power satisfies the classic conditions for market manipulation, see Jarrow (1992). Therefore, the pending settlement over these allegations suggests the possible existence of price manipulation in the CDS market. This brings the pricing efficiency of the market into question. As such, we explore the existence of mispricing in the CDS market and its economic content. It is evidenced in Allen, Litov, and Mei (2006) and Aggarwal and Wu (2006) that price manipulation is typically accompanied by increased volatilities. Therefore, in our empirical analysis we specifically look at the relation between mispricing and volatilities of various risks, e.g., credit and liquidity risks. 3. Model and Estimation Method 3.1 The Model In this section, we develop a one-factor affine model for the term structure of CDS spreads. We use only one factor to capture the dynamics of credit risk because our principal component analysis (PCA) shows that the first principal component captures 96% of the variations of CDS spreads. Our model is similar to that of Longstaff, Mithal, and Neis (2005), Duffie and Singleton (1999), Duffie and Singleton (1997), Duffie, Pedersen, and Singleton (2003), and Zhang (2008). We assume that credit spreads are independent of interest rates and thus avoid estimating a model for the risk-free term structure. As a robustness check, we run and obtain similar results using a two-factor affine model for the risk-free term structure, in which credit spreads are correlated with the two interest rate factors. To investigate mispricing, we equate the “correct” price to the arbitrage-free price by assuming that the market is arbitrage-free and, hence, the existence of an equivalent martingale measure Q. Formally, we assume that the state variable, i.e. the default intensity Zt, follows a square root process (Cox-Ingersoll-Ross (CIR) process) as   dZt=(α−κZt)dt+σZtdwZQ(t), (1) where wZQ(t) is a standard Brownian motion under the equivalent martingale measure Q. While we only need the dynamics of the state variable under the Q measure for pricing purposes, we need its dynamics under the P measure for econometric estimation. Given the extended affine specification for the market price of risk (Cheridito, Filipović, and Kimmel, 2007), we model the state variable’s P measure dynamics as   dZt=(αP−κPZt)dt+σZtdwZP(t). (2) The relation between the Wiener processes under the two measures is given by   wZP(t)=wZQ(t)+α−αPσ∫0t1Zsds−κ−κPσ∫0tZsds. (3) To compute the CDS spread, we assume a constant recovery rate. Since both the buyer and the seller of credit protection in a CDS are exposed to counterparty risk, the quoted recovery rates might differ from the real recovery rates implicit in the CDS spreads. Therefore, unlike the common practice in the literature which fixes the recovery rate to a predetermined constant (see, e.g., Longstaff, Mithal, and Neis, 2005; Zhang, 2008), we estimate the value of the constant recovery rate along with the model parameters from the market prices of CDS spreads. Under the fractional recovery of face value (RFV) framework, which has been widely used for pricing CDS and is consistent with the market practice, the recovery rate and the default intensity can be jointly identified in principle. To this end, we set recovery rate as 1−y= exp ⁡(−β0), where β0>0.5 Then the CDS spread at time t for protection between t and t+τ satisfies   Stτ=∫tt+τP(t,u)E2(t,u){y−Stτ(u−⌞4u⌟4)}du14Σi=14τP(t,t+i4)E1(t,t+i4), (4) where P(t,T) is the time t price of a risk-free zero coupon bond that matures at time T, Stτ(u−⌞4u⌟4) reflects the accrued CDS premium from the previous payment date to the time of default, with ⌞4u⌟ denoting the largest integer smaller than 4u, and   E1(t,u)=EQ[ exp ⁡(−∫tu(c0+Zs)ds)|Ft], (5)  E2(t,u)=EQ[ exp ⁡(−∫tu(c0+Zs)ds)(c0+Zs)|Ft]. (6) Here, following Duffee (1999), the constant c0 is added to improve the fit to the data. The detailed formulae for E1(t,u) and E2(t,u) are presented in Appendix A. In practice, following Berndt et al. (2008) who use the midpoints between the quarterly payments, we discretize (4) as   Stτ=yΣi=14τP(t,t+2i−18)E2(t,t+2i−18)Σi=14τP(t,t+i4)E1(t,t+i4)+18Σi=14τP(t,t+2i−18)E2(t,t+2i−18). (7) In April 2009 ISDA implemented a number of CDS contract and convention changes known as the “CDS Big Bang.” After the “CDS Big Bang,” in general, CDS are quoted in upfront payments (or “upfronts”), which are the initial cash payments that compensate for the difference between a fixed coupon (100 or 500 bps) and the actual par spread. Given (7) and a fixed coupon C, the pricing formula of the upfronts at time t for protection between t and t+τ is   U(Zt,τ)=yΣi=14τP(t,t+2i−18)E2(t,t+2i−18)−C{Σi=14τP(t,t+i4)E1(t,t+i4)+18Σi=14τP(t,t+2i−18)E2(t,t+2i−18)}4. (8) The upfronts can be negative. If they are quoted as a negative then the protection buyer is paid the upfront fee by the protection seller; if the points are positive it is the other way around. Since we will work with the upfronts data in our analysis, we use expression (8). 3.2 Model Estimation In this section, we discuss the econometric method for estimating our affine model using upfronts data. When implementing the model, we first need to back out zero yields from Treasury rates to compute the prices of the risk-free zero coupon bonds P(t,T). Then, we use these zero coupon bond prices multiplied by the estimated discount factors E1(t,T) to calculate the present value of the premium and the protection leg of the CDS contracts. There are different econometric methods that one can use to estimate the affine model. Similar to Li, Wu, and Shi (2017), we use the unscented Kalman filter (UKF) in conjunction with Quasi-Maximum Likelihood Estimate (QMLE) to estimate the credit risk model. This is done because upfronts pricing formula is nonlinear in the state variable Zt (see Appendix B for details of the UKF). 3.2.1 State Space To use the UKF in empirical estimation, we recast our model in the framework of a state-space model. Although the transition density of the state variable in our model is not Gaussian, by applying the UKF with QMLE, we only need to consider the first two moments of the transition density. Therefore, we write down the transition equation as if the state variable is conditionally normally distributed, as long as the first two moments are intact. Duan and Simonato (1999) shows that this approximation is fairly efficient and accurate for estimating models with CIR type of state variables. Based on this approach, we provide the state-space representation of the defaultable term structure model below. Let Δt be the sampling interval in our study, which is a week. Then the transition equation for the default state variable Zt is given as   Et−Δt[Z(t)]=αPκP(1− exp ⁡(−κPΔt))+ exp ⁡(−κPΔt)Z(t−Δt), (9)  Vart−Δt[Z(t)]=αPσ22(κP)2(1− exp ⁡(−κPΔt))2+σ2(exp ⁡(−κPΔt)− exp ⁡(−2κPΔt))κPZ(t−Δt). (10) Let Upfrontstτ be the actual upfronts quoted at t for the protection between t and t+τ. Then the measurement equation becomes   Upfrontstτ=U(Zt,τ)+εtτ, (11) where εtτ∼i.i.d. N(0,vτ2) and τ=1,2,3,5,7,10,15,20, and 30 years. We assume the 5-year upfront is priced without errors and set v5=0. 3.2.2 Likelihood Function We assume that the 5-year upfronts are priced without error, and the measurement errors of other maturities are IID and normally distributed with zero mean. Then the transition density of   S(t)=[Upfrontst1Upfrontst2⋯Upfrontst30]⊺ (12) given the information set Ft−1 is a nine-dimensional normal distribution with mean St|t−1 and covariance matrix Pyt, which are outputs from the UKF. Thus, the transition density of S(t) can be written as   ft−1(S(t))=[(2π)9|Pyt|]−1 exp ⁡[−12(S(t)−St|t−1)⊺Pyt−1(S(t)−St|t−1)], (13) and the log-likelihood function is given by   ln⁡L∝−Σi=1nln⁡|Pyi|−Σi=1n(S(i)−Si|i−1)⊺Pyi−1(S(i)−Si|i−1), (14) where n is the sample size. In the estimation, we restrict αP to be positive to ensure the existence of the CIR process (Feller, 1951). The Q measure parameters are unconstrained. The positivity of the filtered CIR process is ensured by setting the joint likelihood of the entire time series to zero whenever the filtered CIR process is negative. Cheridito, Filipović, and Kimmel (2007) adopt a similar method to impose boundary constraints on implied state variables in their estimation. Also, to ensure the parameter estimates, especially the recovery rate parameter, are not trapped in local optimums, we pre-estimate nine sets of parameters with y fixed at 0.1 to 0.9, then use the set of estimates giving the largest log likelihood value as the starting points in the full estimation. 4. Empirical Analysis 4.1 Data The data used in our primary analysis here are from Markit. We also use GFI data for robustness tests which are presented in Appendix D. Based on the market makers’ official books of record, live indicative quotes, and clearing submissions and results, Markit creates the daily composite spreads for each CDS contract. It also provides implied ratings over time for each firm. In the article we use the implied ratings at the end of the sample period. We focus on US dollar denominated CDS contracts on all North America nonsovereign entities. We only use CDS on senior unsecured issues with modified restructuring (MR) clauses, as they are the most popular CDS contracts in the US market. To obtain accurate estimates of model parameters and to obtain enough observations for out-of-sample analysis, we require all firms included in our study to have a reasonably complete data coverage (over at least 75% daily coverage at each maturity) over the 7 years (from January 2005 to October 2011). After applying this filter, we end up with 500 firms in our sample with daily CDS spreads for maturities of 1, 2, 3, 5, 7, 10, 15, 20, and 30 years. It is also reasonable to assume that these 500 firms are the most traded contracts, i.e., their CDS contracts are the most liquid, given they have the most complete data coverage. To speed up the estimation, we only use weekly (Friday) observations among these daily data. We believe these 500 firms represent the most traded single-name CDS contracts in the US market. All the CDS spreads are converted to upfronts using the ISAD standard CDS converter, assuming the fixed coupon rate C = 100 bps and a recovery rate of 40% (See ISDA, 2009). We bootstrap zero yields from the Constant Maturity Treasury (CMT) yields in release Fed H.15. The default-free discount factors used in our upfront pricing model are the prices of zero coupon bonds computed using these zero yields.6 Descriptive tables of the data are presented in the Online Appendix. Among the six ratings (AA, A, BBB, BB, B, and CCC), AA-, BBB-, and BB-rated firms account for 87% of the 500 firms. The table also contains the distribution of the firms among the ten different sectors, which include basic materials (BM), consumer goods (CG), consumer services (CS), financials (Fin), health care (HC), industrials (Ind), oil & gas (OG), technology (Tec), telecommunication (Tel), and utilities (Uti). Fin have most firms, followed by Ind, CG, CS, OG, and Uti. Other industries have relatively fewer firms. The top three categories that have most firms are BBB-rated Ind (54), Fin (39), and OG (39). Although we estimate our upfront pricing model using the upfronts converted from the CDS spreads, we present the data summary in terms of the CDS spreads for pedagogical reasons, as CDS spreads are closer to a traditional credit risk measure. For most ratings, we see an upward sloping credit curve, which is consistent with the notion that on average default risk is larger for longer maturities. For B- and CCC-rated firms, we observe a hump-shaped curve peaking at the maturity of 5-year. This is consistent with the notion that for speculative grade bonds, the default risk can be high in the near future but if the firm survives long enough, then the default risk actually goes down. The average credit curve for most sectors also slopes upward. One prominent exception is financial firms, whose CDS spread tends to decline with maturity. This is likely due to the fact that our sample covers the financial crisis period. During the crisis many financial firms were in trouble, therefore investors likely believe that these firms had higher credit risks in the short run than in the long run. In contrast, we find that the standard deviation of CDS spread generally declines with maturity. In general, lower rated firms have higher and more volatile CDS spreads than higher rated firms. One exception is that the A-rated firms actually have lower spreads than AA-rated firms. We believe this is probably because of too few A-rated firms (only five of them in total) in our sample. 4.2 Estimation Results We estimate the credit risk model using the whole term structure of upfronts for each of the 500 firms. Table I presents the first, third quarters, median, and mean of variance ratios at different maturities. The variance ratio measures the percentage of variations of the upfronts explained by the model. The model explains the variations of most upfronts very well. At most maturities, the median variance ratios (both the full sample and first the half of the sample) are above 90%. This suggests that our one-factor model does a reasonably good job in capturing the dynamics of the term structure of upfronts. Table I Summary of variance ratios This table shows the first quarter (Q1), median, third quarter (Q3), and mean of the of the variance ratios, which are the proportion of variations of the actual upfronts explained by our upfront pricing model, at nine maturities (1 yr, 2 yr, 3 yr, 5 yr, 7 yr, 10 yr, 15 yr, 20 yr, and 30 yr) as well as the average value (Ave.) across maturities.   Full sample   First subsample     Q1  Median  Q3  Mean  Q1  Median  Q3  Mean  1yr  0.76  0.86  0.92  0.81  0.71  0.84  0.91  0.75  2yr  0.88  0.93  0.96  0.89  0.86  0.92  0.96  0.86  3yr  0.95  0.97  0.98  0.95  0.94  0.97  0.98  0.93  5yr  1.00  1.00  1.00  1.00  1.00  1.00  1.00  1.00  7yr  0.96  0.98  0.99  0.97  0.95  0.98  0.99  0.95  10yr  0.90  0.96  0.98  0.92  0.86  0.94  0.97  0.89  15yr  0.85  0.93  0.97  0.88  0.76  0.86  0.93  0.81  20yr  0.80  0.91  0.96  0.84  0.68  0.81  0.90  0.76  30yr  0.75  0.89  0.94  0.81  0.61  0.74  0.85  0.70  Ave.  0.87  0.93  0.96  0.90  0.82  0.89  0.94  0.85    Full sample   First subsample     Q1  Median  Q3  Mean  Q1  Median  Q3  Mean  1yr  0.76  0.86  0.92  0.81  0.71  0.84  0.91  0.75  2yr  0.88  0.93  0.96  0.89  0.86  0.92  0.96  0.86  3yr  0.95  0.97  0.98  0.95  0.94  0.97  0.98  0.93  5yr  1.00  1.00  1.00  1.00  1.00  1.00  1.00  1.00  7yr  0.96  0.98  0.99  0.97  0.95  0.98  0.99  0.95  10yr  0.90  0.96  0.98  0.92  0.86  0.94  0.97  0.89  15yr  0.85  0.93  0.97  0.88  0.76  0.86  0.93  0.81  20yr  0.80  0.91  0.96  0.84  0.68  0.81  0.90  0.76  30yr  0.75  0.89  0.94  0.81  0.61  0.74  0.85  0.70  Ave.  0.87  0.93  0.96  0.90  0.82  0.89  0.94  0.85  Table I Summary of variance ratios This table shows the first quarter (Q1), median, third quarter (Q3), and mean of the of the variance ratios, which are the proportion of variations of the actual upfronts explained by our upfront pricing model, at nine maturities (1 yr, 2 yr, 3 yr, 5 yr, 7 yr, 10 yr, 15 yr, 20 yr, and 30 yr) as well as the average value (Ave.) across maturities.   Full sample   First subsample     Q1  Median  Q3  Mean  Q1  Median  Q3  Mean  1yr  0.76  0.86  0.92  0.81  0.71  0.84  0.91  0.75  2yr  0.88  0.93  0.96  0.89  0.86  0.92  0.96  0.86  3yr  0.95  0.97  0.98  0.95  0.94  0.97  0.98  0.93  5yr  1.00  1.00  1.00  1.00  1.00  1.00  1.00  1.00  7yr  0.96  0.98  0.99  0.97  0.95  0.98  0.99  0.95  10yr  0.90  0.96  0.98  0.92  0.86  0.94  0.97  0.89  15yr  0.85  0.93  0.97  0.88  0.76  0.86  0.93  0.81  20yr  0.80  0.91  0.96  0.84  0.68  0.81  0.90  0.76  30yr  0.75  0.89  0.94  0.81  0.61  0.74  0.85  0.70  Ave.  0.87  0.93  0.96  0.90  0.82  0.89  0.94  0.85    Full sample   First subsample     Q1  Median  Q3  Mean  Q1  Median  Q3  Mean  1yr  0.76  0.86  0.92  0.81  0.71  0.84  0.91  0.75  2yr  0.88  0.93  0.96  0.89  0.86  0.92  0.96  0.86  3yr  0.95  0.97  0.98  0.95  0.94  0.97  0.98  0.93  5yr  1.00  1.00  1.00  1.00  1.00  1.00  1.00  1.00  7yr  0.96  0.98  0.99  0.97  0.95  0.98  0.99  0.95  10yr  0.90  0.96  0.98  0.92  0.86  0.94  0.97  0.89  15yr  0.85  0.93  0.97  0.88  0.76  0.86  0.93  0.81  20yr  0.80  0.91  0.96  0.84  0.68  0.81  0.90  0.76  30yr  0.75  0.89  0.94  0.81  0.61  0.74  0.85  0.70  Ave.  0.87  0.93  0.96  0.90  0.82  0.89  0.94  0.85  The estimation enables the model to fit the 5-year CDS upfronts as well as possible (see Section 3.2.2) which are believed to be the most liquid and therefore the most informative about the credit risk of the underlying firm.7 The estimation results in Table I confirm that the one-factor model not only fits the 5-year CDS upfronts perfectly (with the variance ratios at 5-year being virtually 100%) but that it is also able to fit the maturities around 5 years very well, for example, 2, 3, 7, 10, and 15 years. Apparently, the one-factor model fits less well at the very short and long ends of the maturity spectrum. Adding more latent factors would improve the fitting performance at both ends. However, our goal here is neither to perfectly fit the whole term structure nor to find the “best” model, but rather to explore the information content implicit in the pricing errors. Specifically, we are interested in determining whether deviations of market prices from our one-factor model indicate any profitable trading opportunities. Since our one-factor model in general captures over 90% of the variation in the term structure dynamics and fits the 5-year maturity perfectly, we use our parsimonious one-factor specification to construct a trading strategy to explore the existence and economic content of any mispricings. It is also worth noting that if a “flawed” model (e.g., only one factor) generates significantly large excess returns, then the true mispricing in the market is even more pronounced. An improved model should generate even larger returns.8 Therefore, our results should be regarded as the conservative documentation of potential mispricing. Summaries of the parameter estimates are presented in Table II. The P measure parameters αP and κP imply a mean-reverting Zt for most of the firms. The median and average long-run mean ( αP/κP) of the processes are around 1.9% and 4.4%. In contrast, the risk-neutral parameters α and κ suggest an explosive Zt for most of the firms. This explains why in most cases we observe higher CDS spreads at the long maturities. This also implies that the expected rate of default intensity is lower under the physical measure than under the risk-neutral measure, indicating that investors require a premium for bearing exposure to varying default risk. This is consistent with the similar findings in several papers (see, e.g., Duffee, 1999; Pan and Singleton, 2008; Jarrow et al., 2010; Filipović and Trolle, 2013). From the estimated recovery rate 1−y, we find that the average is 62.7%. In contrast, the average debt recovery rate measured by post-default trading prices of senior unsecured bonds as reported in Moody’s special comment of “Corporate Default and Recovery Rates” over 2005 to 2011 is around 50.9%. Therefore, our estimated recovery rate slightly overestimates Moody’s recovery rate. Indeed, the Monte Carlo simulation exercise in the Online Appendix shows that the model underestimates the loss by 12% when the true loss rate is 37%. If we take this bias into consideration and adjust the recovery rate using the correction formula provided in the Online Appendix, we find that the corrected average recovery rate is 50.9% which is exactly Moody’s estimate. Table II Summary of parameter estimates This table reports the first quarter (Q1), median, third quarter (Q3), and mean of the parameter estimates of the 500 firms used in our empirical analysis. Panel (a) is the summary of parameter estimates using the full sample (Jan2005–Oct2011), Panel (b) is the summary of parameter estimates using the first half of the sample (Jan2005–Jun2006) (a) Full sample (Jan2005–Oct2011)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.306  0.115  0.003  0.353  0.472  Median  0.002  −0.189  0.143  0.017  0.914  0.572  Q3  0.004  −0.088  0.178  0.048  1.601  0.834  Mean  0.004  −0.221  0.154  0.062  1.416  0.627    (a) Full sample (Jan2005–Oct2011)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.306  0.115  0.003  0.353  0.472  Median  0.002  −0.189  0.143  0.017  0.914  0.572  Q3  0.004  −0.088  0.178  0.048  1.601  0.834  Mean  0.004  −0.221  0.154  0.062  1.416  0.627    (b) First subsample (Jan2005–Jun2008)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.510  0.143  0.005  0.714  0.553  Median  0.002  −0.392  0.173  0.026  2.552  0.841  Q3  0.003  −0.250  0.206  0.082  4.422  0.949  Mean  0.004  −0.386  0.175  0.093  7.172  0.738    (b) First subsample (Jan2005–Jun2008)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.510  0.143  0.005  0.714  0.553  Median  0.002  −0.392  0.173  0.026  2.552  0.841  Q3  0.003  −0.250  0.206  0.082  4.422  0.949  Mean  0.004  −0.386  0.175  0.093  7.172  0.738    Table II Summary of parameter estimates This table reports the first quarter (Q1), median, third quarter (Q3), and mean of the parameter estimates of the 500 firms used in our empirical analysis. Panel (a) is the summary of parameter estimates using the full sample (Jan2005–Oct2011), Panel (b) is the summary of parameter estimates using the first half of the sample (Jan2005–Jun2006) (a) Full sample (Jan2005–Oct2011)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.306  0.115  0.003  0.353  0.472  Median  0.002  −0.189  0.143  0.017  0.914  0.572  Q3  0.004  −0.088  0.178  0.048  1.601  0.834  Mean  0.004  −0.221  0.154  0.062  1.416  0.627    (a) Full sample (Jan2005–Oct2011)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.306  0.115  0.003  0.353  0.472  Median  0.002  −0.189  0.143  0.017  0.914  0.572  Q3  0.004  −0.088  0.178  0.048  1.601  0.834  Mean  0.004  −0.221  0.154  0.062  1.416  0.627    (b) First subsample (Jan2005–Jun2008)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.510  0.143  0.005  0.714  0.553  Median  0.002  −0.392  0.173  0.026  2.552  0.841  Q3  0.003  −0.250  0.206  0.082  4.422  0.949  Mean  0.004  −0.386  0.175  0.093  7.172  0.738    (b) First subsample (Jan2005–Jun2008)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.510  0.143  0.005  0.714  0.553  Median  0.002  −0.392  0.173  0.026  2.552  0.841  Q3  0.003  −0.250  0.206  0.082  4.422  0.949  Mean  0.004  −0.386  0.175  0.093  7.172  0.738    Panel (a) of Table III presents the median of the adjusted long-term default intensity mean, which is given by c0+αPκP, in ratings and sectors. Except for the rating of CCC, where the median adjusted long-term mean is slightly lower than that of the B rating, we find that the lower the rating, the higher the adjusted long-term mean. So the results confirm that default risks are larger for lower rated firms. From this panel, we also see that the telecommunication sector has the lowest adjusted long-term mean, and that the consumer goods sector has the highest adjusted long-term mean. The ranking among the sectors is different, however, during the first half of the sample period due to probably different macroeconomic environments. The ranking among the ratings is similar in the first half of the sample period. Only the A rating seems to have an abnormally large adjusted long-term mean relative to other ratings during the first half sample. However, the A rating has only five firms, so the results might not be representative. Table III Median adjusted long-term mean of default intensity in different ratings and sectors This table reports the median of the adjusted long-term mean of the default intensities (c0 plus the long-term mean of Zt under measure P, αP/κP) of different ratings and sectors. Panel (a) is based on the full sample (Jan2005–Dec2011) estimates, Panel (b) is based the first half of the sample (Jan2005–Jun2008) estimates. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) Full sample (Jan2005–Dec2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.000  0.009  0.011  0.067  0.053  0.011  CG  0.007    0.013  0.031  0.039  0.044  0.031  CS  0.005    0.006  0.017  0.019  0.033  0.017  Fin  0.013    0.022  0.024    0.000  0.017  HC      0.013  0.026  0.029  0.017  0.021  Ind  0.009    0.014      0.001  0.009  OG    0.017  0.009  0.029    0.050  0.023  Tec      0.009  0.003  0.028  0.048  0.019  Tel  0.003    0.006  0.024      0.006  Uti  0.012  0.011  0.015    0.046  0.012  0.012  Median  0.008  0.011  0.011  0.024  0.034  0.033  0.013    (a) Full sample (Jan2005–Dec2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.000  0.009  0.011  0.067  0.053  0.011  CG  0.007    0.013  0.031  0.039  0.044  0.031  CS  0.005    0.006  0.017  0.019  0.033  0.017  Fin  0.013    0.022  0.024    0.000  0.017  HC      0.013  0.026  0.029  0.017  0.021  Ind  0.009    0.014      0.001  0.009  OG    0.017  0.009  0.029    0.050  0.023  Tec      0.009  0.003  0.028  0.048  0.019  Tel  0.003    0.006  0.024      0.006  Uti  0.012  0.011  0.015    0.046  0.012  0.012  Median  0.008  0.011  0.011  0.024  0.034  0.033  0.013    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.003  0.008  0.018  0.007  0.008  0.008  CG  0.005    0.012  0.009  0.015  0.033  0.012  CS  0.002    0.002  0.002  0.005  0.023  0.002  Fin  0.006    0.006  0.004      0.006  HC      0.004  0.016  0.044  0.009  0.012  Ind  0.013    0.011      0.011  0.011  OG    0.012  0.009  0.007    0.009  0.009  Tec      0.005  0.000  0.021    0.005  Tel  0.005    0.005  0.007      0.005  Uti  0.018  0.016  0.006    0.007  0.007  0.007  Median  0.005  0.012  0.006  0.007  0.011  0.009  0.008    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.003  0.008  0.018  0.007  0.008  0.008  CG  0.005    0.012  0.009  0.015  0.033  0.012  CS  0.002    0.002  0.002  0.005  0.023  0.002  Fin  0.006    0.006  0.004      0.006  HC      0.004  0.016  0.044  0.009  0.012  Ind  0.013    0.011      0.011  0.011  OG    0.012  0.009  0.007    0.009  0.009  Tec      0.005  0.000  0.021    0.005  Tel  0.005    0.005  0.007      0.005  Uti  0.018  0.016  0.006    0.007  0.007  0.007  Median  0.005  0.012  0.006  0.007  0.011  0.009  0.008    Table III Median adjusted long-term mean of default intensity in different ratings and sectors This table reports the median of the adjusted long-term mean of the default intensities (c0 plus the long-term mean of Zt under measure P, αP/κP) of different ratings and sectors. Panel (a) is based on the full sample (Jan2005–Dec2011) estimates, Panel (b) is based the first half of the sample (Jan2005–Jun2008) estimates. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) Full sample (Jan2005–Dec2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.000  0.009  0.011  0.067  0.053  0.011  CG  0.007    0.013  0.031  0.039  0.044  0.031  CS  0.005    0.006  0.017  0.019  0.033  0.017  Fin  0.013    0.022  0.024    0.000  0.017  HC      0.013  0.026  0.029  0.017  0.021  Ind  0.009    0.014      0.001  0.009  OG    0.017  0.009  0.029    0.050  0.023  Tec      0.009  0.003  0.028  0.048  0.019  Tel  0.003    0.006  0.024      0.006  Uti  0.012  0.011  0.015    0.046  0.012  0.012  Median  0.008  0.011  0.011  0.024  0.034  0.033  0.013    (a) Full sample (Jan2005–Dec2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.000  0.009  0.011  0.067  0.053  0.011  CG  0.007    0.013  0.031  0.039  0.044  0.031  CS  0.005    0.006  0.017  0.019  0.033  0.017  Fin  0.013    0.022  0.024    0.000  0.017  HC      0.013  0.026  0.029  0.017  0.021  Ind  0.009    0.014      0.001  0.009  OG    0.017  0.009  0.029    0.050  0.023  Tec      0.009  0.003  0.028  0.048  0.019  Tel  0.003    0.006  0.024      0.006  Uti  0.012  0.011  0.015    0.046  0.012  0.012  Median  0.008  0.011  0.011  0.024  0.034  0.033  0.013    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.003  0.008  0.018  0.007  0.008  0.008  CG  0.005    0.012  0.009  0.015  0.033  0.012  CS  0.002    0.002  0.002  0.005  0.023  0.002  Fin  0.006    0.006  0.004      0.006  HC      0.004  0.016  0.044  0.009  0.012  Ind  0.013    0.011      0.011  0.011  OG    0.012  0.009  0.007    0.009  0.009  Tec      0.005  0.000  0.021    0.005  Tel  0.005    0.005  0.007      0.005  Uti  0.018  0.016  0.006    0.007  0.007  0.007  Median  0.005  0.012  0.006  0.007  0.011  0.009  0.008    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.003  0.008  0.018  0.007  0.008  0.008  CG  0.005    0.012  0.009  0.015  0.033  0.012  CS  0.002    0.002  0.002  0.005  0.023  0.002  Fin  0.006    0.006  0.004      0.006  HC      0.004  0.016  0.044  0.009  0.012  Ind  0.013    0.011      0.011  0.011  OG    0.012  0.009  0.007    0.009  0.009  Tec      0.005  0.000  0.021    0.005  Tel  0.005    0.005  0.007      0.005  Uti  0.018  0.016  0.006    0.007  0.007  0.007  Median  0.005  0.012  0.006  0.007  0.011  0.009  0.008    Panel (a) of Table IV presents the median of the recovery rates across ratings and sectors. For ratings, the estimated recovery rates range from 41% (CCC) to 93% (A). Except for the A rating, we find a monotone relation between the median recovery rates and the ratings, i.e., the higher the rating the larger the median recovery. This is consistent with the intuition that the firms with higher ratings typically have larger recovery rates. The deviation of the median recovery rates among different sectors is smaller, as the range is from 46% (CG) to 68% (Ind). Again, the interpretation of these estimated recovery rates comes with a caveat that although the relative ranking among the recovery rates is accurate, the absolute level of the recovery rates is somewhat overestimated. According to the correction formula in the Online Appendix, the recovery rates of 41% (CCC), 93% (A), 46% (CG), and 68% (Ind) should be 33% (CCC), 76% (A), 37% (CG), and 55% (Ind), respectively. Table IV Median recovery rate in different ratings and sectors This table reports the median of the recovery rates in different ratings and sectors. Panel (a) is based on the full sample (Jan2005–Oct2011) estimates, Panel (b) is based the first half of the sample (Jan2005–Jun2008) estimates. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) Full sample (Jan2005–Oct2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.925  0.709  0.421  0.471  0.467  0.471  CG  0.880    0.794  0.460  0.410  0.416  0.460  CS  0.869    0.755  0.501  0.456  0.412  0.501  Fin  0.571    0.499  0.416    0.427  0.463  HC      0.831  0.547  0.513  0.436  0.530  Ind  0.888    0.684      0.408  0.684  OG    0.931  0.532  0.468    0.401  0.500  Tec      0.726  0.598  0.389  0.401  0.499  Tel  0.871    0.647  0.472      0.647  Uti  0.905  0.857  0.543    0.451  0.272  0.543  Median  0.876  0.925  0.697  0.470  0.454  0.412  0.572    (a) Full sample (Jan2005–Oct2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.925  0.709  0.421  0.471  0.467  0.471  CG  0.880    0.794  0.460  0.410  0.416  0.460  CS  0.869    0.755  0.501  0.456  0.412  0.501  Fin  0.571    0.499  0.416    0.427  0.463  HC      0.831  0.547  0.513  0.436  0.530  Ind  0.888    0.684      0.408  0.684  OG    0.931  0.532  0.468    0.401  0.500  Tec      0.726  0.598  0.389  0.401  0.499  Tel  0.871    0.647  0.472      0.647  Uti  0.905  0.857  0.543    0.451  0.272  0.543  Median  0.876  0.925  0.697  0.470  0.454  0.412  0.572    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.940  0.919  0.503  0.459  0.508  0.508  CG  0.959    0.958  0.556  0.277  0.456  0.556  CS  0.899    0.853  0.457  0.392  0.471  0.471  Fin  0.837    0.778  0.505    0.514  0.646  HC      0.900  0.598  0.671  0.392  0.634  Ind  0.958    0.849      0.522  0.849  OG    0.973  0.947  0.389    0.225  0.668  Tec      0.945  0.737  0.338  0.497  0.617  Tel  0.896    0.607  0.309      0.607  Uti  0.975  0.972  0.872    0.012  0.147  0.872  Median  0.928  0.972  0.886  0.504  0.365  0.471  0.845    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.940  0.919  0.503  0.459  0.508  0.508  CG  0.959    0.958  0.556  0.277  0.456  0.556  CS  0.899    0.853  0.457  0.392  0.471  0.471  Fin  0.837    0.778  0.505    0.514  0.646  HC      0.900  0.598  0.671  0.392  0.634  Ind  0.958    0.849      0.522  0.849  OG    0.973  0.947  0.389    0.225  0.668  Tec      0.945  0.737  0.338  0.497  0.617  Tel  0.896    0.607  0.309      0.607  Uti  0.975  0.972  0.872    0.012  0.147  0.872  Median  0.928  0.972  0.886  0.504  0.365  0.471  0.845    Table IV Median recovery rate in different ratings and sectors This table reports the median of the recovery rates in different ratings and sectors. Panel (a) is based on the full sample (Jan2005–Oct2011) estimates, Panel (b) is based the first half of the sample (Jan2005–Jun2008) estimates. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) Full sample (Jan2005–Oct2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.925  0.709  0.421  0.471  0.467  0.471  CG  0.880    0.794  0.460  0.410  0.416  0.460  CS  0.869    0.755  0.501  0.456  0.412  0.501  Fin  0.571    0.499  0.416    0.427  0.463  HC      0.831  0.547  0.513  0.436  0.530  Ind  0.888    0.684      0.408  0.684  OG    0.931  0.532  0.468    0.401  0.500  Tec      0.726  0.598  0.389  0.401  0.499  Tel  0.871    0.647  0.472      0.647  Uti  0.905  0.857  0.543    0.451  0.272  0.543  Median  0.876  0.925  0.697  0.470  0.454  0.412  0.572    (a) Full sample (Jan2005–Oct2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.925  0.709  0.421  0.471  0.467  0.471  CG  0.880    0.794  0.460  0.410  0.416  0.460  CS  0.869    0.755  0.501  0.456  0.412  0.501  Fin  0.571    0.499  0.416    0.427  0.463  HC      0.831  0.547  0.513  0.436  0.530  Ind  0.888    0.684      0.408  0.684  OG    0.931  0.532  0.468    0.401  0.500  Tec      0.726  0.598  0.389  0.401  0.499  Tel  0.871    0.647  0.472      0.647  Uti  0.905  0.857  0.543    0.451  0.272  0.543  Median  0.876  0.925  0.697  0.470  0.454  0.412  0.572    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.940  0.919  0.503  0.459  0.508  0.508  CG  0.959    0.958  0.556  0.277  0.456  0.556  CS  0.899    0.853  0.457  0.392  0.471  0.471  Fin  0.837    0.778  0.505    0.514  0.646  HC      0.900  0.598  0.671  0.392  0.634  Ind  0.958    0.849      0.522  0.849  OG    0.973  0.947  0.389    0.225  0.668  Tec      0.945  0.737  0.338  0.497  0.617  Tel  0.896    0.607  0.309      0.607  Uti  0.975  0.972  0.872    0.012  0.147  0.872  Median  0.928  0.972  0.886  0.504  0.365  0.471  0.845    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.940  0.919  0.503  0.459  0.508  0.508  CG  0.959    0.958  0.556  0.277  0.456  0.556  CS  0.899    0.853  0.457  0.392  0.471  0.471  Fin  0.837    0.778  0.505    0.514  0.646  HC      0.900  0.598  0.671  0.392  0.634  Ind  0.958    0.849      0.522  0.849  OG    0.973  0.947  0.389    0.225  0.668  Tec      0.945  0.737  0.338  0.497  0.617  Tel  0.896    0.607  0.309      0.607  Uti  0.975  0.972  0.872    0.012  0.147  0.872  Median  0.928  0.972  0.886  0.504  0.365  0.471  0.845    Since our main focus, which is to develop a trading strategy, does not require an unbiased estimate of the recovery rate, and because our model fits the CDS upfront data reasonably well, we proceed using these parameter estimates without correction. 5. Statistical Arbitrage Strategy In this section, we first briefly review the definitions of statistical arbitrage and explain how CDS contracts are traded. We then discuss the design and implementation of the strategy. 5.1 Definitions of Statistical Arbitrage Statistical arbitrage is typically referred to as trading strategies that rely on mathematical modeling techniques seeking profit opportunities from pricing inefficiencies (see, e.g., Whistler, 2004; Pole, 2007; Avellaneda and Lee, 2010). In the academic community, statistical arbitrage is defined more rigorously and often employed to develop statistical tools for testing market efficiency (Bondarenko, 2003; Hogan et al., 2004; Jarrow et al., 2012). For example, Hogan et al. (2004) define a statistical arbitrage to satisfy four conditions: (i) it is a zero initial cost self-financing strategy,9 (ii) it has positive expected profits in the limit as time goes to infinity, (iii) the probability of a loss converges to zero, and (iv) if loss the probability is nonzero in finite time, a time-average variance converges to zero. Notice that the academic definition does not conflict with the practical one. The strategy we develop is consistent with the both definitions. First, the practical implications of our strategy make it potentially useful for hedge funds and investment banks engaging in quantitative trading in the CDS market, while we admit that the technical details involved in actual trading is not our focus and depends on an idiosyncratic set of constraints faced by the arbitrageur (Fleckenstein, Longstaff, and Lustig, 2014); second, the previously developed statistical arbitrage test can be applied to the profits generated by our strategy to examine the extent of any mispricing in the CDS market. 5.2 Trading CDS Contracts As mentioned before, after the “CDS Big Bang” the quoting convention for CDS contracts changed from quoting par spreads10 to quoting upfronts. With this quoting convention, the upfronts can be directly regarded as the price of the contract. The capital gain before transaction costs (CGb) of a trade is simply the notional amount (NA) times the positive or negative difference between the starting and ending upfronts of the trade depending on whether the trade is initially a long or a short position:   CGt+Δt=NAtIt(Upfrontt+Δt−Upfrontt), (15) where Δt is the duration of the trade; It is an indicator variable that equals 1 if the trade is initiated in a long position and −1 if it is initiated in a short position. The return of a trade is given by   rt=CGtIMt−Δt, (16) where IMt−Δt is the initial margin posted at time t−Δt. In this section, we abstract from transaction costs and focus primarily on the idea of the strategy itself while leaving the detailed discussion of transaction costs to Section 7. 5.3 Design and Implementation The idea behind our trading strategy is to exploit the predictability in the pricing residuals produced by an arbitrage-free term structure model. Economically significant predictability in the pricing residuals in government bonds and LIBOR/Swap markets has been documented in Sercu and Wu (1997) and Bali, Heidari, and Wu (2009). Here we develop a trading strategy similar to Bali, Heidari, and Wu (2009) to explore mispricings in the CDS market. We rely on our estimated CDS pricing model to extract the state variable of the individual default risk (under the risk-neutral measure) and produce pricing residuals. We construct market-neutral portfolios of CDS contracts that are immune to both the first- and second-order changes in the state variable. Then, we long (short) under (over) valued hedged portfolios. Here are the details of the strategy. We consider the second-order expansion of the upfronts pricing function U(Zt,τ) around the state variable Zt with the following first- and second-order derivatives H1(Zt,τ)=∂U(Z,τ)∂Z|Z=Zt and H2(Zt,τ)=∂2U(Z,τ)∂Z2|Z=Zt. The closed-form formulae of H1 and H2 are presented in Appendix C. Specifically, we have   U(Zt+Δt,τ)=U(Zt,τ)+H1(Zt,τ)(Zt+Δt−Zt)+12H2(Zt,τ)(Zt+Δt−Zt)2+O((Zt+Δt−Zt)3). (17) We assume for Δt= 1 week,11 U(Z^t+Δt,τ) can approximate U(Z^t+Δt,τ−Δt) well, i.e.,   U(Zt+Δt,τ−Δt)≈U(Zt+Δt,τ). (18) Then, by the above approximation and ignoring high order terms, (17) can be rewritten as   U(Zt+Δt,τ−Δt)≈U(Zt,τ)+H1(Zt,τ)(Zt+Δt−Zt)+12H2(Zt,τ)(Zt+Δt−Zt)2. (19) Denote any maturity in ( τ=1,2,3,5,7,10,15,20, and 30) by τ0, and the two adjacent maturities of τ0 by τ1 and τ2. To see how τ1 and τ2 are chosen, let us look at a few examples: if τ0=1, then τ1=2, and τ2=3; if τ0=7, then τ1=5, and τ2=10; and if τ0=30, then τ1=15, and τ2=20. Therefore, given Δt, at time t for a unit NA of a CDS contract with maturity of τ0, we employ two other CDS contracts with maturities of τ1 and τ2 to form a hedged portfolio, whose value is immune to the variation of Zt up to the second order but subject to the variation of the pricing residuals. To see this, let us recall our measurement Equation (11) in Section 3.2.1 for τ0, τ1, and τ2:   Upfrontstτi⃗︸3×1=U(Zt,τi⃗)︸3×1+εtτi⃗︸3×1, (20) where τi⃗=[τ0,τ1,τ2]⊺. Denote the portfolio weight by M(Zt,τi⃗)︸3×1. We fix the first element in M(Zt,τi⃗) at one for any τi⃗. So the future value of the portfolio at time t+Δt is given by   M(Zt,τi⃗)⊺Upfrontst+Δtτi⃗−Δt=M(Zt,τi⃗)⊺U(Zt+Δt,τi⃗−Δt)+M(Zt,τi⃗)⊺εt+Δtτi⃗−Δt≈M(Zt,τi⃗)⊺U(Zt,τi⃗)+M(Zt,τi⃗)⊺[H1(Zt,τi⃗)(Zt+Δt−Zt)+12H2(Zt,τi⃗)(Zt+Δt−Zt)2]+M(Zt,τi⃗)⊺εt+Δtτi⃗. (21) If the second and third elements of M(Zt,τi⃗) are specified such that   M(Zt,τi⃗)⊺H1(Zt,τi⃗)=0,M(Zt,τi⃗)⊺H2(Zt,τi⃗)=0, (22) i.e.,   M(Zt,τi⃗)⊺=[1,m1(Zt,τi⃗),m2(Zt,τi⃗)],m1(Zt,τi⃗)=H2(Zt,τ0)H1(Zt,τ2)−H1(Zt,τ0)H2(Zt,τ2)H1(Zt,τ1)H2(Zt,τ2)−H2(Zt,τ1)H1(Zt,τ2),m2(Zt,τi⃗)=H2(Zt,τ0)H1(Zt,τ1)−H1(Zt,τ0)H2(Zt,τ1)H1(Zt,τ2)H2(Zt,τ1)−H2(Zt,τ2)H1(Zt,τ1), (23) we have the hedged portfolio. That is   M(Zt,τi⃗)⊺Upfrontst+Δtτi⃗−Δt≈M(Zt,τi⃗)⊺U(Zt,τ)+M(Zt,τi⃗)⊺εt+Δtτi⃗=M(Zt,τi⃗)⊺Upfrontstτi⃗+M(Zt,τi⃗)⊺Δεt+Δtτi⃗, (24) where Δεt+Δtτi⃗=εt+Δtτi⃗−εtτi⃗. For each τi⃗, we can either long or short this hedged portfolio. The long (1) and short (−1) indicators are summarized in a 9×1 vector, It. Given It and M(Zt,τi⃗), the portfolio weight vector W(Zt,I,τ⃗) for all the nine CDS contracts is given by   W(Zt,It,τ⃗)︸9×1=[I1,tM⃗(Zt,τ1⃗),I2,tM⃗(Zt,τ2⃗),⋯,I9,tM⃗(Zt,τ9⃗)]︸9×9=1︸9×1, (25) where τ⃗=[1,2,3,5,7,10,15,20,30]⊺, Ii,t is the ith element in It, M⃗(Zt,τi⃗) is a 9 × 1 vector in which the elements corresponding to τi⃗ in τ⃗ are M(Zt,τi⃗), and others are zero,12 and 1 is the vector of ones. Therefore, the expected capital gain after transaction costs at time t from holding the hedged portfolio of the whole term structure of CDS contracts for a period of Δt, given It, W(Zt,It,τ⃗), and the forecast of Δεt+Δtτi⃗, Et(Δεt+Δtτ⃗), is   Et(CGt+Δt)=W(Zt,It,τ⃗)⊺Et(Upfrontst+Δtτ⃗−Δt−Upfrontstτ⃗)≈W(Zt,It,τ⃗)⊺Et(Δεt+Δtτ⃗). (26) We use the negative exponential moving average to forecast Δεt+Δtτ⃗, specifically,   Et(Δεt+Δtτ⃗)=−Σj=0n(1−2n+1)j2n+1Δεt−jΔtτ⃗,n=4. (27) This is an exponential moving average of the past 5 weeks of Δεtτ⃗. The negative exponential moving average captures three characteristics of Δεtτ⃗: (i) the conditional mean is close to zero; (ii) the speed of mean reverting is large; and (iii) the recent lags are informative in forecasting next period’s value. Since we have nine maturities, at each point in time t, It can be chosen from Σi=199!i!(9−i)!=511 different vectors of −1 and 1 combinations. Denote the entire set of these 511 vectors by I. At each point in time t, the portfolio weight W(Zt,It,τ⃗) is chosen over I to maximize Et(CGt+Δta):   W˜t={W(Zt,I˜t,τ⃗)09×1if W(Zt,I˜t,τ⃗)⊺Et(Δεt+Δtτ⃗)>0if W(Zt,I˜t,τ⃗)⊺Et(Δεt+Δtτ⃗)≤0, (28) where13 I˜t maximizes (26), i.e.,   I˜t=arg⁡max⁡It∈I{W(Zt,It,τ⃗)⊺Et(Δεt+Δtτ⃗)}. (29) A real-life example of this weight calculation can be found in the Online Appendix. The realized capital gain CGt+Δt is given by   CGt+Δt=st[W˜t⊺( Upfrontst+Δtτ⃗−Δt−Upfrontstτ⃗)], (30) where st is a scale factor and defined as   st={1L¯W˜t−Δt⊺1−S̲W˜t−Δt⊺1if 0≤W˜t−Δt⊺1≤L¯ or −S̲≤W˜t−Δt⊺1≤0if W˜t−Δt⊺1>L¯if W˜t−Δt⊺1<−S̲, and L¯ ( S̲) is the cap on the net NA for a net long (short) position. Once the model is estimated, Zt can be backed out by fitting the estimated model to the term structure of the observed Upfronts. Given the estimated model parameters and Zt, H1(Zt,τ), H2(Zt,τ), εt can be calculated for any τ and t, which then allow us to implement the above optimization procedure. In the main test, we assume that we invest $1 M as an initial deposit, i.e., IM0= $1 M which covers the initial margin for counterparty risk and collaterals for compensating the loss in case of an underlying default. With this $1M, we are allowed to short with no more than the initial deposit, $1M, and long with no more than three times the initial deposit, 3×$1M = $3M. That is L¯=3×$1M = $3 M and S̲= $1M. During the sample period, we add cash to restore the account to $1 M whenever the total money in the account is less than $1M, which ensures no margin calls. We never withdraw cash from the account during the sample period, implying that if we are consistently making profit then the returns will be lower for a given capital gain. 6. Evidence and the Economic Contents of Mispricings In this section, we first show evidence of the existence and persistence CDS market mispricings in terms of positive trading strategy returns. Then, we explore the economic contents of the mispricings by regressing the trading strategy returns onto various economic variables. 6.1 Trading Strategy Returns as a Mispricing Measure It is natural to use the returns generated from the trading strategy as a measure of mispricing in the sense of statistical arbitrage. If the strategy consistently generates positive returns, then the mispricing is persistent. We focus only on the out-of-sample performance of the strategy. We include the in-sample performance for comparison purposes in the next subsection and Section 7. Specifically, the model is estimated only using the first half of the full sample (January 2005 to June 2008); then based on the estimated model, the strategy is implemented on the second half of the full sample (July 2008 to October 2011). During each week, the weekly returns (16) are averaged cross-sectionally to form a time series of aggregate returns. We use these aggregate returns to analyze the overall mispricing in the CDS market. The larger the returns, the larger the mispricing. It can be seen from the first panel in Figure 1 that these returns are positive. The annualized return is 71.5% and annualized SR is 7.8. This is evidence supporting the existence of mispricings. The mispricing is more severe during the financial crisis period. From the second panel in Figure 1, we can see that the aggregate returns are significantly smaller when the financial crisis period (December 2007 to June 2009) is excluded.14 This is also evident in the third panel of Figure 1 where the returns cumulate faster before June 2009 than after. Indeed, without the financial crisis period, the annualized return reduces by almost half to 30.8%. However, it is worth noting that although the annualized return reduces, the annualized SR more than doubles to 16.2, indicating the mispricings persist after the financial crisis with a lower volatility. These results are robust to the choice of the holding period and different data sources. Similar results are obtained using GFI CDS data.15 The robustness tests are presented in Appendix D and the Online Appendix. Figure 1. View largeDownload slide Aggregate returns: histograms and cumulative path. The first (second) panel shows the histogram of the aggregate weekly returns from July 2008 to October 2011 (July 2009 to October 2011). The third panel shows the cumulative returns over time. All returns are weekly and the unit of x-axes in the first and second panel, and y-axis in the third panel is percent. Figure 1. View largeDownload slide Aggregate returns: histograms and cumulative path. The first (second) panel shows the histogram of the aggregate weekly returns from July 2008 to October 2011 (July 2009 to October 2011). The third panel shows the cumulative returns over time. All returns are weekly and the unit of x-axes in the first and second panel, and y-axis in the third panel is percent. 6.2 Risk Factors Explaining the Returns The previous analysis confirms the existence and persistence of mispricings. We also investigate whether the aggregate returns are correlated with certain systematic risks, especially the variance risk per the discussion in Section 2. To this end, we try to explain the aggregate returns using market-wide credit risk and liquidity risk. Specifically, we consider the following factors:16 CRDTSPRD: the weekly changes of the yield spread between Baa and Aaa bond indices. We use this series to represent the market-wide credit risk. Bali, Heidari, and Wu (2009) and Hu, Pan, and Wang (2013) use the same factor. NOISE: the weekly changes of the “Noise” measure of Hu et al. (2013). This is a market-wide liquidity measure which exploits the connection between the amount of arbitrage capital in the market and observed “noise” in US Treasury bonds. To control for sovereign default risk, business cycle risk, and systemic risk to the financial system, we also include the following variables in the regression as control variables:17 USCDS: the weekly changes of the 5-year USA sovereign CDS spreads. We use this series as a proxy for market-based sovereign default risk. NFCI: the weekly changes of National Financial Conditions Index (NFCI). This is a high-frequency proxy for business cycle risk (see Brave and Butters, 2011, 2012). MES: the weekly changes of average Marginal Expected Shortfall of 96 US banks/financial firm. We use this series as a proxy for systemic risk to the financial system. To study the relation between the aggregate returns of our trading strategy and these factors, we regress the returns on these factors. To account for the variance of the main factors, we also include the squared terms of CRDTSPRD and NOISE on the right-hand side (RHS). The regression results are reported in Table V.18 From the results of regression (1), in which the RHS has only CRDTSPRD, NOISE, and the three control variables, we see the RHS has a fairly limited explanatory power, as the adjusted R2 is less than 1% for the full sample returns; and in the second half sample, it is less than 5% for the in-sample returns and 12% for the out-of-sample returns. Moreover, the coefficients of CRDTSPRD and NOISE in regression (1) are never significant. Interestingly, the adjusted R2’s increase dramatically after adding CRDTSPRD2 and NOISE2 to the RHS: the adjusted R2 increases from 1% to 16% for the full sample returns, from 5% to 39% for the in-sample returns, and from 11% to 40% for the out-of-sample returns. Also, the coefficients of CRDTSPRD2 and NOISE2 are all highly significant in regression (2), in which the RHS has CRDTSPRD, CRDTSPRD2, NOISE, NOISE2, and the three control variables in the three samples. The statistical significance of the coefficients and the adjusted R2 remain almost the same when moving from regression (2) to regression (3), where the RHS has only CRDTSPRD2, NOISE2, and the control variables. Looking at the control variables, none of them are significant in any of the regressions and samples. The same conclusion holds when the GFI CDS data are employed (see Appendix D). Table V Results of the return factor regressions This table reports the OLS coefficients of the return factor regressions. Newey–West standard errors are in brackets. ***, **, * represent the significance levels at 1%, 5%, and 10%, respectively. The LHS variable is the aggregate returns of the trading strategy from in-sample and out-of-sample. Besides the control variables, USCDS, NFCI, and MES, the RHS variables include CRDTSPRD and NOISE for regression (1); CRDTSPRD, CRDTSPRD2, NOISE, and NOISE2 for regression (2); CRDTSPRD2 and NOISE2 for regression (3). The upper panel reports full sample (in-sample Jan 2005–Oct 2011) results, the middle panel reports in-sample (Jul 2008–Oct 2011) results, and the lower panel reports out-of-sample (Jul 2008–Oct 2011) results. To make the coefficients comparable in magnitude, CRDTSPRD2 and NOISE2 are multiplied by 1,000 and 100,000 in the regressions.     Const.  CRDTSPRD  CRDTSPRD2  NOISE  NOISE2  US CDS  NFCI  MES  Adj. R2(%)  in-sample (Jan2005–Oct2011)  (1)  0.006***  0.630  –  5.083  –  0.138  –0.314  0.562  0.68    (0.00)  (0.51)    (4.44)    (0.50)  (0.27)  (0.64)      (2)  0.005***  –0.122  0.400***  –0.929  0.911***  0.037  –0.408  0.150  16.45      (0.00)  (0.33)  (0.10)  (3.41)  (0.24)  (0.33)  (0.28)  (0.62)      (3)  0.006***  –  0.381***  –  0.890***  0.031  –0.455  0.168  17.22      (0.00)    (0.07)    (0.26)  (0.33)  (0.28)  (0.61)    in-sample (Jul2008–Oct2011)  (1)  0.011***  2.475  –  17.593  –  1.011  –1.019  0.427  4.78    (0.00)  (1.62)    (13.55)    (1.41)  (2.23)  (1.62)      (2)  0.009***  –0.842  1.682***  –7.402  3.029***  0.685  0.314  1.205  38.72      (0.00)  (0.83)  (0.25)  (9.88)  (0.54)  (1.26)  (2.60)  (1.91)      (3)  0.009***  –  1.541***  –  2.829***  0.663  –0.227  1.085  38.89      (0.00)    (0.29)    (0.63)  (1.26)  (2.34)  (1.83)    out-of-sample (Jul2008–Oct2011)  (1)  0.012***  3.910**  –  28.887*  –  2.245  –1.666  –0.233  11.54    (0.00)  (1.78)    (15.35)    (2.12)  (2.66)  (1.87)      (2)  0.009***  0.167  1.904***  0.831  3.392***  1.881  –0.168  0.657  40.38      (0.00)  (1.00)  (0.28)  (15.02)  (0.76)  (1.59)  (2.40)  (1.86)      (3)  0.009***  –  1.931***  –  3.422***  1.880  –0.079  0.680  41.21      (0.00)    (0.25)    (0.91)  (1.66)  (2.11)  (1.83)        Const.  CRDTSPRD  CRDTSPRD2  NOISE  NOISE2  US CDS  NFCI  MES  Adj. R2(%)  in-sample (Jan2005–Oct2011)  (1)  0.006***  0.630  –  5.083  –  0.138  –0.314  0.562  0.68    (0.00)  (0.51)    (4.44)    (0.50)  (0.27)  (0.64)      (2)  0.005***  –0.122  0.400***  –0.929  0.911***  0.037  –0.408  0.150  16.45      (0.00)  (0.33)  (0.10)  (3.41)  (0.24)  (0.33)  (0.28)  (0.62)      (3)  0.006***  –  0.381***  –  0.890***  0.031  –0.455  0.168  17.22      (0.00)    (0.07)    (0.26)  (0.33)  (0.28)  (0.61)    in-sample (Jul2008–Oct2011)  (1)  0.011***  2.475  –  17.593  –  1.011  –1.019  0.427  4.78    (0.00)  (1.62)    (13.55)    (1.41)  (2.23)  (1.62)      (2)  0.009***  –0.842  1.682***  –7.402  3.029***  0.685  0.314  1.205  38.72      (0.00)  (0.83)  (0.25)  (9.88)  (0.54)  (1.26)  (2.60)  (1.91)      (3)  0.009***  –  1.541***  –  2.829***  0.663  –0.227  1.085  38.89      (0.00)    (0.29)    (0.63)  (1.26)  (2.34)  (1.83)    out-of-sample (Jul2008–Oct2011)  (1)  0.012***  3.910**  –  28.887*  –  2.245  –1.666  –0.233  11.54    (0.00)  (1.78)    (15.35)    (2.12)  (2.66)  (1.87)      (2)  0.009***  0.167  1.904***  0.831  3.392***  1.881  –0.168  0.657  40.38      (0.00)  (1.00)  (0.28)  (15.02)  (0.76)  (1.59)  (2.40)  (1.86)      (3)  0.009***  –  1.931***  –  3.422***  1.880  –0.079  0.680  41.21      (0.00)    (0.25)    (0.91)  (1.66)  (2.11)  (1.83)    Table V Results of the return factor regressions This table reports the OLS coefficients of the return factor regressions. Newey–West standard errors are in brackets. ***, **, * represent the significance levels at 1%, 5%, and 10%, respectively. The LHS variable is the aggregate returns of the trading strategy from in-sample and out-of-sample. Besides the control variables, USCDS, NFCI, and MES, the RHS variables include CRDTSPRD and NOISE for regression (1); CRDTSPRD, CRDTSPRD2, NOISE, and NOISE2 for regression (2); CRDTSPRD2 and NOISE2 for regression (3). The upper panel reports full sample (in-sample Jan 2005–Oct 2011) results, the middle panel reports in-sample (Jul 2008–Oct 2011) results, and the lower panel reports out-of-sample (Jul 2008–Oct 2011) results. To make the coefficients comparable in magnitude, CRDTSPRD2 and NOISE2 are multiplied by 1,000 and 100,000 in the regressions.     Const.  CRDTSPRD  CRDTSPRD2  NOISE  NOISE2  US CDS  NFCI  MES  Adj. R2(%)  in-sample (Jan2005–Oct2011)  (1)  0.006***  0.630  –  5.083  –  0.138  –0.314  0.562  0.68    (0.00)  (0.51)    (4.44)    (0.50)  (0.27)  (0.64)      (2)  0.005***  –0.122  0.400***  –0.929  0.911***  0.037  –0.408  0.150  16.45      (0.00)  (0.33)  (0.10)  (3.41)  (0.24)  (0.33)  (0.28)  (0.62)      (3)  0.006***  –  0.381***  –  0.890***  0.031  –0.455  0.168  17.22      (0.00)    (0.07)    (0.26)  (0.33)  (0.28)  (0.61)    in-sample (Jul2008–Oct2011)  (1)  0.011***  2.475  –  17.593  –  1.011  –1.019  0.427  4.78    (0.00)  (1.62)    (13.55)    (1.41)  (2.23)  (1.62)      (2)  0.009***  –0.842  1.682***  –7.402  3.029***  0.685  0.314  1.205  38.72      (0.00)  (0.83)  (0.25)  (9.88)  (0.54)  (1.26)  (2.60)  (1.91)      (3)  0.009***  –  1.541***  –  2.829***  0.663  –0.227  1.085  38.89      (0.00)    (0.29)    (0.63)  (1.26)  (2.34)  (1.83)    out-of-sample (Jul2008–Oct2011)  (1)  0.012***  3.910**  –  28.887*  –  2.245  –1.666  –0.233  11.54    (0.00)  (1.78)    (15.35)    (2.12)  (2.66)  (1.87)      (2)  0.009***  0.167  1.904***  0.831  3.392***  1.881  –0.168  0.657  40.38      (0.00)  (1.00)  (0.28)  (15.02)  (0.76)  (1.59)  (2.40)  (1.86)      (3)  0.009***  –  1.931***  –  3.422***  1.880  –0.079  0.680  41.21      (0.00)    (0.25)    (0.91)  (1.66)  (2.11)  (1.83)        Const.  CRDTSPRD  CRDTSPRD2  NOISE  NOISE2  US CDS  NFCI  MES  Adj. R2(%)  in-sample (Jan2005–Oct2011)  (1)  0.006***  0.630  –  5.083  –  0.138  –0.314  0.562  0.68    (0.00)  (0.51)    (4.44)    (0.50)  (0.27)  (0.64)      (2)  0.005***  –0.122  0.400***  –0.929  0.911***  0.037  –0.408  0.150  16.45      (0.00)  (0.33)  (0.10)  (3.41)  (0.24)  (0.33)  (0.28)  (0.62)      (3)  0.006***  –  0.381***  –  0.890***  0.031  –0.455  0.168  17.22      (0.00)    (0.07)    (0.26)  (0.33)  (0.28)  (0.61)    in-sample (Jul2008–Oct2011)  (1)  0.011***  2.475  –  17.593  –  1.011  –1.019  0.427  4.78    (0.00)  (1.62)    (13.55)    (1.41)  (2.23)  (1.62)      (2)  0.009***  –0.842  1.682***  –7.402  3.029***  0.685  0.314  1.205  38.72      (0.00)  (0.83)  (0.25)  (9.88)  (0.54)  (1.26)  (2.60)  (1.91)      (3)  0.009***  –  1.541***  –  2.829***  0.663  –0.227  1.085  38.89      (0.00)    (0.29)    (0.63)  (1.26)  (2.34)  (1.83)    out-of-sample (Jul2008–Oct2011)  (1)  0.012***  3.910**  –  28.887*  –  2.245  –1.666  –0.233  11.54    (0.00)  (1.78)    (15.35)    (2.12)  (2.66)  (1.87)      (2)  0.009***  0.167  1.904***  0.831  3.392***  1.881  –0.168  0.657  40.38      (0.00)  (1.00)  (0.28)  (15.02)  (0.76)  (1.59)  (2.40)  (1.86)      (3)  0.009***  –  1.931***  –  3.422***  1.880  –0.079  0.680  41.21      (0.00)    (0.25)    (0.91)  (1.66)  (2.11)  (1.83)    These results indicate that the aggregate returns of the trading strategy are significantly related to the variance of market-wide credit and liquidity risk. The significantly positive coefficients of CRDTSPRD2 and NOISE2 show that when the credit market is volatile due to either credit or liquidity risk the mispricing is pronounced. Our analysis indicates that the mispricing stem from the strong predictive power of these nonlinear factors. Our trading strategy potentially provides an indirect tool that takes advantage of predictive power. This finding also echoes the price manipulation concern in Section 2 that motivates our study. See more discussion on this in the next subsection. 6.3 Mispricing, Misspecification, or Risk Premium The fact that our trading strategy is based on pricing residuals of a reduced-form affine model might give an incorrect impression that the identified mispricing is due to model misspecification. However, we emphasize that this is not the case. We use the one-factor model to identify a common factor across different maturities and form portfolios that are immune to the changes in this factor. This is similar to Bali, Heidari, and Wu (2009)’s idea of hedging away principal interest rate factors to explore predictability in residual factors. If the abnormal returns are due to model misspecification, then a better specified model should produce less prominent returns. However, we show in the Online Appendix that when based on a two factor affine model which fits the data better, our strategy produces even higher returns. Our findings also relate to the well-known debate about whether abnormal returns are associated with mispricing or risk, see, e.g., Hirshleifer (2001) and Bloomfield and Michaely (2004).19 Although the two sides of this debate are not mutually exclusive,20 they are traditionally considered to be competing alternatives. Two recent studies, Fleckenstein, Longstaff, and Lustig (2014) and Mueller, Tahbaz-salehi, and Vedolin (2017), on trading strategies exhibiting large excess returns distinctly classify themselves into these two sides. Fleckenstein, Longstaff, and Lustig (2014) attribute the excess returns they find to TIPS-Treasury mispricing as they show that the mispricing is a violation of the law of one price and therefore cannot be reconciled with an equilibrium model of asset pricing. They further provide direct evidence supportive of the slow-moving-capital explanation of arbitrage persistence. In contrast, Mueller, Tahbaz-salehi, and Vedolin (2017) first develop a theoretical model showing that an increase in uncertainty regarding future interest rates in the USA results in higher excess returns for other currencies. Based on this model, they hypothesize that an increase in monetary policy uncertainty due to an upcoming FOMC announcement results in the depreciation of foreign currencies against the US dollar, followed by an expected appreciation in the future. In their empirical analysis, they develop a trading strategy to verify the hypothesis and attribute the significantly larger excess returns on FOMC announcement days to monetary policy uncertainty premium. Our study is consistent with Fleckenstein, Longstaff, and Lustig (2014)’s mispricing story: we develop a market neutral trading strategy using portfolios of CDS contracts that consistently delivers positive excess returns. We provide further evidence that the realized excess returns are associated with the variances/volatilities of market-wide credit and liquidity factors and have small correlation with other systematic risk factors. Since increased volatility could be associated with price manipulation per Allen, Litov, and Mei (2006) and Aggarwal and Wu (2006), we conjecture that the identified mispricing could be attributed to potential manipulative behavior in the CDS market. 7. After Fee Performance and Economic Significance In this section, we present the performance and discuss economic significance of the trading strategy taking into account transaction costs. We show that our trading strategy is potentially useful for hedge funds and investment banks engaging in quantitative trading in the CDS market. However, we emphasize that this section is not to offer a complete and practical guide to the implementation of a statistical arbitrage strategy, which depends very much on the idiosyncratic set of constraints faced by any arbitrageur, and is beyond the scope of our academic study. Similar to Fleckenstein, Longstaff, and Lustig (2014), our focal point is the identification of mispricings in the CDS market and the linkage between the mispricings and risk factors. 7.1 After Fee Performance When there are transaction costs, the trading strategy needs to be slightly revised. The revised version and the transaction cost assumptions are presented in the Online Appendix. Given the assumptions, while the Markit data still carries economically significant profitability, the GFI data, which has much lower weekly volatilities,21 fails to deliver decent performance. The analysis in this section is therefore only based on Markit data. We provide in-sample and out-of-sample analysis on the profitability of our trading strategy. We divide our data into two parts, with the first part covering January 2005 to June 2008 and the second part covering July 2008 to October 2011. In the in-sample analysis, we estimate model parameters and construct trading portfolios using data that cover the entire sample period. In the out-of-sample analysis, we estimate the model parameters using the first half of the data, implement the strategy, and compute the in-sample performance using the second half of the data. We consider three important performance measures: the annualized total excess return (XR), the annualized SR, and the maximum draw-down (MDD). XR measures how a trading strategy outperforms a benchmark investment in terms of returns, but does not adjust for the trading risks. SR measures the risk (standard deviation of excess returns) adjusted excess returns of a trading strategy. MDD measures the downside risk of a trading strategy. They are defined as:   XR={∏i=1T(1+riΔta)}52T−1−r0, a, risk-freeT52, (31)  SR=52⟨riΔta−r(i−1)Δt, w, risk-free152⟩i=1Tstd(riΔta−r(i−1)Δt, w, risk-free152)i=1T, (32)  MDD=max⁡i=1:T{max⁡j=1:i(∑k=1jCGkΔta)−∑k=1iCGkΔtamax⁡j=1:i(∑k=1jCGkΔta)+IM0}, (33) where T is the total number of weeks during the period; rta is the after transaction cost weekly return at time t and given in the Online Appendix; rt, a, risk-freeτ and rt, w, risk-freeτ are respectively the risk-free annual and weekly return for the period of τ at t; ⟨·⟩i=1T and std(·)i=1T denote respectively the average and standard deviations of the variable ·i for i=1,⋯,T; CGta is the after transaction cost weekly capital gain at time t and given in the Online Appendix. A general summary of these measures is reported in Table VI. Among the 500 firms the median and mean of the full sample XR are 15% and 17%, respectively. As for the full sample SR, the median and mean are 1.29 and 1.27, respectively. The full sample MDD is on average only 11%. In general, the full sample results show that the trading strategy makes profits in-sample from 2005 to 2011. These results carry over to the out-of-sample period, as is evident from the left panels of Table VI. During the second half of the sample (June 2008 to October 2011), out-of-sample performance is comparable to the in-sample performance with only small differences in the three measures: the median XR is 26% out-of-sample and 27% in-sample, the median SR is 1.5 out-of-sample and 1.7 in-sample, and the median MDD is 9% out-of-sample and 7% in-sample. Table VI Summary of the trading performance after fee This table reports the first quarter (Q1), median, third quarter (Q3), and mean of three measures of the trading performance: annualized total excess return (XR), annualized SR, and MDD. Panel (a) contains both the in-sample (upper subpanel) and out-of-sample (lower subpanel) results of the second half of the sample (Jul2008–Oct2011), Panel (b) contains the full sample (Jan2005–Oct2011) results. The benchmark risk-free return for XR is based on the US treasury rate. On average it is 5% for the full sample, which is roughly the return of holding a 60 to 72 months zero coupon government bond issued at the beginning 2005 till its maturity, and 3% for the second half of the sample, which is roughly the return of holding a 40 months zero coupon government bond issued in Jul 2008 till its maturity. (a) Second subsample (Jul2008–Oct2011)       Q1  Median  Q3  Mean  in-sample  XR  0.163  0.265  0.452  0.331    SR  1.350  1.662  2.045  1.642    MDD  0.045  0.072  0.125  0.097  out-of-sample  XR  0.139  0.259  0.443  0.326    SR  1.134  1.503  1.860  1.447    MDD  0.054  0.088  0.130  0.157    (a) Second subsample (Jul2008–Oct2011)       Q1  Median  Q3  Mean  in-sample  XR  0.163  0.265  0.452  0.331    SR  1.350  1.662  2.045  1.642    MDD  0.045  0.072  0.125  0.097  out-of-sample  XR  0.139  0.259  0.443  0.326    SR  1.134  1.503  1.860  1.447    MDD  0.054  0.088  0.130  0.157    (b) Full sample (Jan2005–Oct2011)     Q1  Median  Q3  Mean  XR  0.094  0.152  0.238  0.174  SR  0.979  1.293  1.574  1.269  MDD  0.045  0.068  0.111  0.107    (b) Full sample (Jan2005–Oct2011)     Q1  Median  Q3  Mean  XR  0.094  0.152  0.238  0.174  SR  0.979  1.293  1.574  1.269  MDD  0.045  0.068  0.111  0.107    Table VI Summary of the trading performance after fee This table reports the first quarter (Q1), median, third quarter (Q3), and mean of three measures of the trading performance: annualized total excess return (XR), annualized SR, and MDD. Panel (a) contains both the in-sample (upper subpanel) and out-of-sample (lower subpanel) results of the second half of the sample (Jul2008–Oct2011), Panel (b) contains the full sample (Jan2005–Oct2011) results. The benchmark risk-free return for XR is based on the US treasury rate. On average it is 5% for the full sample, which is roughly the return of holding a 60 to 72 months zero coupon government bond issued at the beginning 2005 till its maturity, and 3% for the second half of the sample, which is roughly the return of holding a 40 months zero coupon government bond issued in Jul 2008 till its maturity. (a) Second subsample (Jul2008–Oct2011)       Q1  Median  Q3  Mean  in-sample  XR  0.163  0.265  0.452  0.331    SR  1.350  1.662  2.045  1.642    MDD  0.045  0.072  0.125  0.097  out-of-sample  XR  0.139  0.259  0.443  0.326    SR  1.134  1.503  1.860  1.447    MDD  0.054  0.088  0.130  0.157    (a) Second subsample (Jul2008–Oct2011)       Q1  Median  Q3  Mean  in-sample  XR  0.163  0.265  0.452  0.331    SR  1.350  1.662  2.045  1.642    MDD  0.045  0.072  0.125  0.097  out-of-sample  XR  0.139  0.259  0.443  0.326    SR  1.134  1.503  1.860  1.447    MDD  0.054  0.088  0.130  0.157    (b) Full sample (Jan2005–Oct2011)     Q1  Median  Q3  Mean  XR  0.094  0.152  0.238  0.174  SR  0.979  1.293  1.574  1.269  MDD  0.045  0.068  0.111  0.107    (b) Full sample (Jan2005–Oct2011)     Q1  Median  Q3  Mean  XR  0.094  0.152  0.238  0.174  SR  0.979  1.293  1.574  1.269  MDD  0.045  0.068  0.111  0.107    We also break down the results of the these three measures into different ratings and sectors, and present them in Tables VII–IX. Since it is more meaningful to focus on the out-of-sample performance, we mostly discuss the out-of-sample results hereafter. In Table VII, we see a monotone increase (except for CCC) in the median out-of-sample XR’s from 18% for AA to 45% for B. As shown in Table VIII, we observe the same pattern for the median out-of-sample SR’s, they range from 1.2 for AA to 1.5 for CCC. Firms with lower ratings normally have larger probabilities of default than those with higher ratings. The same pattern also exists in the median out-of-sample MDD’s across ratings. They are all quite low, ranging from 7% for AA to 11% for CCC. This shows that the trading strategy has very little downside risk. Since our XR and SR do not adjust for credit risk, these monotone results indicate that profits from trading lowly rated firms carry a risk premium. Table VII Median annualized total excess return (XR) in different ratings and sectors This table reports the median of the annualized total excess return (XR) in different ratings and sectors in the second half of the sample (Jul2008–Oct2011). The benchmark risk-free return for XR is based on the US treasury rate. On average it is 3% for the second half of the sample, which is roughly the return of holding a 40 months zero coupon government bond issued in Jul 2008 till its maturity. Panel (a) contains the in-sample results, Panel (b) contains the out-of-sample results. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    −0.020  0.243  0.466  0.418  0.322  0.322  CG  0.214    0.239  0.375  0.773  0.544  0.375  CS  0.207    0.327  0.562  0.867  0.480  0.480  Fin  0.247    0.327  0.198    0.434  0.287  HC      0.123  0.389  0.297  0.798  0.343  Ind  0.137    0.335      0.812  0.335  OG    0.171  0.254  0.352    0.519  0.303  Tec      0.175  0.078  0.547  0.746  0.361  Tel  0.307    0.282  0.253      0.282  Uti  0.205  0.304  0.260    0.223  0.409  0.260  Median  0.210  0.171  0.257  0.363  0.483  0.519  0.265    (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    −0.020  0.243  0.466  0.418  0.322  0.322  CG  0.214    0.239  0.375  0.773  0.544  0.375  CS  0.207    0.327  0.562  0.867  0.480  0.480  Fin  0.247    0.327  0.198    0.434  0.287  HC      0.123  0.389  0.297  0.798  0.343  Ind  0.137    0.335      0.812  0.335  OG    0.171  0.254  0.352    0.519  0.303  Tec      0.175  0.078  0.547  0.746  0.361  Tel  0.307    0.282  0.253      0.282  Uti  0.205  0.304  0.260    0.223  0.409  0.260  Median  0.210  0.171  0.257  0.363  0.483  0.519  0.265    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.327  0.231  0.442  0.464  0.310  0.327  CG  0.169    0.233  0.540  0.575  0.430  0.430  CS  0.205    0.323  0.420  0.634  0.398  0.398  Fin  0.221    0.346  0.229    0.364  0.287  HC      0.122  0.466  0.245  0.967  0.355  Ind  0.126    0.333      0.691  0.333  OG    0.129  0.250  0.367    0.467  0.309  Tec      0.131  0.032  0.416  0.285  0.208  Tel  0.188    0.293  0.167      0.188  Uti  0.179  0.232  0.254    0.432  0.550  0.254  Median  0.183  0.232  0.252  0.393  0.448  0.430  0.259    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.327  0.231  0.442  0.464  0.310  0.327  CG  0.169    0.233  0.540  0.575  0.430  0.430  CS  0.205    0.323  0.420  0.634  0.398  0.398  Fin  0.221    0.346  0.229    0.364  0.287  HC      0.122  0.466  0.245  0.967  0.355  Ind  0.126    0.333      0.691  0.333  OG    0.129  0.250  0.367    0.467  0.309  Tec      0.131  0.032  0.416  0.285  0.208  Tel  0.188    0.293  0.167      0.188  Uti  0.179  0.232  0.254    0.432  0.550  0.254  Median  0.183  0.232  0.252  0.393  0.448  0.430  0.259    Table VII Median annualized total excess return (XR) in different ratings and sectors This table reports the median of the annualized total excess return (XR) in different ratings and sectors in the second half of the sample (Jul2008–Oct2011). The benchmark risk-free return for XR is based on the US treasury rate. On average it is 3% for the second half of the sample, which is roughly the return of holding a 40 months zero coupon government bond issued in Jul 2008 till its maturity. Panel (a) contains the in-sample results, Panel (b) contains the out-of-sample results. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    −0.020  0.243  0.466  0.418  0.322  0.322  CG  0.214    0.239  0.375  0.773  0.544  0.375  CS  0.207    0.327  0.562  0.867  0.480  0.480  Fin  0.247    0.327  0.198    0.434  0.287  HC      0.123  0.389  0.297  0.798  0.343  Ind  0.137    0.335      0.812  0.335  OG    0.171  0.254  0.352    0.519  0.303  Tec      0.175  0.078  0.547  0.746  0.361  Tel  0.307    0.282  0.253      0.282  Uti  0.205  0.304  0.260    0.223  0.409  0.260  Median  0.210  0.171  0.257  0.363  0.483  0.519  0.265    (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    −0.020  0.243  0.466  0.418  0.322  0.322  CG  0.214    0.239  0.375  0.773  0.544  0.375  CS  0.207    0.327  0.562  0.867  0.480  0.480  Fin  0.247    0.327  0.198    0.434  0.287  HC      0.123  0.389  0.297  0.798  0.343  Ind  0.137    0.335      0.812  0.335  OG    0.171  0.254  0.352    0.519  0.303  Tec      0.175  0.078  0.547  0.746  0.361  Tel  0.307    0.282  0.253      0.282  Uti  0.205  0.304  0.260    0.223  0.409  0.260  Median  0.210  0.171  0.257  0.363  0.483  0.519  0.265    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.327  0.231  0.442  0.464  0.310  0.327  CG  0.169    0.233  0.540  0.575  0.430  0.430  CS  0.205    0.323  0.420  0.634  0.398  0.398  Fin  0.221    0.346  0.229    0.364  0.287  HC      0.122  0.466  0.245  0.967  0.355  Ind  0.126    0.333      0.691  0.333  OG    0.129  0.250  0.367    0.467  0.309  Tec      0.131  0.032  0.416  0.285  0.208  Tel  0.188    0.293  0.167      0.188  Uti  0.179  0.232  0.254    0.432  0.550  0.254  Median  0.183  0.232  0.252  0.393  0.448  0.430  0.259    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.327  0.231  0.442  0.464  0.310  0.327  CG  0.169    0.233  0.540  0.575  0.430  0.430  CS  0.205    0.323  0.420  0.634  0.398  0.398  Fin  0.221    0.346  0.229    0.364  0.287  HC      0.122  0.466  0.245  0.967  0.355  Ind  0.126    0.333      0.691  0.333  OG    0.129  0.250  0.367    0.467  0.309  Tec      0.131  0.032  0.416  0.285  0.208  Tel  0.188    0.293  0.167      0.188  Uti  0.179  0.232  0.254    0.432  0.550  0.254  Median  0.183  0.232  0.252  0.393  0.448  0.430  0.259    Table VIII Median annualized SR in different ratings and sectors This table reports the median of the annualized SR in different ratings and sectors in the second half of the sample (Jul2008–Oct2011). Panel (a) contains the in-sample results. Panel (b) contains the out-of-sample results. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.140  1.768  1.647  2.033  1.285  1.647  CG  1.683    1.640  1.912  2.436  2.116  1.912  CS  1.498    1.686  1.984  1.768  1.712  1.712  Fin  1.508    1.702  1.659    1.784  1.680  HC      1.167  2.341  1.229  2.645  1.785  Ind  1.316    1.590      1.499  1.499  OG    1.136  1.835  1.978    2.660  1.906  Tec      1.382  1.162  1.882  2.471  1.632  Tel  1.450    1.561  2.387      1.561  Uti  1.787  1.848  1.675    1.446  1.932  1.787  Median  1.503  1.136  1.658  1.945  1.825  1.932  1.662    (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.140  1.768  1.647  2.033  1.285  1.647  CG  1.683    1.640  1.912  2.436  2.116  1.912  CS  1.498    1.686  1.984  1.768  1.712  1.712  Fin  1.508    1.702  1.659    1.784  1.680  HC      1.167  2.341  1.229  2.645  1.785  Ind  1.316    1.590      1.499  1.499  OG    1.136  1.835  1.978    2.660  1.906  Tec      1.382  1.162  1.882  2.471  1.632  Tel  1.450    1.561  2.387      1.561  Uti  1.787  1.848  1.675    1.446  1.932  1.787  Median  1.503  1.136  1.658  1.945  1.825  1.932  1.662    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    1.848  1.472  1.690  1.657  1.502  1.657  CG  −0.936    1.524  1.502  2.154  1.565  1.524  CS  0.950    1.638  1.445  1.832  1.534  1.534  Fin  1.336    1.664  1.576    1.544  1.560  HC      1.178  1.636  1.106  2.390  1.407  Ind  1.159    1.467      1.150  1.159  OG    1.066  1.599  1.867    2.034  1.733  Tec      0.925  0.737  1.610  1.183  1.054  Tel  1.534    1.491  1.541      1.534  Uti  1.632  1.236  1.635    2.049  1.708  1.635  Median  1.248  1.236  1.508  1.558  1.744  1.544  1.503    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    1.848  1.472  1.690  1.657  1.502  1.657  CG  −0.936    1.524  1.502  2.154  1.565  1.524  CS  0.950    1.638  1.445  1.832  1.534  1.534  Fin  1.336    1.664  1.576    1.544  1.560  HC      1.178  1.636  1.106  2.390  1.407  Ind  1.159    1.467      1.150  1.159  OG    1.066  1.599  1.867    2.034  1.733  Tec      0.925  0.737  1.610  1.183  1.054  Tel  1.534    1.491  1.541      1.534  Uti  1.632  1.236  1.635    2.049  1.708  1.635  Median  1.248  1.236  1.508  1.558  1.744  1.544  1.503    Table VIII Median annualized SR in different ratings and sectors This table reports the median of the annualized SR in different ratings and sectors in the second half of the sample (Jul2008–Oct2011). Panel (a) contains the in-sample results. Panel (b) contains the out-of-sample results. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.140  1.768  1.647  2.033  1.285  1.647  CG  1.683    1.640  1.912  2.436  2.116  1.912  CS  1.498    1.686  1.984  1.768  1.712  1.712  Fin  1.508    1.702  1.659    1.784  1.680  HC      1.167  2.341  1.229  2.645  1.785  Ind  1.316    1.590      1.499  1.499  OG    1.136  1.835  1.978    2.660  1.906  Tec      1.382  1.162  1.882  2.471  1.632  Tel  1.450    1.561  2.387      1.561  Uti  1.787  1.848  1.675    1.446  1.932  1.787  Median  1.503  1.136  1.658  1.945  1.825  1.932  1.662    (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.140  1.768  1.647  2.033  1.285  1.647  CG  1.683    1.640  1.912  2.436  2.116  1.912  CS  1.498    1.686  1.984  1.768  1.712  1.712  Fin  1.508    1.702  1.659    1.784  1.680  HC      1.167  2.341  1.229  2.645  1.785  Ind  1.316    1.590      1.499  1.499  OG    1.136  1.835  1.978    2.660  1.906  Tec      1.382  1.162  1.882  2.471  1.632  Tel  1.450    1.561  2.387      1.561  Uti  1.787  1.848  1.675    1.446  1.932  1.787  Median  1.503  1.136  1.658  1.945  1.825  1.932  1.662    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    1.848  1.472  1.690  1.657  1.502  1.657  CG  −0.936    1.524  1.502  2.154  1.565  1.524  CS  0.950    1.638  1.445  1.832  1.534  1.534  Fin  1.336    1.664  1.576    1.544  1.560  HC      1.178  1.636  1.106  2.390  1.407  Ind  1.159    1.467      1.150  1.159  OG    1.066  1.599  1.867    2.034  1.733  Tec      0.925  0.737  1.610  1.183  1.054  Tel  1.534    1.491  1.541      1.534  Uti  1.632  1.236  1.635    2.049  1.708  1.635  Median  1.248  1.236  1.508  1.558  1.744  1.544  1.503    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    1.848  1.472  1.690  1.657  1.502  1.657  CG  −0.936    1.524  1.502  2.154  1.565  1.524  CS  0.950    1.638  1.445  1.832  1.534  1.534  Fin  1.336    1.664  1.576    1.544  1.560  HC      1.178  1.636  1.106  2.390  1.407  Ind  1.159    1.467      1.150  1.159  OG    1.066  1.599  1.867    2.034  1.733  Tec      0.925  0.737  1.610  1.183  1.054  Tel  1.534    1.491  1.541      1.534  Uti  1.632  1.236  1.635    2.049  1.708  1.635  Median  1.248  1.236  1.508  1.558  1.744  1.544  1.503    Table IX Median MDD in different ratings and sectors This table reports the median of the MDD in different ratings and sectors in the second half of the sample (Jul2008–Oct2011). Panel (a) contains the in-sample results. Panel (b) contains the out-of-sample results. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.201  0.069  0.077  0.074  0.362  0.077  CG  0.036    0.064  0.056  0.138  0.084  0.064  CS  0.085    0.076  0.059  0.130  0.133  0.085  Fin  0.061    0.098  0.078    0.084  0.081  HC      0.056  0.054  0.174  0.086  0.071  Ind  0.105    0.084      0.104  0.104  OG    0.081  0.049  0.068    0.033  0.058  Tec      0.054  0.052  0.070  0.157  0.062  Tel  0.042    0.075  0.073      0.073  Uti  0.049  0.045  0.062    0.136  0.063  0.062  Median  0.055  0.081  0.066  0.063  0.133  0.086  0.072    (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.201  0.069  0.077  0.074  0.362  0.077  CG  0.036    0.064  0.056  0.138  0.084  0.064  CS  0.085    0.076  0.059  0.130  0.133  0.085  Fin  0.061    0.098  0.078    0.084  0.081  HC      0.056  0.054  0.174  0.086  0.071  Ind  0.105    0.084      0.104  0.104  OG    0.081  0.049  0.068    0.033  0.058  Tec      0.054  0.052  0.070  0.157  0.062  Tel  0.042    0.075  0.073      0.073  Uti  0.049  0.045  0.062    0.136  0.063  0.062  Median  0.055  0.081  0.066  0.063  0.133  0.086  0.072    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.153  0.099  0.159  0.073  0.261  0.153  CG  0.025    0.081  0.099  0.112  0.106  0.099  CS  0.081    0.080  0.137  0.109  0.108  0.108  Fin  0.074    0.102  0.060    0.114  0.088  HC      0.073  0.102  0.187  0.073  0.088  Ind  0.131    0.094      0.106  0.106  OG    0.084  0.073  0.059    0.124  0.078  Tec      0.087  0.063  0.118  0.243  0.103  Tel  0.056    0.080  0.082      0.080  Uti  0.074  0.084  0.083    0.023  0.085  0.083  Median  0.074  0.084  0.082  0.090  0.110  0.108  0.088    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.153  0.099  0.159  0.073  0.261  0.153  CG  0.025    0.081  0.099  0.112  0.106  0.099  CS  0.081    0.080  0.137  0.109  0.108  0.108  Fin  0.074    0.102  0.060    0.114  0.088  HC      0.073  0.102  0.187  0.073  0.088  Ind  0.131    0.094      0.106  0.106  OG    0.084  0.073  0.059    0.124  0.078  Tec      0.087  0.063  0.118  0.243  0.103  Tel  0.056    0.080  0.082      0.080  Uti  0.074  0.084  0.083    0.023  0.085  0.083  Median  0.074  0.084  0.082  0.090  0.110  0.108  0.088    Table IX Median MDD in different ratings and sectors This table reports the median of the MDD in different ratings and sectors in the second half of the sample (Jul2008–Oct2011). Panel (a) contains the in-sample results. Panel (b) contains the out-of-sample results. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.201  0.069  0.077  0.074  0.362  0.077  CG  0.036    0.064  0.056  0.138  0.084  0.064  CS  0.085    0.076  0.059  0.130  0.133  0.085  Fin  0.061    0.098  0.078    0.084  0.081  HC      0.056  0.054  0.174  0.086  0.071  Ind  0.105    0.084      0.104  0.104  OG    0.081  0.049  0.068    0.033  0.058  Tec      0.054  0.052  0.070  0.157  0.062  Tel  0.042    0.075  0.073      0.073  Uti  0.049  0.045  0.062    0.136  0.063  0.062  Median  0.055  0.081  0.066  0.063  0.133  0.086  0.072    (a) In-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.201  0.069  0.077  0.074  0.362  0.077  CG  0.036    0.064  0.056  0.138  0.084  0.064  CS  0.085    0.076  0.059  0.130  0.133  0.085  Fin  0.061    0.098  0.078    0.084  0.081  HC      0.056  0.054  0.174  0.086  0.071  Ind  0.105    0.084      0.104  0.104  OG    0.081  0.049  0.068    0.033  0.058  Tec      0.054  0.052  0.070  0.157  0.062  Tel  0.042    0.075  0.073      0.073  Uti  0.049  0.045  0.062    0.136  0.063  0.062  Median  0.055  0.081  0.066  0.063  0.133  0.086  0.072    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.153  0.099  0.159  0.073  0.261  0.153  CG  0.025    0.081  0.099  0.112  0.106  0.099  CS  0.081    0.080  0.137  0.109  0.108  0.108  Fin  0.074    0.102  0.060    0.114  0.088  HC      0.073  0.102  0.187  0.073  0.088  Ind  0.131    0.094      0.106  0.106  OG    0.084  0.073  0.059    0.124  0.078  Tec      0.087  0.063  0.118  0.243  0.103  Tel  0.056    0.080  0.082      0.080  Uti  0.074  0.084  0.083    0.023  0.085  0.083  Median  0.074  0.084  0.082  0.090  0.110  0.108  0.088    (b) Out-of-sample     AA  A  BBB  BB  B  CCC  Median  BM    0.153  0.099  0.159  0.073  0.261  0.153  CG  0.025    0.081  0.099  0.112  0.106  0.099  CS  0.081    0.080  0.137  0.109  0.108  0.108  Fin  0.074    0.102  0.060    0.114  0.088  HC      0.073  0.102  0.187  0.073  0.088  Ind  0.131    0.094      0.106  0.106  OG    0.084  0.073  0.059    0.124  0.078  Tec      0.087  0.063  0.118  0.243  0.103  Tel  0.056    0.080  0.082      0.080  Uti  0.074  0.084  0.083    0.023  0.085  0.083  Median  0.074  0.084  0.082  0.090  0.110  0.108  0.088    When we look at the performance across different sectors, we find that our strategy performs equally well in all sectors. For the median out-of-sample XR, Tel has the lowest XR of 19% and CG has the largest XR of 43%. For the median out-of-sample SR, Tec has the lowest SR of 1.1 and OG has the largest SR of 1.7. For the median out-of-sample MDD, OG has the lowest MDD of 8% and BM has the largest MDD of 15%. To save space, more results on the performance are contained in the Online Appendix. 7.2 The Economic Significance of the Trading Strategy To conduct an economic analysis of the trading strategy, we look at the returns and capital gains on an aggregate level. Specifically, we average (add up) the cross-sectional after fee returns (after fee capital gains) from trading the individual term structure of CDS contracts over the sample period. Therefore, the time series of the aggregate returns rArb (capital gain CG) can be viewed as the returns (profits) of a fund investing in the CDS market by applying the trading strategy to the 500 single-names. We measure the economic significance of the trading strategy using break-even performance fees with respect to ETF funds and formally test the extent of mispricing using statistical measures. To measure the economic significance of the trading strategy, we compare the performance of the dynamic trading strategy, to that of a static strategy that has the same target volatility. Following Fleming, Kirby, and Ostdiek (2001), we estimate the economic significance using a utility-based measure, which measures the performance fee Δ that an investor would be willing to pay to switch from a static strategy to a dynamic one.22 Specifically, we find the value of Δ that solves   Σt=0T−1(Rarb, t+1−Δ)−γ2(1+γ)(Rarb, t+1−Δ)2=Σt=0T−1Rsta, t+1−γ2(1+γ)Rsta, t+12, (34) where γ represents the value of the investor’s relative risk aversion; Rarb, t and Rsta, t are adjusted total returns, which have the same target volatility σtrgt from investing in the dynamic trading strategy and the static strategy, respectively. Specifically,   Rarb, t=1+(1−σtrgtσ^arb, t)rf+σtrgtσ^arb, trarb, t, (35)  Rsta, t=1+(1−σtrgtσ^sta)rf+σtrgtσ^ starsta, t, (36) where rsta, t is the gross return from a static strategy, σ^arb, t is the ex post estimate of conditional volatility of rarb, t, and σ^sta is the sample standard deviation of rsta, t.23 Here we consider three popular ETF funds in different markets (featuring the bond, equity, and volatility risks) as benchmarks: (i) iShares iBoxx Investment Grade Corporate Bond ETF (LQD), which seeks to track the investment results of an index composed of US dollar-denominated, investment grade corporate bonds and therefore represents the performance of the US corporate bond market; (ii) SPDR S&P 500 ETF (SPY), which seeks to track the investment results of the S&P 500 Index and therefore represents the overall performance of the US equity market; and (iii) iPath S&P 500 VIX ST Futures ETN (VXX), which is designed to provide exposure to the S&P 500 VIX Short-Term Futures Index Total Return and therefore represents the returns of bearing the overall volatility risk. These are the most heavily traded ETFs in the market. We obtain the historical data for these ETF funds from Yahoo Finance, and calculate the weekly gross returns (before expenses) during the period matching our out-of-sample period. Based on our out-of-sample returns and those of the ETF funds, we estimate the performance fees according to (34) for γ= 1 and 10 with σtrgt= 3% to 8%. The results are summarized in Figure 2. Our trading strategy outperforms all three benchmarks in terms of performance fees. For example, when γ = 1 and the target volatility σ trgt=6%, the investor is willing to pay a weekly performance fee of 66 bps, 70 bps, and 74 bps to switch to our trading strategy from statically trading LDQ, SPY, and VXX, respectively. It is also worth noting that the performance fees reported in Figure 2 are likely understated as the returns of the EFT funds used in the estimation are before expenses, therefore the actual performance fees might be even higher. Figure 2. View largeDownload slide Summary of the performance fee estimates. This figure shows the estimates of the weekly performance fee Δ with γ = 1 (a) and γ = 10 (b) when the target volatility varies from 3%, 4%,…, to 8%. The aggregate out-of-sample returns of the trading strategy are compared to the returns (before expenses) of the three benchmarks ETF funds: LQD, SPY, and VXX. Figure 2. View largeDownload slide Summary of the performance fee estimates. This figure shows the estimates of the weekly performance fee Δ with γ = 1 (a) and γ = 10 (b) when the target volatility varies from 3%, 4%,…, to 8%. The aggregate out-of-sample returns of the trading strategy are compared to the returns (before expenses) of the three benchmarks ETF funds: LQD, SPY, and VXX. The economic significance is further confirmed by looking at the statistical arbitrage test developed by Hogan et al. (2004). We use the Constrained Mean (CM) statistical arbitrage test as in Hogan et al. (2004) to see whether our strategy gives rise to statistical arbitrage. Specifically, the discounted capital gain after transaction cost at the ith period is assumed to be   e−y0iΔtCGiΔta=μ+ζiλxi, (37) where y0iΔt is the risk-free zero yield with maturity of iΔt at zero period; xi is an i.i.d. N(0,1) random variable; μ, ζ, and λ are unknown parameters estimated from CGiΔta using maximum likelihood. We normalize the capital gains to be the dollar-denomination profits as in Hogan et al. (2004). In other words, we divide the aggregate capital gains by 500 since the (borrowed) initial margin for each firm is one million dollars and we have 500 firms in the portfolio. The joint hypothesis test for statistical arbitrage is μ>0 and λ<0. For out-of-sample (in-sample) profits, μ^=0.008 ( μ^=0.010) with a standard error of 0.001 (0.001), and λ^=−2.43 ( λ^=−0.87) with a standard error of 0.217 (0.117). Therefore, the null hypothesis of no statistical arbitrage is strongly rejected. 8. Conclusion The rapid growth of the CDS market makes it possible to speculate on the relative pricing of a company’s credit risk across a wide range of maturities. Based on a reduced-form model of credit risk, we explore mispricing in the term structure of CDS upfronts for a large number of North American companies, its economic content, and it implications. Specifically, we estimate an affine model for the term structure of CDS upfronts of a given company and identify misvalued CDS contracts along the credit curve. We develop a strategy to trade market-neutral portfolios of misvalued CDS contracts relative to our model, betting that the misvaluation will disappear over time. The empirical analysis shows that our trading strategy generates significantly large excess returns documenting the existence of mispricing in the term structure of CDS spreads. We also find that the mispricings are positively correlated with the variance of market-wide credit and liquidity risk factors. The mispricings are more pronounced when the market is more volatile and during the financial crisis period. The fact that the extent of mispricing is related to the variance factors reinforces our concern about the pricing inefficiency of the CDS market due to market manipulation, since the market manipulation often comes with increased variance/volatility risk (see, e.g., Allen, Litov, and Mei, 2006; Aggarwal and Wu, 2006). To provide insights to the practical issues for practitioners interested in implementing the strategy, we provide some preliminary analysis of the strategy performance after transaction fees. Based on the Markit data, we show that the trading strategy is very profitable even after controlling for transaction costs. For most firms, both in-sample and out-of-sample Sharpe ratios are higher than one, and for some firms, they are even above two. Using rigorous statistical measures, we also show that the strategy is economically significant. We hope that our study will generate more research that benefits the long-run development of the CDS market. One interesting direction would be to check if the correlation structure of credit spreads has any impact on the mispricings, e.g., Luo, Li, and Ouyang (2016). Supplementary Material Supplementary data are available at Review of Finance online. Appendices A. Formulae of E1(t,u)and E2(t,u) Following Duffie, Pan, and Singleton (2000), we consider the “Transform” and the “Extended Transform” respectively below,   Ψ(w,Zt,t,u)=EQ[ exp ⁡(−∫tuc0+Zsds)ew(c0+Zu)|Ft], (A.1)  Φ(v,w,Zt,t,u)=EQ[ exp ⁡(−∫tuc0+Zsds)v(c0+Zu)ew(c0+Zu)|Ft]. (A.2) Proposition 1 of Duffie, Pan, and Singleton (2000) indicates that (A.1) has the following form:   Ψ(w,Zt,t,u)= exp ⁡{(w+t−u)c0+A(t,u;w)+B(t,u;w)Zt}, (A.3) where A and B satisfy the ODEs   B˙(t,u;w)=1+κB(t,u;w)−12B(t,u;w)2σ2,A˙(t,u;w)=−αB(t,u;w), with boundary conditions B(u,u)=w and A(u,u)=0, and   B(t,u;w)=κ−(ζ+κ−wσ2)e(u−t)ζ−(ζ−κ+wσ2)(ζ+κ−wσ2)e(u−t)ζ+(ζ−κ+wσ2)ζσ2,A(t,u;w)=α∫tuB(s,u;w)ds,ζ=2σ2+κ2.  Similarly, (A.2) is given by   Φ(v,w,Zt,t,u)=∂Ψ(φv+w,Zt,t,u)∂φ|φ=0=Ψ(w,Zt,t,u)[vc0+C(t,u;v,w)+D(t,u;v,w)Zt], (A.4) where C and D satisfy the ODEs   D˙(t,u;v,w)=κD(t,u;v,w)−12D(t,u;v,w)B(t,u;v,w)σ2,C˙(t,u;v,w)=−αD(t,u;v,w), with boundary conditions D(u,u;v,w)=v and C(u,u;v,w)=0, and   D(t,u;v,w)=vζ2−{(ζ+κ−wσ2)e(u−t)ζ−(ζ−κ+wσ2)(ζ+κ−wσ2)e(u−t)ζ+(ζ−κ+wσ2)}2vζ2ζ2−(κ−wσ2)2,C(t,u;v,w)=α∫tuD(s,u;v,w)ds. Then we have   E1(t,u)=Ψ(0,Zt,t,u), (A.5)  E2(t,u)=Φ(1,0,Zt,t,u). (A.6) B. UKF In this appendix, we briefly discuss the implementations of the UKF. More detailed discussions can be found in Harvey (1991) and Haykin et al. (2001). One challenge in applying the Kalman filter to estimate the credit risk model is that the CDS spread is a nonlinear function of the state variable. One solution to this problem, the so called extended Kalman filter (EKF), is to consider a first order Taylor expansion of the measurement equation around the predicted state Zt|t−1. Unlike the EKF, the UKF uses the exact nonlinear function Stτ(Zt) and does not linearize the measurement equation. Instead, the UKF approximates the conditional distribution of Zt using a scaled unscented transformation. The essence of the UKF (Chow, Ferrer, and Nesselroade, 2007) used in this article can be summarized briefly as follows. For each measurement occasion t, a set of deterministically selected points, termed sigma points, are used to approximate the distribution of the current state24 estimates at time t using a normal distribution with a mean vector Zt|t−1, and a covariance matrix, which is a function in the state variance PZ,t−1|t−1 (for notational clarity, we normalize the time interval to one) and conditional variance Vart−1[Z(t)]. Sigma points are specifically selected to capture the dispersion around Zt|t−1. They are then projected using the measurement function Stτ(·), weighted, and used to update the estimates in conjunction with the newly observed measurements at time t to obtain Zt|t and PZ,t|t. We start the UKF by choosing the initial values of the state variable and its variance as their steady state values:   Z0|0=αPκP, PZ,0|0=αP2(κP)2σ2. (A.7) Given Zt−1|t−1 and PZ,t−1|t−1, the ex ante prediction of the state and its variance are given by   Zt|t−1=αPκP(1− exp ⁡(−κPΔt))+ exp ⁡(−κPΔt)Zt−1|t−1, (A.8)  PZ,t|t−1=e−2κPΔtPZ,t−1|t−1+αPσ22(κP)2(1−e−κPΔt)2+σ2(e−κPΔt−e−2κPΔt)κPZt−1|t−1. (A.9) Given an ex ante prediction of state Zt|t−1, a set of three sigma points are selected as   χt|t−1=[χ0,t−1χ+,t−1χ−,t−1], (A.10) where   χ0,t−1=Zt|t−1,χ+,t−1=Zt|t−1+(1+ρ)(exp ⁡(−κPΔt)PZ,t−1|t−1+Vart−1[Z(t)]),χ−,t−1=Zt|t−1−(1+ρ)(exp ⁡(−κPΔt)PZ,t−1|t−1+Vart−1[Z(t)]). The term ρ is a scaling constant and given by   ρ=φ2(1+κ)−1, (A.11) where φ and κ are user-specified constants. In this article, we choose φ=0.001 and κ = 2. Since the values of these constants are not critical in our analysis, we do not provide more detailed discussions for brevity. Curious readers are referred to Chow, Ferrer, and Nesselroade (2007) or Chapter 7 in Haykin et al. (2001) for details. χt|t−1 is propagated through the nonlinear measurement function U(·,τ) (i.e., nonlinear transformation of the sigma points through measurement function)   St|t−1=U(χt|t−1,τ), (A.12) where the dimension of St|t−1 is 9 × 3. We define the set of weights for covariance matrix estimates as   W(c)=diag[ρ1+ρ+1−φ2+2,12(1+ρ),12(1+ρ)]; (A.13) and the weights for mean estimates as W(m)=[ρ1+ρ12(1+ρ)12(ω+ρ)]⊺. Predicted measurements and the associated covariance matrix are computed as   St|t−1=St|t−1W(m), (A.14)  Pyt=[St|t−1−11×3⊗St|t−1]W(c)[St|t−1−11×3⊗St|t−1]⊺+V, (A.15)  PZt,yt=[χt|t−1−11×3⊗Zt|t−1]W(c)[St|t−1−11×3⊗St|t−1]⊺, (A.16) where V=diag[v12,v22,⋯,v302]9×9. Finally, the discrepancy between model prediction and actual observations is weighted by a Kalman gain Ξt function to yield ex post state and variance estimates as   Zt|t=Zt|t−1+Ξt(St−St|t−1), (A.17)  PZ,t|t=PZ,t|t−1−ΞtPytΞt⊺, (A.18) where Ξt=PZt,ytPyt−1. C. Closed-form Formulae of H1 and H2 In this appendix, we derive the closed-form formulae for H1(Zt,τ)=∂U(Z,τ)∂Z|Z=Zt, and H2(Zt,τ)=∂2U(Z,τ)∂Z2|Z=Zt. Given (8), we have   H1(Zt,τ)=yΣi=14τP(t,t+2i−18)∂E2(t,t+2i−18)∂Zt−C{Σi=14τP(t,t+i4)∂E1(t,t+2i−18)∂Zt+18Σi=14τP(t,t+2i−18)∂E2(t,t+2i−18)∂Zt}4 (A.19)  H2(Zt,τ)=yΣi=14τP(t,t+2i−18)∂2E2(t,t+2i−18)∂Zt2−C{Σi=14τP(t,t+i4)∂2E1(t,t+2i−18)∂Zt2+18Σi=14τP(t,t+2i−18)∂2E2(t,t+2i−18)∂Zt2}4 (A.20) Therefore we just need to derive ∂E1(t,u)∂Zt, ∂E2(t,u)∂Zt, ∂2E1(t,u)∂Zt2, and ∂2E2(t,u)∂Zt2.  ∂E1(t,u)∂Zt=∂Ψ(0,Zt,t,u)∂Zt=Ψ(0,Zt,t,u)B(t,u;0)=E1(t,u)B(t,u;0), (A.21)  ∂E2(t,u)∂Zt=∂Φ(1,0,Zt,t,u)∂Zt=∂Ψ(0,Zt,t,u)[c0+C(t,u;1,0)+D(t,u;1,0)Zt]∂Zt={[c0+C(t,u;1,0)+D(t,u;1,0)Zt]B(t,u;0)+D(t,u;1,0)}Ψ(0,Zt,t,u)={Φ(1,0,Zt,t,u)B(t,u;0)+Ψ(0,Zt,t,u)D(t,u;1,0)}=E2(t,u)B(t,u;0)+E1(t,u)D(t,u;1,0), (A.22)  ∂2E1(t,u)∂Zt2=∂E1(t,u)∂ZtB(t,u;0), (A.23)  ∂2E2(t,u)∂Zt2=∂E2(t,u)∂ZtB(t,u;0)+∂E1(t,u)∂ZtD(t,u;1,0). (A.24) D. Results using GFI CDS Data The CDS curve data from GFI are similar to those from Markit but with fewer maturities: each curve consists of seven points for tenors from 1 to 5-year, 7-year, and 10-year. Same as in Markit data, we use CDS on senior unsecured issues with MR clauses. We require the firms to have complete data coverage over the 7 years (from January 2005 to October 2011) and end up with 200 firms in the sample with daily CDS premia on the seven tenors. All the CDS premia are converted to upfronts using ISAD standard CDS converter, assuming the fixed coupon rate C = 100 bps and recovery rate 40% (See ISDA, 2009). We bootstrap zero yields from Constant Maturity Treasury (CMT) yields from release Fed H.15. The default-free discount factor used in our upfront pricing model is prices of zero coupon bonds computed using these zero yields. In a preliminary analysis of the GFI data, we find that the one-factor model fail to adequately capture the dynamic of the upfront term structures. To better fit the data, we develop a two-factor upfront model. The details of the model can be found in the Online Appendix of this article. Given the two-factor, a two-factor strategy is also developed to ensure the portfolios of CDS contract are immune to the first two orders of the changes of two factors. The Online Appendix also has the details. Similar to the Markit data case, the two-factor model is estimated only using the first half of the full sample (January 2005 to June 2008); then based on the estimated model, the two-factor strategy is implemented on the second half of the full sample (July 2008 to October 2011). Analogous to Figure 1, the three panels in Figure A1 show the histograms and the cumulative path of the aggregate returns based on the GFI data. The annualized return is 25.2% and annualized SR is 15.6. The trading strategy again identifies the mispricing in the GFI data. While the mispricing persists regardless of the financial crisis, in the GFI data the observation of more severe mispricing during the financial crisis period is less evident. Indeed, without the financial crisis period, the annualized return only reduces by 2% to 22.7%, and the annualized SR only increases by slightly more than one to 17. The analogy of Table V is presented in Table A1. The same conclusion about the nonlinear risk factors holds in the GFI data with lower adjusted R2s and a more (less) significant coefficient for CRDTSPRD ( CRDTSPRD2) in the second (third) regression. Table A1. Results of the return factor regressions using the GFI data This table reports the OLS coefficients of the return factor regressions. Newey–West standard errors are in brackets. ***, **, * represent the significance levels at 1%, 5%, and 10%, respectively. The LHS variable is the aggregate returns of the trading strategy from in-sample and out-of-sample. Besides the control variables, USCDS, NFCI, and MES, the RHS variables include CRDTSPRD and NOISE for regression (1); CRDTSPRD, CRDTSPRD2, NOISE, and NOISE2 for regression (2); CRDTSPRD2 and NOISE2 for regression (3). The upper panel reports full sample (in-sample Jan 2005–Oct 2011) results, the middle panel reports in-sample (Jul 2008–Oct 2011) results, and the lower panel reports out-of-sample (Jul 2008–Oct 2011) results. To make the coefficients comparable in magnitude, CRDTSPRD2 and NOISE2 are multiplied by 1,000 and 100,000 in the regressions.     Const.  CRDTSPRD  CRDTSPRD2  NOISE  NOISE2  US CDS  NFCI  MES  Adj. R2(%)  in-sample (Jan2005–Oct2011)  (1)  0.003***  –0.035  –  0.074  –  –0.108  0.055  –0.078  −2.05      (0.00)  (0.11)    (2.05)    (0.20)  (0.07)  (0.36)      (2)  0.003***  –0.140  0.029  –1.291  0.243**  –0.141  0.025  –0.210  0.31      (0.00)  (0.13)  (0.04)  (1.95)  (0.11)  (0.19)  (0.07)  (0.37)      (3)  0.003***  –  0.007  –  0.217*  –0.146  –0.035  –0.192  0.66      (0.00)    (0.03)    (0.11)  (0.19)  (0.08)  (0.37)    in-sample (Jul2008–Oct2011)  (1)  0.004***  –0.168  –  0.491  –  –0.243  –0.088  0.033  −1.62      (0.00)  (0.27)    (3.99)    (0.35)  (0.33)  (0.45)      (2)  0.004***  −0.557***  0.139***  –3.421  0.559***  –0.335  0.112  0.056  10.70      (0.00)  (0.21)  (0.05)  (2.17)  (0.15)  (0.31)  (0.27)  (0.41)      (3)  0.004***  –  0.051  –  0.455***  –0.319  –0.214  –0.041  7.40      (0.00)    (0.05)    (0.17)  (0.35)  (0.25)  (0.44)    out-of-sample (Jul2008–Oct2011)  (1)  0.005***  –0.239  –  2.366  –  –0.310  –0.299  0.154  −1.63      (0.00)  (0.33)    (4.86)    (0.40)  (0.35)  (0.58)      (2)  0.004***  −0.626**  0.163***  –1.028  0.464***  –0.383  –0.119  0.212  0.56      (0.00)  (0.29)  (0.06)  (4.59)  (0.16)  (0.39)  (0.31)  (0.57)      (3)  0.004***  –  0.070  –  0.389***  –0.349  –0.406  0.123  0.38      (0.00)    (0.04)    (0.14)  (0.38)  (0.28)  (0.59)        Const.  CRDTSPRD  CRDTSPRD2  NOISE  NOISE2  US CDS  NFCI  MES  Adj. R2(%)  in-sample (Jan2005–Oct2011)  (1)  0.003***  –0.035  –  0.074  –  –0.108  0.055  –0.078  −2.05      (0.00)  (0.11)    (2.05)    (0.20)  (0.07)  (0.36)      (2)  0.003***  –0.140  0.029  –1.291  0.243**  –0.141  0.025  –0.210  0.31      (0.00)  (0.13)  (0.04)  (1.95)  (0.11)  (0.19)  (0.07)  (0.37)      (3)  0.003***  –  0.007  –  0.217*  –0.146  –0.035  –0.192  0.66      (0.00)    (0.03)    (0.11)  (0.19)  (0.08)  (0.37)    in-sample (Jul2008–Oct2011)  (1)  0.004***  –0.168  –  0.491  –  –0.243  –0.088  0.033  −1.62      (0.00)  (0.27)    (3.99)    (0.35)  (0.33)  (0.45)      (2)  0.004***  −0.557***  0.139***  –3.421  0.559***  –0.335  0.112  0.056  10.70      (0.00)  (0.21)  (0.05)  (2.17)  (0.15)  (0.31)  (0.27)  (0.41)      (3)  0.004***  –  0.051  –  0.455***  –0.319  –0.214  –0.041  7.40      (0.00)    (0.05)    (0.17)  (0.35)  (0.25)  (0.44)    out-of-sample (Jul2008–Oct2011)  (1)  0.005***  –0.239  –  2.366  –  –0.310  –0.299  0.154  −1.63      (0.00)  (0.33)    (4.86)    (0.40)  (0.35)  (0.58)      (2)  0.004***  −0.626**  0.163***  –1.028  0.464***  –0.383  –0.119  0.212  0.56      (0.00)  (0.29)  (0.06)  (4.59)  (0.16)  (0.39)  (0.31)  (0.57)      (3)  0.004***  –  0.070  –  0.389***  –0.349  –0.406  0.123  0.38      (0.00)    (0.04)    (0.14)  (0.38)  (0.28)  (0.59)    Table A1. Results of the return factor regressions using the GFI data This table reports the OLS coefficients of the return factor regressions. Newey–West standard errors are in brackets. ***, **, * represent the significance levels at 1%, 5%, and 10%, respectively. The LHS variable is the aggregate returns of the trading strategy from in-sample and out-of-sample. Besides the control variables, USCDS, NFCI, and MES, the RHS variables include CRDTSPRD and NOISE for regression (1); CRDTSPRD, CRDTSPRD2, NOISE, and NOISE2 for regression (2); CRDTSPRD2 and NOISE2 for regression (3). The upper panel reports full sample (in-sample Jan 2005–Oct 2011) results, the middle panel reports in-sample (Jul 2008–Oct 2011) results, and the lower panel reports out-of-sample (Jul 2008–Oct 2011) results. To make the coefficients comparable in magnitude, CRDTSPRD2 and NOISE2 are multiplied by 1,000 and 100,000 in the regressions.     Const.  CRDTSPRD  CRDTSPRD2  NOISE  NOISE2  US CDS  NFCI  MES  Adj. R2(%)  in-sample (Jan2005–Oct2011)  (1)  0.003***  –0.035  –  0.074  –  –0.108  0.055  –0.078  −2.05      (0.00)  (0.11)    (2.05)    (0.20)  (0.07)  (0.36)      (2)  0.003***  –0.140  0.029  –1.291  0.243**  –0.141  0.025  –0.210  0.31      (0.00)  (0.13)  (0.04)  (1.95)  (0.11)  (0.19)  (0.07)  (0.37)      (3)  0.003***  –  0.007  –  0.217*  –0.146  –0.035  –0.192  0.66      (0.00)    (0.03)    (0.11)  (0.19)  (0.08)  (0.37)    in-sample (Jul2008–Oct2011)  (1)  0.004***  –0.168  –  0.491  –  –0.243  –0.088  0.033  −1.62      (0.00)  (0.27)    (3.99)    (0.35)  (0.33)  (0.45)      (2)  0.004***  −0.557***  0.139***  –3.421  0.559***  –0.335  0.112  0.056  10.70      (0.00)  (0.21)  (0.05)  (2.17)  (0.15)  (0.31)  (0.27)  (0.41)      (3)  0.004***  –  0.051  –  0.455***  –0.319  –0.214  –0.041  7.40      (0.00)    (0.05)    (0.17)  (0.35)  (0.25)  (0.44)    out-of-sample (Jul2008–Oct2011)  (1)  0.005***  –0.239  –  2.366  –  –0.310  –0.299  0.154  −1.63      (0.00)  (0.33)    (4.86)    (0.40)  (0.35)  (0.58)      (2)  0.004***  −0.626**  0.163***  –1.028  0.464***  –0.383  –0.119  0.212  0.56      (0.00)  (0.29)  (0.06)  (4.59)  (0.16)  (0.39)  (0.31)  (0.57)      (3)  0.004***  –  0.070  –  0.389***  –0.349  –0.406  0.123  0.38      (0.00)    (0.04)    (0.14)  (0.38)  (0.28)  (0.59)        Const.  CRDTSPRD  CRDTSPRD2  NOISE  NOISE2  US CDS  NFCI  MES  Adj. R2(%)  in-sample (Jan2005–Oct2011)  (1)  0.003***  –0.035  –  0.074  –  –0.108  0.055  –0.078  −2.05      (0.00)  (0.11)    (2.05)    (0.20)  (0.07)  (0.36)      (2)  0.003***  –0.140  0.029  –1.291  0.243**  –0.141  0.025  –0.210  0.31      (0.00)  (0.13)  (0.04)  (1.95)  (0.11)  (0.19)  (0.07)  (0.37)      (3)  0.003***  –  0.007  –  0.217*  –0.146  –0.035  –0.192  0.66      (0.00)    (0.03)    (0.11)  (0.19)  (0.08)  (0.37)    in-sample (Jul2008–Oct2011)  (1)  0.004***  –0.168  –  0.491  –  –0.243  –0.088  0.033  −1.62      (0.00)  (0.27)    (3.99)    (0.35)  (0.33)  (0.45)      (2)  0.004***  −0.557***  0.139***  –3.421  0.559***  –0.335  0.112  0.056  10.70      (0.00)  (0.21)  (0.05)  (2.17)  (0.15)  (0.31)  (0.27)  (0.41)      (3)  0.004***  –  0.051  –  0.455***  –0.319  –0.214  –0.041  7.40      (0.00)    (0.05)    (0.17)  (0.35)  (0.25)  (0.44)    out-of-sample (Jul2008–Oct2011)  (1)  0.005***  –0.239  –  2.366  –  –0.310  –0.299  0.154  −1.63      (0.00)  (0.33)    (4.86)    (0.40)  (0.35)  (0.58)      (2)  0.004***  −0.626**  0.163***  –1.028  0.464***  –0.383  –0.119  0.212  0.56      (0.00)  (0.29)  (0.06)  (4.59)  (0.16)  (0.39)  (0.31)  (0.57)      (3)  0.004***  –  0.070  –  0.389***  –0.349  –0.406  0.123  0.38      (0.00)    (0.04)    (0.14)  (0.38)  (0.28)  (0.59)    Figure A1. View largeDownload slide Aggregate returns: histograms and cumulative path GFI data. The first (second) panel shows the histogram of the aggregate weekly returns from July 2008 to October 2011 (July 2009 to October 2011). The third panel shows the cumulative returns over time. All returns are weekly and the unit of x-axes in the first and second panel, and y-axis in the third panel is percent. Figure A1. View largeDownload slide Aggregate returns: histograms and cumulative path GFI data. The first (second) panel shows the histogram of the aggregate weekly returns from July 2008 to October 2011 (July 2009 to October 2011). The third panel shows the cumulative returns over time. All returns are weekly and the unit of x-axes in the first and second panel, and y-axis in the third panel is percent. Footnotes 1 See Oehmke and Zawadowski (2016) for a comprehensive discussion on the economic role of the CDS market. See Cui, Liu, and Zhang (2013) for a study on the interaction between credit spreads, asset allocation, and liquidity risk. 2 By “competitive” we mean that all traders act as price takers, that is, traders have no control on prices (Jarrow and Larsson, 2012). 3 van Deventer (2015)’s results are based on data provide by Depository Trust & Clearing Corporation (DTCC) from July 16, 2010 through December 26, 2014. DTCC started providing trading volume data only after 2010. However, the trading volume was similar during our sample period (from 2005 to 2010). 4 The twelve banks are Bank of America Corp., Barclays PLC, BNP Paribas SA, Citigroup Inc., Credit Suisse Group AG, Deutsche Bank AG, Goldman Sachs Group Inc., HSBC Holdings PLC, J.P. Morgan Chase & Co., Morgan Stanley, Royal Bank of Scotland Group PLC, and UBS Group AG. The two industry groups are the International Swaps and Derivatives Association and data provider Markit Group Ltd. 5 The positiveness of parameter β0 ensures that y∈(0,1). 6 We do not use LIBOR/Swap to construct benchmark zero yields as Dai and Singleton (2003), Duffie, Pedersen, and Singleton (2003), and Zhang (2008) because our data cover the period after 2008 in which the LIBOR manipulation was occurring. See, e.g., Abrantes-Metz et al. (2012), McConnell (2013), Fouquau and Spieser (2015). This scandal resulted in the gradual elimination after 2008 of LIBOR as a benchmark rate due to a Wall Street Journal article by Mollenkamp and Whitehouse (2008). Therefore, it is improper to use LIBOR/Swap for CDS modeling especially during our sample period. We show in the Online Appendix the main results in Section 6 are robust to using Overnight Index Swaps (OIS) and Interest Rate Swaps (IRS) as discount rates. 7 The fact that the five-year CDS are fitted perfectly does not mean that the trading strategy will not trade five-year CDS contracts. The five-year CDS contracts are traded as often as the other contracts because every contract is deployed in the trading strategy to form hedged portfolios. This will become clearer in Section 5. 8 This is supported in the robustness test in the appendix where a two-factor model is developed to fit another CDS dataset. The two-factor model significantly outperforms the one-factor model. 9 Condition (i) can be enforced by investing trading profits and borrowing trading losses and collateral (initial margin) at the risk-free rate (see Hogan et al., 2004, p. 541). This is a part of the implementation of our trading strategy. Therefore, our trading strategy fits the formal definition while fulfilling a realistic collateral setup. 10 This is the CDS spreads in the traditional sense. That is the par spread is the coupon a buyer is willing to pay on a quarterly basis over the life of the contract in exchange for protection against the default of the reference obligation, with no initial cash exchange except the cash exchanged as a collateral on the trade. 11 The accuracy of the approximation is justified using simulation results. The numerical results are given in the Online Appendix. 12 For an illustration,   M⃗(Zt,τ1⃗)=[1,m1(Zt,τ1⃗),m2(Zt,τ1⃗),0,0,0,0,0,0]⊺M⃗(Zt,τ2⃗)=[m1(Zt,τ2⃗),1,m2(Zt,τ2⃗),0,0,0,0,0,0]⊺M⃗(Zt,τ3⃗)=[0,m1(Zt,τ3⃗),1,m2(Zt,τ3⃗),0,0,0,0,0]⊺. 13 Notice that when W˜t=0 there is no trade at t, which does happen from time to time. Therefore even though the trading strategy monitors the dynamics of the CDS term structure weekly, it does not assume weekly trading in all CDS contracts. In other words, the inability to trade contracts weekly does not constrain the performance of our trading strategy in any way. 14 We use the NBER Business Cycles: http://www.nber.org/cycles.html to define the most recent financial crisis period. 15 This is also true when changing the holding period to 1 month. See the Online Appendix of this article for details. 16 CRDTSPRD are obtained from Federal Reserve Statistical Release; NOISE are generously provided by Jun Pan on her website. 17 USCDS are from Credit Market Analysis (CMA); NFCI are from Federal Reserve Bank of Chicago: https://www.chicagofed.org/publications/nfci/index; MES are generously provided by the Volatility Laboratory of the NYU Stern Volatility Institute (https://vlab.stern.nyu.edu). 18 The intercepts in the regressions should not be interpreted as alphas since the squared terms of the factors are not tradable. The purpose of running the regressions is to check if the aggregate returns are correlated with any of the factors. 19 We thank the anonymous editor for pointing out this topic. 20 Mispricing per se can be a risk that carries premium, see, e.g., Brennan and Wang (2010) and Stambaugh and Yuan (2016). Fleckenstein, Longstaff, and Lustig (2014) also admit that the difference between “mispricing” and what others call “risk premium” is a semantic one. 21 We are not in the position to discuss why the GFI data have much lower volatilities than the Markit data. The lower volatilities, nevertheless, explains why the GFI data fails to deliver a decent after fee performance: if the average weekly changes of the CDS upfronts are smaller than the assumed bid-ask spreads, then any potential profits will be shadowed by the cost. Indeed, the average weekly changes in the Markit data are of two to three times the size of the assumed bid-ask spreads, while that of the GFI data are similar to if not much smaller than the assumed bid-ask spreads. 22 For studies following this approach, see also Fleming, Kirby, and Ostdiek (2001); Marquering and Verbeek (2004); Han (2006); Della Corte, Sarno, and Thornton (2008). 23 We estimate the conditional volatility σ^arb, t using GARCH(1,1). rf is the risk-free rate and set to 3%. 24 In a typical UKF setting, both transition and measurement equations are nonlinear. Hence, the sigma points are needed to approximate the distribution of previous state estimates, to compute the ex ante predictions of state variables’ mean and variance. However, in our article, the transition equations are linear, so we can directly compute the ex ante predictions as in the classic Kalman Filter, and do not need the sigma points at this stage. References Abrantes-Metz R. M., Kraten M., Metz A. D., Seow G. S. ( 2012): Libor manipulation?, Journal of Banking & Finance  36, 136– 150. Google Scholar CrossRef Search ADS   Aggarwal R. K., Wu G. ( 2006): Stock market manipulations, The Journal of Business  79, 1915– 1953. Google Scholar CrossRef Search ADS   Allen F., Litov L., Mei J. ( 2006): Large investors, price manipulation, and limits to arbitrage: an anatomy of market corners, Review of Finance  10, 645– 693. Google Scholar CrossRef Search ADS   Arakelyan A., Serrano P. ( 2012). Liquidity in credit default swap markets. Working paper, Universidad Carlos III de Madrid. Avellaneda M., Lee J.-H. ( 2010): Statistical arbitrage in the US equities market, Quantitative Finance  10, 761– 782. Google Scholar CrossRef Search ADS   Bali T., Heidari M., Wu L. ( 2009): Predictability of interest rates and interest-rate portfolios, Journal of Business & Economic Statistics  27, 517– 527. Google Scholar CrossRef Search ADS   Berndt A., Douglas R., Duffie D., Ferguson M., Schranz D. ( 2008). Measuring default risk premia from default swap rates and EDFs . Working paper, Carnegie Mellon University. Biswas G., Nikolova S., Stahel C. W. ( 2014). The transaction costs of trading corporate credit. Available at SSRN 2532805. Blanco R., Brennan S., Marsh I. ( 2005): An empirical analysis of the dynamic relationship between investment-grade bonds and credit default swaps, The Journal of Finance  60, 2255– 2281. Google Scholar CrossRef Search ADS   Bloomfield R., Michaely R. ( 2004): Risk or mispricing? From the mouths of professionals, Financial Management  61– 81. Bondarenko O. ( 2003): Statistical arbitrage and securities prices, Review of Financial Studies  16, 875– 919. Google Scholar CrossRef Search ADS   Brave S., Butters R. A. ( 2011): Monitoring financial stability: a financial conditions index approach. Economic Perspectives  22– 43. Brave S., Butters R. A. ( 2012): Diagnosing the financial system: financial conditions and financial stress, International Journal of Central Banking  8, 191– 239. Brennan M. J., Wang A. W. ( 2010): The mispricing return premium, The Review of Financial Studies  23, 3437– 3468. Google Scholar CrossRef Search ADS   Brunnermeier M. K., Pedersen L. H. ( 2009): Funding liquidity and market liquidity, Review of Financial Studies  22, 2201– 2238. Google Scholar CrossRef Search ADS   Burne K. ( 2015): Banks near pact on swaps suit, The Wall Street Journal , September 12. Burne K. ( 2016): Swaps payout is a windfall for funds, The Wall Street Journal , January 11. Chen R.-R., Cheng X., Wu L. ( 2013): Dynamic interactions between interest-rate and credit risk: theory and evidence on the credit default swap term structure, Review of Finance  17, 403– 441. Google Scholar CrossRef Search ADS   Cheridito P., Filipović D., Kimmel R. ( 2007): Market price of risk specifications for affine models: theory and evidence, Journal of Financial Economics  83, 123– 170. Google Scholar CrossRef Search ADS   Chow S., Ferrer E., Nesselroade J. ( 2007): An unscented Kalman Filter approach to the estimation of nonlinear dynamical systems models, Multivariate Behavioral Research  42, 283– 321. Google Scholar CrossRef Search ADS   Cui C., Liu H., Zhang Y. ( 2013): On credit spread change of Chinese corporate bonds: credit risk or asset allocation effect?, China Finance Review International  3, 250– 263. Google Scholar CrossRef Search ADS   Dai Q., Singleton K. ( 2003): Term structure dynamics in theory and reality, Review of Financial Studies  16, 631– 678. Google Scholar CrossRef Search ADS   Della Corte P., Sarno L., Thornton D. L. ( 2008): The expectation hypothesis of the term structure of very short-term rates: statistical tests and economic value, Journal of Financial Economics  89, 158– 174. Google Scholar CrossRef Search ADS   Demirgüç-Kunt A. ( 2013): Re-examining the role of the state in the financial sector, Journal of Financial Stability  9, 451– 830. Google Scholar CrossRef Search ADS   Du S., Zhu H. ( 2015). Are CDS auctions biased and inefficient? Available at SSRN 1804610. Duan J., Simonato J. ( 1999): Estimating and testing exponential-affine term structure models by Kalman filter, Review of Quantitative Finance and Accounting  13, 111– 135. Google Scholar CrossRef Search ADS   Duffee G. ( 1999): Estimating the price of default risk, Review of Financial Studies  12, 197– 226. Google Scholar CrossRef Search ADS   Duffie D. ( 2010): Presidential address: asset price dynamics with slow-moving capital, The Journal of Finance  65, 1237– 1267. Google Scholar CrossRef Search ADS   Duffie D., Pan J., Singleton K. ( 2000): Transform analysis and asset pricing for affine jump-diffusions, Econometrica  68, 1343– 1376. Google Scholar CrossRef Search ADS   Duffie D., Pedersen L. H., Singleton K. J. ( 2003): Modeling sovereign yield spreads: a case study of Russian debt, The Journal of Finance  58, 119– 159. Google Scholar CrossRef Search ADS   Duffie D., Singleton K. J. ( 1997): An econometric model of the term structure of interest-rate swap yields, Journal of Finance  52, 1287– 1321. Google Scholar CrossRef Search ADS   Duffie D., Singleton K. J. ( 1999): Modeling term structures of defaultable bonds, Review of Financial Studies  12, 687– 720. Google Scholar CrossRef Search ADS   Feller W. ( 1951): Two singular diffusion problems, Annals of Mathematics  54, second series, 173– 182. Google Scholar CrossRef Search ADS   Filipović D., Trolle A. B. ( 2013): The term structure of interbank risk, Journal of Financial Economics  109, 707– 733. Google Scholar CrossRef Search ADS   Fleckenstein M., Longstaff F. A., Lustig H. ( 2014): The tips-treasury bond puzzle, The Journal of Finance  69, 2151– 2197. Google Scholar CrossRef Search ADS   Fleming J., Kirby C., Ostdiek B. ( 2001): The economic value of volatility timing, The Journal of Finance  56, 329– 352. Google Scholar CrossRef Search ADS   Fleming J., Kirby C., Ostdiek B. ( 2003): The economic value of volatility timing using “realized” volatility, Journal of Financial Economics  67, 473– 509. Google Scholar CrossRef Search ADS   Fouquau J., Spieser P. K. ( 2015): Statistical evidence about LIBOR manipulation: a “sherlock holmes” investigation, Journal of Banking & Finance  50, 632– 643. Google Scholar CrossRef Search ADS   Gibson M. S. ( 2007): Credit Derivatives and Risk Management, Economic Review-Federal Reserve Bank of Atlanta  92, 25. Han Y. ( 2006): Asset allocation with a high dimensional latent factor stochastic volatility model, Review of Financial Studies  19, 237– 271. Google Scholar CrossRef Search ADS   Harvey A. ( 1991): Forecasting, structural time series models and the Kalman filter . Cambridge University Press, UK. Haykin S. et al.   ( 2001): Kalman filtering and neural networks . Wiley, New York, pp. 221– 269. Google Scholar CrossRef Search ADS   Hirshleifer D. ( 2001): Investor psychology and asset pricing, The Journal of Finance  56, 1533– 1597. Google Scholar CrossRef Search ADS   Hogan S., Jarrow R., Teo M., Warachka M. ( 2004): Testing market efficiency using statistical arbitrage with applications to momentum and value strategies, Journal of Financial Economics  73, 525– 565. Google Scholar CrossRef Search ADS   Hu G. X., Pan J., Wang J. ( 2013): Noise as information for illiquidity, The Journal of Finance  68, 2341– 2382. Google Scholar CrossRef Search ADS   ISDA ( 2009). ISDA standard CDS converter specification. Technical report, www.isda.com. Jarrow R., Li H., Liu S., Wu C. ( 2010): Reduced-form valuation of callable corporate bonds: theory and evidence, Journal of Financial Economics  95, 227– 248. Google Scholar CrossRef Search ADS   Jarrow R., Teo M., Tse Y. K., Warachka M. ( 2012): An improved test for statistical arbitrage, Journal of Financial Markets  15, 47– 80. Google Scholar CrossRef Search ADS   Jarrow R. A. ( 1992): Market manipulation, bubbles, corners, and short squeezes, Journal of Financial and Quantitative Analysis  27, 311– 336. Google Scholar CrossRef Search ADS   Jarrow R. A. ( 2012): Problems with using CDS to infer default probabilities, Journal of Fixed Income  21, 6. Google Scholar CrossRef Search ADS   Jarrow R. A., Larsson M. ( 2012): The meaning of market efficiency, Mathematical Finance  22, 1– 30. Google Scholar CrossRef Search ADS   Levine R. ( 2012): The governance of financial regulation: reform lessons from the recent crisis, International Review of Finance  12, 39– 56. Google Scholar CrossRef Search ADS   Li H., Wu C., Shi J. ( 2017): Estimating liquidity premium of corporate bonds using the spread information in on-and off-the-run treasury securities, China Finance Review International  7, 134– 162. Google Scholar CrossRef Search ADS   Longstaff F., Mithal S., Neis E. ( 2005): Corporate yield spreads: default risk or liquidity? New evidence from the credit default swap market, Journal of Finance  55, 2213– 2253. Google Scholar CrossRef Search ADS   Luo C., Li M., Ouyang Z. ( 2016): An empirical study on the correlation structure of credit spreads based on the dynamic and pair copula functions, China Finance Review International  6, 284– 303. Google Scholar CrossRef Search ADS   Marquering W., Verbeek M. ( 2004): The economic value of predicting stock index returns and volatility, Journal of Financial and Quantitative Analysis  39, 407– 429. Google Scholar CrossRef Search ADS   Mayordomo S., Peña J. I., Romo J. ( 2014): Testing for statistical arbitrage in credit derivatives markets, Journal of Empirical Finance  26, 59– 75. Google Scholar CrossRef Search ADS   McConnell P. J. ( 2013): Systemic operational risk: the LIBOR manipulation scandal, Journal of Operational Risk  8, 59– 99. Mitchell M., Pedersen L. H., Pulvino T. ( 2007): Slow moving capital, American Economic Review  97, 215– 220. Google Scholar CrossRef Search ADS   Mollenkamp C., Whitehouse M. ( 2008). Study casts doubt on key rate: WSJ analysis suggests banks may have reported flawed interest data for libor, The Wall Street Journal, May. Mueller P., Tahbaz-salehi A., Vedolin A. ( 2017): Exchange rates and monetary policy uncertainty, The Journal of Finance  72, 1213– 1252. Google Scholar CrossRef Search ADS   Norden L., Weber M. ( 2004): Informational efficiency of credit default swap and stock markets: the impact of credit rating announcements, Journal of Banking and Finance  28, 2813– 2843. Google Scholar CrossRef Search ADS   Oehmke M., Zawadowski A. ( 2016): The anatomy of the CDS market. The Review of Financial Studies  30, 80– 119. Google Scholar CrossRef Search ADS   Pan J., Singleton K. ( 2008): Default and recovery implicit in the term structure of sovereign CDS spreads, The Journal of Finance  63, 2345– 2384. Google Scholar CrossRef Search ADS   Phillips S. M., Smith C. W. ( 1980): Trading costs for listed options: the implications for market efficiency, Journal of Financial Economics  8, 179– 201. Google Scholar CrossRef Search ADS   Pole A. ( 2007): Statistical arbitrage: algorithmic trading insights and techniques . John Wiley & Sons, New Jersey. Sercu P., Wu X. ( 1997): The information content in bond model residuals: an empirical study on the Belgian bond market, Journal of Banking & Finance  21, 685– 720. Google Scholar CrossRef Search ADS   Stambaugh R. F., Yuan Y. ( 2016): Mispricing factors, The Review of Financial Studies  30, 1270– 1315. Google Scholar CrossRef Search ADS   Stulz R. M. ( 2010): Credit default swaps and the credit crisis, The Journal of Economic Perspectives  24, 73. Google Scholar CrossRef Search ADS   van Deventer D. ( 2015): CDS trading volume for 1, 239 reference names from 2010 to 2014, Kamakura Corporate Blog, January. Whistler M. ( 2004): Trading pairs: capturing profits and hedging risk with statistical arbitrage strategies , Vol. 216, John Wiley & Sons, New Jersey. Zhang F. X. ( 2008): Market expectations and default risk premium in credit default swap prices, The Journal of Fixed Income  18, 37– 55. Google Scholar CrossRef Search ADS   Zhang G. ( 2009): Informational efficiency of credit default swap and stock markets: the impact of adverse credit events, International Review of Accounting, Banking and Finance  1, 107– 128. Zhang G., Zhang S. ( 2013): Information efficiency of the U.S. credit default swap market: evidence from earnings surprises, Journal of Financial Stability  9, 720– 730. Google Scholar CrossRef Search ADS   © The Author(s) 2018. Published by Oxford University Press on behalf of the European Finance Association. All rights reserved. For permissions, please email: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Review of Finance Oxford University Press

Exploring Mispricing in the Term Structure of CDS Spreads

Loading next page...
 
/lp/ou_press/exploring-mispricing-in-the-term-structure-of-cds-spreads-c0xmTxVQ3q
Publisher
Oxford University Press
Copyright
© The Author(s) 2018. 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/rfy014
Publisher site
See Article on Publisher Site

Abstract

Abstract Based on a reduced-form model of credit risk, we explore mispricing in the credit default swaps (CDS) spreads of North American companies and its economic content. Specifically, we develop a trading strategy using the model to trade out of sample market-neutral portfolios across the term structure of CDS contracts. Our empirical results show that the trading strategy exhibits abnormally large returns, confirming the existence and persistence of a mispricing. The aggregate returns of the trading strategy are positively related to the square of market-wide credit and liquidity risks, indicating that the mispricing is more pronounced when the market is more volatile. When implemented on the Markit data, the strategy shows significant economic value even after controlling for realistic transaction costs. 1. Introduction The credit derivatives markets have experienced tremendous growth in the last decade. According to the Bank for International Settlements (BIS), the notional value of outstanding credit derivatives peaked at the end of 2007 with $58 trillion, then dropped sharply in 2008, and gradually stabilized at around $26 trillion in recent years. The single-name credit default swaps (CDS) are believed to be the most liquid and popular product, as they account for more than two thirds of all outstanding credit derivatives. Though some exotic credit derivatives, such as subprime collateralized debt obligations (CDO) caused tremendous problems in the financial crisis, the vanilla CDS contracts play important economic roles. The newly proposed regulations, such as the establishment of a central clearing house for CDS, would help to reduce systemic risk and improve transparency in the CDS market. Therefore, the CDS contracts are likely to remain the preferred vehicle for investing, speculating, and managing single name credit risk.1 The rapid growth of the CDS market makes it possible to speculate on the relative pricing of the credit risk of a company across a wide range of maturities. Although 5-year CDS have historically been the most liquid contracts, nowadays a complete credit curve (CDS spreads over different maturities) is available for many companies. As a result, it is possible to buy and sell protections on a given firm at different maturities. An interesting question to both academics and practitioners arises as to whether the credit risk of a firm is consistently priced across maturities. From an academic perspective, an important issue is whether existing credit risk models, either structural or reduced-form, can capture the rich term structure behaviors of credit spreads. From a practical perspective, one challenging issue is whether one can design trading strategies to exploit potential mispricings along the credit curve. In two recent WSJ articles, Burne (2015, 2016) reports “Wall Street’s biggest banks have agreed to a tentative settlement over allegations that they colluded to influence the market for credit derivatives. […] The lawsuit brought by a group of investors accused the banks, the International Swaps and Derivatives Association and data provider Markit Group Ltd. of colluding to block competing providers, including exchanges, from entering the market for derivatives called credit-default swaps.” The likely settlement over these allegations casts doubt on the price discovery function and efficiency of the CDS market. Instead of directly studying the liquidity and competitiveness of the CDS market, our article provides a test of the hypothesis that the CDS market is efficient given its current liquidity and competitiveness. As pointed out by Phillips and Smith (1980) and Hogan et al. (2004), the existence of abnormal returns after taking into consideration of proper trading costs rejects market efficiency. We provide strong evidence of the existence of mispricing in our empirical section below. Using a reduced-form model of credit risk, we explore mispricing in the term structure of CDS spreads for a large number of North American companies. Specifically, we consider 500 firms with almost continuous daily observations of CDS spreads with maturities of 1, 2, 3, 5, 7, 10, 15, 20, and 30 years between January 2005 and October 2011. We estimate an affine model of credit risk for each company based on its term structure of CDS spreads and identify “mis-valued” CDS contracts relative to the model. Based on the estimated model parameters, we construct a portfolio of CDS contracts that are both delta- and gamma-neutral to the potential changes in credit spreads. Then we long (short) the portfolio if it is under (over) valued relative to our model and unwind the portfolio a week later. We conduct both in-sample and out-of-sample analysis. In the in-sample analysis, we estimate model parameters, construct trading portfolios, and calculate trading profits using all of the data. In the out-of-sample analysis, we estimate model parameters using the first half of the sample, and we construct trading portfolios and calculate trading profits using the second half of the sample. In both in-sample and out-of-sample exercises, the strategy generates significantly positive excess returns. These results, especially the out-of-sample tests, provide strong evidence of CDS mispricing. To explore the economic content of the mispricing, we average the out-of-sample returns across all firms and look at the relation between aggregate returns and various risk factors. We find that the aggregate returns are significantly and positively correlated with the square of market-wide credit and liquidity risks. This also means that the mispricing is more pronounced when the market is more volatile. Our robustness tests show the mispricing persists not only during the financial crisis period but in normal times. We do find evidence that the mispricing is more significant during the financial crisis period. Although not our main focus, we also conduct a performance analysis accounting for realistic transaction costs. The purpose of the analysis is not to offer a complete and practical guide to the implementation of a statistical arbitrage strategy, but merely to shed some light on its economic significance. This might be potentially useful: given the positive relation between aggregate returns and nonlinear risk factors, our trading strategy might provide quantitative CDS investors a means to obtain a highly profitable risk position. In this analysis, we find that our trading strategy is quite profitable: for half of the firms, the annualized Sharpe ratio (SR) is well above one in out-of-sample tests.The transaction costs included are as realistic as possible. For example, the included bid-ask spreads applying to the term structure of the CDS contracts are based on the evidence documented in recent CDS empirical studies (see Biswas, Nikolova, and Stahel, 2014; Arakelyan and Serrano, 2012). We also find that the trading strategy has significant economic value in terms of the maximum performance fee due to Fleming, Kirby, and Ostdiek (2001) with respect to popular exchange-traded funds (ETF), and the computed statistical arbitrage opportunities are statistically significant in terms of the formal test developed by Hogan et al. (2004). Our article contributes to three areas in the literature. First, our study relates to the informational efficiency of CDS markets. There are only a few studies looking at the efficiency of the CDS market. Norden and Weber (2004), Zhang (2009), and Zhang and Zhang (2013) study the informational efficiency of the CDS market by analyzing the response of CDS to rating announcements, a variety of credit events, and earnings surprises, and they find evidence supportive of an efficient market. Recently, Du and Zhu (2015) study the design of CDS auctions and find that the current design leads to biased and inefficient prices. Although these studies have shed some light on the efficiency issue, and provide some interesting insights, they do not confront the core issue. As pointed out by Stulz (2010), the ultimate way to test the CDS market is to look at “the inefficiency brought about by limits of arbitrage—because otherwise arbitrageurs could exploit any mispricing of risk.” We take the direct route by developing a trading strategy to test the pricing efficiency of the CDS market. Second, our article also contributes to the growing literature of persistent mispricing. For example, Mitchell, Pedersen, and Pulvino (2007) and Duffie (2010) discuss the role that slow-moving capital may play in allowing arbitrage opportunities to exist for extended periods of time. Fleckenstein, Longstaff, and Lustig (2014) show the persistent mispricing driven by slow-moving capital exists in the TIPS-Treasury market. Meanwhile, Brunnermeier and Pedersen (2009) show that funding availability to intermediaries in financial markets is a potential explanation for deviations of security prices from the no-arbitrage conditions. Mayordomo, Peña, and Romo (2014) study the relative pricing between CDS and Asset Swap Packages (ASPs) and find persistent mispricing; however, they only consider the relative pricing between CDS and ASPs. Our study documents persistent mispricing in the CDS market that might be due to market manipulations (see, e.g., Allen, Litov, and Mei, 2006; Aggarwal and Wu, 2006). Our third contribution lies in the credit risk literature that has arisen in the past decade. There are many empirical studies on CDS involving the modeling of the entire credit curve given the increased availability of CDS spreads for a wide range of maturities. A few studies that are closely related to ours are Zhang (2008), Pan and Singleton (2008), and Chen, Cheng, and Wu (2013), who estimate default risk models using the entire credit curve of CDS spreads. One contribution of our article is that we are among the first to explore mispricing in the term structure of CDS spreads and its economic content. The rest of this article is organized as follows. In Section 2 we motivate our study. In Section 3, we develop a one-factor affine term structure model for CDS and discuss the econometric methods for estimating the model. Section 4 discusses the data and empirical results from the estimation. In Section 5 we discuss our trading strategy. We show the evidence of mispricing, explore and interpret its economic content in Section 6. Section 7 shows the net fee performance and economic significance of the trading strategy. Section 8 concludes. Appendices and Online Appendices contain technical details and supplementary results. 2. Motivation The CDS market boomed after the Fed permitted banks to use CDSs to reduce capital reserves in 1996 (Levine, 2012). The growth of this market was exponentially fast from 2005 to 2007. The overall CDS market reached a notional value of $58 trillion in 2007. Due to this explosive growth, market participants extensively use CDS for risk management and speculation (Gibson, 2007). This CDS usage might be partially due to the promotion of the CDS market by both the academic (e.g., Longstaff, Mithal, and Neis, 2005; Blanco, Brennan, and Marsh, 2005) and regulatory communities (e.g., CDS implied default probabilities are being considered to replace credit ratings in US financial regulations, Jarrow (2012)). This rapid rise of the CDS market (unintentionally) gave rise to a common impression that the market is liquid, competitive,2 and efficient. This impression is debatable. van Deventer (2015) notes that 72.48% of the trading volume in single name CDS consists of only dealer to dealer trades.3 A problem with a dealer dominated market is that it is nearly costless for dealers to inflate gross trading volume by trading among themselves. However, this article does not directly examine the liquidity of the CDS market since there are already extensive studies in the literature looking at this issue. Instead, we take the current liquidity of this market as given and focus on market efficiency. It is reported in recent WSJ articles (Burne, 2015, 2016) that twelve big banks and two industry groups4 have tentatively agreed to pay $1.87 billion to settle allegations that they conspired to rig the CDS market. The accusations are that there was collusion among the banks and industry groups to block competing providers, e.g., exchanges from entering the CDS market. The simplest way of colluding is to manipulate the prices of CDS contracts. Indeed, the current CDS market with a majority of the players being institutional traders with large market power satisfies the classic conditions for market manipulation, see Jarrow (1992). Therefore, the pending settlement over these allegations suggests the possible existence of price manipulation in the CDS market. This brings the pricing efficiency of the market into question. As such, we explore the existence of mispricing in the CDS market and its economic content. It is evidenced in Allen, Litov, and Mei (2006) and Aggarwal and Wu (2006) that price manipulation is typically accompanied by increased volatilities. Therefore, in our empirical analysis we specifically look at the relation between mispricing and volatilities of various risks, e.g., credit and liquidity risks. 3. Model and Estimation Method 3.1 The Model In this section, we develop a one-factor affine model for the term structure of CDS spreads. We use only one factor to capture the dynamics of credit risk because our principal component analysis (PCA) shows that the first principal component captures 96% of the variations of CDS spreads. Our model is similar to that of Longstaff, Mithal, and Neis (2005), Duffie and Singleton (1999), Duffie and Singleton (1997), Duffie, Pedersen, and Singleton (2003), and Zhang (2008). We assume that credit spreads are independent of interest rates and thus avoid estimating a model for the risk-free term structure. As a robustness check, we run and obtain similar results using a two-factor affine model for the risk-free term structure, in which credit spreads are correlated with the two interest rate factors. To investigate mispricing, we equate the “correct” price to the arbitrage-free price by assuming that the market is arbitrage-free and, hence, the existence of an equivalent martingale measure Q. Formally, we assume that the state variable, i.e. the default intensity Zt, follows a square root process (Cox-Ingersoll-Ross (CIR) process) as   dZt=(α−κZt)dt+σZtdwZQ(t), (1) where wZQ(t) is a standard Brownian motion under the equivalent martingale measure Q. While we only need the dynamics of the state variable under the Q measure for pricing purposes, we need its dynamics under the P measure for econometric estimation. Given the extended affine specification for the market price of risk (Cheridito, Filipović, and Kimmel, 2007), we model the state variable’s P measure dynamics as   dZt=(αP−κPZt)dt+σZtdwZP(t). (2) The relation between the Wiener processes under the two measures is given by   wZP(t)=wZQ(t)+α−αPσ∫0t1Zsds−κ−κPσ∫0tZsds. (3) To compute the CDS spread, we assume a constant recovery rate. Since both the buyer and the seller of credit protection in a CDS are exposed to counterparty risk, the quoted recovery rates might differ from the real recovery rates implicit in the CDS spreads. Therefore, unlike the common practice in the literature which fixes the recovery rate to a predetermined constant (see, e.g., Longstaff, Mithal, and Neis, 2005; Zhang, 2008), we estimate the value of the constant recovery rate along with the model parameters from the market prices of CDS spreads. Under the fractional recovery of face value (RFV) framework, which has been widely used for pricing CDS and is consistent with the market practice, the recovery rate and the default intensity can be jointly identified in principle. To this end, we set recovery rate as 1−y= exp ⁡(−β0), where β0>0.5 Then the CDS spread at time t for protection between t and t+τ satisfies   Stτ=∫tt+τP(t,u)E2(t,u){y−Stτ(u−⌞4u⌟4)}du14Σi=14τP(t,t+i4)E1(t,t+i4), (4) where P(t,T) is the time t price of a risk-free zero coupon bond that matures at time T, Stτ(u−⌞4u⌟4) reflects the accrued CDS premium from the previous payment date to the time of default, with ⌞4u⌟ denoting the largest integer smaller than 4u, and   E1(t,u)=EQ[ exp ⁡(−∫tu(c0+Zs)ds)|Ft], (5)  E2(t,u)=EQ[ exp ⁡(−∫tu(c0+Zs)ds)(c0+Zs)|Ft]. (6) Here, following Duffee (1999), the constant c0 is added to improve the fit to the data. The detailed formulae for E1(t,u) and E2(t,u) are presented in Appendix A. In practice, following Berndt et al. (2008) who use the midpoints between the quarterly payments, we discretize (4) as   Stτ=yΣi=14τP(t,t+2i−18)E2(t,t+2i−18)Σi=14τP(t,t+i4)E1(t,t+i4)+18Σi=14τP(t,t+2i−18)E2(t,t+2i−18). (7) In April 2009 ISDA implemented a number of CDS contract and convention changes known as the “CDS Big Bang.” After the “CDS Big Bang,” in general, CDS are quoted in upfront payments (or “upfronts”), which are the initial cash payments that compensate for the difference between a fixed coupon (100 or 500 bps) and the actual par spread. Given (7) and a fixed coupon C, the pricing formula of the upfronts at time t for protection between t and t+τ is   U(Zt,τ)=yΣi=14τP(t,t+2i−18)E2(t,t+2i−18)−C{Σi=14τP(t,t+i4)E1(t,t+i4)+18Σi=14τP(t,t+2i−18)E2(t,t+2i−18)}4. (8) The upfronts can be negative. If they are quoted as a negative then the protection buyer is paid the upfront fee by the protection seller; if the points are positive it is the other way around. Since we will work with the upfronts data in our analysis, we use expression (8). 3.2 Model Estimation In this section, we discuss the econometric method for estimating our affine model using upfronts data. When implementing the model, we first need to back out zero yields from Treasury rates to compute the prices of the risk-free zero coupon bonds P(t,T). Then, we use these zero coupon bond prices multiplied by the estimated discount factors E1(t,T) to calculate the present value of the premium and the protection leg of the CDS contracts. There are different econometric methods that one can use to estimate the affine model. Similar to Li, Wu, and Shi (2017), we use the unscented Kalman filter (UKF) in conjunction with Quasi-Maximum Likelihood Estimate (QMLE) to estimate the credit risk model. This is done because upfronts pricing formula is nonlinear in the state variable Zt (see Appendix B for details of the UKF). 3.2.1 State Space To use the UKF in empirical estimation, we recast our model in the framework of a state-space model. Although the transition density of the state variable in our model is not Gaussian, by applying the UKF with QMLE, we only need to consider the first two moments of the transition density. Therefore, we write down the transition equation as if the state variable is conditionally normally distributed, as long as the first two moments are intact. Duan and Simonato (1999) shows that this approximation is fairly efficient and accurate for estimating models with CIR type of state variables. Based on this approach, we provide the state-space representation of the defaultable term structure model below. Let Δt be the sampling interval in our study, which is a week. Then the transition equation for the default state variable Zt is given as   Et−Δt[Z(t)]=αPκP(1− exp ⁡(−κPΔt))+ exp ⁡(−κPΔt)Z(t−Δt), (9)  Vart−Δt[Z(t)]=αPσ22(κP)2(1− exp ⁡(−κPΔt))2+σ2(exp ⁡(−κPΔt)− exp ⁡(−2κPΔt))κPZ(t−Δt). (10) Let Upfrontstτ be the actual upfronts quoted at t for the protection between t and t+τ. Then the measurement equation becomes   Upfrontstτ=U(Zt,τ)+εtτ, (11) where εtτ∼i.i.d. N(0,vτ2) and τ=1,2,3,5,7,10,15,20, and 30 years. We assume the 5-year upfront is priced without errors and set v5=0. 3.2.2 Likelihood Function We assume that the 5-year upfronts are priced without error, and the measurement errors of other maturities are IID and normally distributed with zero mean. Then the transition density of   S(t)=[Upfrontst1Upfrontst2⋯Upfrontst30]⊺ (12) given the information set Ft−1 is a nine-dimensional normal distribution with mean St|t−1 and covariance matrix Pyt, which are outputs from the UKF. Thus, the transition density of S(t) can be written as   ft−1(S(t))=[(2π)9|Pyt|]−1 exp ⁡[−12(S(t)−St|t−1)⊺Pyt−1(S(t)−St|t−1)], (13) and the log-likelihood function is given by   ln⁡L∝−Σi=1nln⁡|Pyi|−Σi=1n(S(i)−Si|i−1)⊺Pyi−1(S(i)−Si|i−1), (14) where n is the sample size. In the estimation, we restrict αP to be positive to ensure the existence of the CIR process (Feller, 1951). The Q measure parameters are unconstrained. The positivity of the filtered CIR process is ensured by setting the joint likelihood of the entire time series to zero whenever the filtered CIR process is negative. Cheridito, Filipović, and Kimmel (2007) adopt a similar method to impose boundary constraints on implied state variables in their estimation. Also, to ensure the parameter estimates, especially the recovery rate parameter, are not trapped in local optimums, we pre-estimate nine sets of parameters with y fixed at 0.1 to 0.9, then use the set of estimates giving the largest log likelihood value as the starting points in the full estimation. 4. Empirical Analysis 4.1 Data The data used in our primary analysis here are from Markit. We also use GFI data for robustness tests which are presented in Appendix D. Based on the market makers’ official books of record, live indicative quotes, and clearing submissions and results, Markit creates the daily composite spreads for each CDS contract. It also provides implied ratings over time for each firm. In the article we use the implied ratings at the end of the sample period. We focus on US dollar denominated CDS contracts on all North America nonsovereign entities. We only use CDS on senior unsecured issues with modified restructuring (MR) clauses, as they are the most popular CDS contracts in the US market. To obtain accurate estimates of model parameters and to obtain enough observations for out-of-sample analysis, we require all firms included in our study to have a reasonably complete data coverage (over at least 75% daily coverage at each maturity) over the 7 years (from January 2005 to October 2011). After applying this filter, we end up with 500 firms in our sample with daily CDS spreads for maturities of 1, 2, 3, 5, 7, 10, 15, 20, and 30 years. It is also reasonable to assume that these 500 firms are the most traded contracts, i.e., their CDS contracts are the most liquid, given they have the most complete data coverage. To speed up the estimation, we only use weekly (Friday) observations among these daily data. We believe these 500 firms represent the most traded single-name CDS contracts in the US market. All the CDS spreads are converted to upfronts using the ISAD standard CDS converter, assuming the fixed coupon rate C = 100 bps and a recovery rate of 40% (See ISDA, 2009). We bootstrap zero yields from the Constant Maturity Treasury (CMT) yields in release Fed H.15. The default-free discount factors used in our upfront pricing model are the prices of zero coupon bonds computed using these zero yields.6 Descriptive tables of the data are presented in the Online Appendix. Among the six ratings (AA, A, BBB, BB, B, and CCC), AA-, BBB-, and BB-rated firms account for 87% of the 500 firms. The table also contains the distribution of the firms among the ten different sectors, which include basic materials (BM), consumer goods (CG), consumer services (CS), financials (Fin), health care (HC), industrials (Ind), oil & gas (OG), technology (Tec), telecommunication (Tel), and utilities (Uti). Fin have most firms, followed by Ind, CG, CS, OG, and Uti. Other industries have relatively fewer firms. The top three categories that have most firms are BBB-rated Ind (54), Fin (39), and OG (39). Although we estimate our upfront pricing model using the upfronts converted from the CDS spreads, we present the data summary in terms of the CDS spreads for pedagogical reasons, as CDS spreads are closer to a traditional credit risk measure. For most ratings, we see an upward sloping credit curve, which is consistent with the notion that on average default risk is larger for longer maturities. For B- and CCC-rated firms, we observe a hump-shaped curve peaking at the maturity of 5-year. This is consistent with the notion that for speculative grade bonds, the default risk can be high in the near future but if the firm survives long enough, then the default risk actually goes down. The average credit curve for most sectors also slopes upward. One prominent exception is financial firms, whose CDS spread tends to decline with maturity. This is likely due to the fact that our sample covers the financial crisis period. During the crisis many financial firms were in trouble, therefore investors likely believe that these firms had higher credit risks in the short run than in the long run. In contrast, we find that the standard deviation of CDS spread generally declines with maturity. In general, lower rated firms have higher and more volatile CDS spreads than higher rated firms. One exception is that the A-rated firms actually have lower spreads than AA-rated firms. We believe this is probably because of too few A-rated firms (only five of them in total) in our sample. 4.2 Estimation Results We estimate the credit risk model using the whole term structure of upfronts for each of the 500 firms. Table I presents the first, third quarters, median, and mean of variance ratios at different maturities. The variance ratio measures the percentage of variations of the upfronts explained by the model. The model explains the variations of most upfronts very well. At most maturities, the median variance ratios (both the full sample and first the half of the sample) are above 90%. This suggests that our one-factor model does a reasonably good job in capturing the dynamics of the term structure of upfronts. Table I Summary of variance ratios This table shows the first quarter (Q1), median, third quarter (Q3), and mean of the of the variance ratios, which are the proportion of variations of the actual upfronts explained by our upfront pricing model, at nine maturities (1 yr, 2 yr, 3 yr, 5 yr, 7 yr, 10 yr, 15 yr, 20 yr, and 30 yr) as well as the average value (Ave.) across maturities.   Full sample   First subsample     Q1  Median  Q3  Mean  Q1  Median  Q3  Mean  1yr  0.76  0.86  0.92  0.81  0.71  0.84  0.91  0.75  2yr  0.88  0.93  0.96  0.89  0.86  0.92  0.96  0.86  3yr  0.95  0.97  0.98  0.95  0.94  0.97  0.98  0.93  5yr  1.00  1.00  1.00  1.00  1.00  1.00  1.00  1.00  7yr  0.96  0.98  0.99  0.97  0.95  0.98  0.99  0.95  10yr  0.90  0.96  0.98  0.92  0.86  0.94  0.97  0.89  15yr  0.85  0.93  0.97  0.88  0.76  0.86  0.93  0.81  20yr  0.80  0.91  0.96  0.84  0.68  0.81  0.90  0.76  30yr  0.75  0.89  0.94  0.81  0.61  0.74  0.85  0.70  Ave.  0.87  0.93  0.96  0.90  0.82  0.89  0.94  0.85    Full sample   First subsample     Q1  Median  Q3  Mean  Q1  Median  Q3  Mean  1yr  0.76  0.86  0.92  0.81  0.71  0.84  0.91  0.75  2yr  0.88  0.93  0.96  0.89  0.86  0.92  0.96  0.86  3yr  0.95  0.97  0.98  0.95  0.94  0.97  0.98  0.93  5yr  1.00  1.00  1.00  1.00  1.00  1.00  1.00  1.00  7yr  0.96  0.98  0.99  0.97  0.95  0.98  0.99  0.95  10yr  0.90  0.96  0.98  0.92  0.86  0.94  0.97  0.89  15yr  0.85  0.93  0.97  0.88  0.76  0.86  0.93  0.81  20yr  0.80  0.91  0.96  0.84  0.68  0.81  0.90  0.76  30yr  0.75  0.89  0.94  0.81  0.61  0.74  0.85  0.70  Ave.  0.87  0.93  0.96  0.90  0.82  0.89  0.94  0.85  Table I Summary of variance ratios This table shows the first quarter (Q1), median, third quarter (Q3), and mean of the of the variance ratios, which are the proportion of variations of the actual upfronts explained by our upfront pricing model, at nine maturities (1 yr, 2 yr, 3 yr, 5 yr, 7 yr, 10 yr, 15 yr, 20 yr, and 30 yr) as well as the average value (Ave.) across maturities.   Full sample   First subsample     Q1  Median  Q3  Mean  Q1  Median  Q3  Mean  1yr  0.76  0.86  0.92  0.81  0.71  0.84  0.91  0.75  2yr  0.88  0.93  0.96  0.89  0.86  0.92  0.96  0.86  3yr  0.95  0.97  0.98  0.95  0.94  0.97  0.98  0.93  5yr  1.00  1.00  1.00  1.00  1.00  1.00  1.00  1.00  7yr  0.96  0.98  0.99  0.97  0.95  0.98  0.99  0.95  10yr  0.90  0.96  0.98  0.92  0.86  0.94  0.97  0.89  15yr  0.85  0.93  0.97  0.88  0.76  0.86  0.93  0.81  20yr  0.80  0.91  0.96  0.84  0.68  0.81  0.90  0.76  30yr  0.75  0.89  0.94  0.81  0.61  0.74  0.85  0.70  Ave.  0.87  0.93  0.96  0.90  0.82  0.89  0.94  0.85    Full sample   First subsample     Q1  Median  Q3  Mean  Q1  Median  Q3  Mean  1yr  0.76  0.86  0.92  0.81  0.71  0.84  0.91  0.75  2yr  0.88  0.93  0.96  0.89  0.86  0.92  0.96  0.86  3yr  0.95  0.97  0.98  0.95  0.94  0.97  0.98  0.93  5yr  1.00  1.00  1.00  1.00  1.00  1.00  1.00  1.00  7yr  0.96  0.98  0.99  0.97  0.95  0.98  0.99  0.95  10yr  0.90  0.96  0.98  0.92  0.86  0.94  0.97  0.89  15yr  0.85  0.93  0.97  0.88  0.76  0.86  0.93  0.81  20yr  0.80  0.91  0.96  0.84  0.68  0.81  0.90  0.76  30yr  0.75  0.89  0.94  0.81  0.61  0.74  0.85  0.70  Ave.  0.87  0.93  0.96  0.90  0.82  0.89  0.94  0.85  The estimation enables the model to fit the 5-year CDS upfronts as well as possible (see Section 3.2.2) which are believed to be the most liquid and therefore the most informative about the credit risk of the underlying firm.7 The estimation results in Table I confirm that the one-factor model not only fits the 5-year CDS upfronts perfectly (with the variance ratios at 5-year being virtually 100%) but that it is also able to fit the maturities around 5 years very well, for example, 2, 3, 7, 10, and 15 years. Apparently, the one-factor model fits less well at the very short and long ends of the maturity spectrum. Adding more latent factors would improve the fitting performance at both ends. However, our goal here is neither to perfectly fit the whole term structure nor to find the “best” model, but rather to explore the information content implicit in the pricing errors. Specifically, we are interested in determining whether deviations of market prices from our one-factor model indicate any profitable trading opportunities. Since our one-factor model in general captures over 90% of the variation in the term structure dynamics and fits the 5-year maturity perfectly, we use our parsimonious one-factor specification to construct a trading strategy to explore the existence and economic content of any mispricings. It is also worth noting that if a “flawed” model (e.g., only one factor) generates significantly large excess returns, then the true mispricing in the market is even more pronounced. An improved model should generate even larger returns.8 Therefore, our results should be regarded as the conservative documentation of potential mispricing. Summaries of the parameter estimates are presented in Table II. The P measure parameters αP and κP imply a mean-reverting Zt for most of the firms. The median and average long-run mean ( αP/κP) of the processes are around 1.9% and 4.4%. In contrast, the risk-neutral parameters α and κ suggest an explosive Zt for most of the firms. This explains why in most cases we observe higher CDS spreads at the long maturities. This also implies that the expected rate of default intensity is lower under the physical measure than under the risk-neutral measure, indicating that investors require a premium for bearing exposure to varying default risk. This is consistent with the similar findings in several papers (see, e.g., Duffee, 1999; Pan and Singleton, 2008; Jarrow et al., 2010; Filipović and Trolle, 2013). From the estimated recovery rate 1−y, we find that the average is 62.7%. In contrast, the average debt recovery rate measured by post-default trading prices of senior unsecured bonds as reported in Moody’s special comment of “Corporate Default and Recovery Rates” over 2005 to 2011 is around 50.9%. Therefore, our estimated recovery rate slightly overestimates Moody’s recovery rate. Indeed, the Monte Carlo simulation exercise in the Online Appendix shows that the model underestimates the loss by 12% when the true loss rate is 37%. If we take this bias into consideration and adjust the recovery rate using the correction formula provided in the Online Appendix, we find that the corrected average recovery rate is 50.9% which is exactly Moody’s estimate. Table II Summary of parameter estimates This table reports the first quarter (Q1), median, third quarter (Q3), and mean of the parameter estimates of the 500 firms used in our empirical analysis. Panel (a) is the summary of parameter estimates using the full sample (Jan2005–Oct2011), Panel (b) is the summary of parameter estimates using the first half of the sample (Jan2005–Jun2006) (a) Full sample (Jan2005–Oct2011)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.306  0.115  0.003  0.353  0.472  Median  0.002  −0.189  0.143  0.017  0.914  0.572  Q3  0.004  −0.088  0.178  0.048  1.601  0.834  Mean  0.004  −0.221  0.154  0.062  1.416  0.627    (a) Full sample (Jan2005–Oct2011)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.306  0.115  0.003  0.353  0.472  Median  0.002  −0.189  0.143  0.017  0.914  0.572  Q3  0.004  −0.088  0.178  0.048  1.601  0.834  Mean  0.004  −0.221  0.154  0.062  1.416  0.627    (b) First subsample (Jan2005–Jun2008)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.510  0.143  0.005  0.714  0.553  Median  0.002  −0.392  0.173  0.026  2.552  0.841  Q3  0.003  −0.250  0.206  0.082  4.422  0.949  Mean  0.004  −0.386  0.175  0.093  7.172  0.738    (b) First subsample (Jan2005–Jun2008)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.510  0.143  0.005  0.714  0.553  Median  0.002  −0.392  0.173  0.026  2.552  0.841  Q3  0.003  −0.250  0.206  0.082  4.422  0.949  Mean  0.004  −0.386  0.175  0.093  7.172  0.738    Table II Summary of parameter estimates This table reports the first quarter (Q1), median, third quarter (Q3), and mean of the parameter estimates of the 500 firms used in our empirical analysis. Panel (a) is the summary of parameter estimates using the full sample (Jan2005–Oct2011), Panel (b) is the summary of parameter estimates using the first half of the sample (Jan2005–Jun2006) (a) Full sample (Jan2005–Oct2011)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.306  0.115  0.003  0.353  0.472  Median  0.002  −0.189  0.143  0.017  0.914  0.572  Q3  0.004  −0.088  0.178  0.048  1.601  0.834  Mean  0.004  −0.221  0.154  0.062  1.416  0.627    (a) Full sample (Jan2005–Oct2011)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.306  0.115  0.003  0.353  0.472  Median  0.002  −0.189  0.143  0.017  0.914  0.572  Q3  0.004  −0.088  0.178  0.048  1.601  0.834  Mean  0.004  −0.221  0.154  0.062  1.416  0.627    (b) First subsample (Jan2005–Jun2008)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.510  0.143  0.005  0.714  0.553  Median  0.002  −0.392  0.173  0.026  2.552  0.841  Q3  0.003  −0.250  0.206  0.082  4.422  0.949  Mean  0.004  −0.386  0.175  0.093  7.172  0.738    (b) First subsample (Jan2005–Jun2008)     α  κ  σ  αP  κP  1−y  Q1  0.001  −0.510  0.143  0.005  0.714  0.553  Median  0.002  −0.392  0.173  0.026  2.552  0.841  Q3  0.003  −0.250  0.206  0.082  4.422  0.949  Mean  0.004  −0.386  0.175  0.093  7.172  0.738    Panel (a) of Table III presents the median of the adjusted long-term default intensity mean, which is given by c0+αPκP, in ratings and sectors. Except for the rating of CCC, where the median adjusted long-term mean is slightly lower than that of the B rating, we find that the lower the rating, the higher the adjusted long-term mean. So the results confirm that default risks are larger for lower rated firms. From this panel, we also see that the telecommunication sector has the lowest adjusted long-term mean, and that the consumer goods sector has the highest adjusted long-term mean. The ranking among the sectors is different, however, during the first half of the sample period due to probably different macroeconomic environments. The ranking among the ratings is similar in the first half of the sample period. Only the A rating seems to have an abnormally large adjusted long-term mean relative to other ratings during the first half sample. However, the A rating has only five firms, so the results might not be representative. Table III Median adjusted long-term mean of default intensity in different ratings and sectors This table reports the median of the adjusted long-term mean of the default intensities (c0 plus the long-term mean of Zt under measure P, αP/κP) of different ratings and sectors. Panel (a) is based on the full sample (Jan2005–Dec2011) estimates, Panel (b) is based the first half of the sample (Jan2005–Jun2008) estimates. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) Full sample (Jan2005–Dec2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.000  0.009  0.011  0.067  0.053  0.011  CG  0.007    0.013  0.031  0.039  0.044  0.031  CS  0.005    0.006  0.017  0.019  0.033  0.017  Fin  0.013    0.022  0.024    0.000  0.017  HC      0.013  0.026  0.029  0.017  0.021  Ind  0.009    0.014      0.001  0.009  OG    0.017  0.009  0.029    0.050  0.023  Tec      0.009  0.003  0.028  0.048  0.019  Tel  0.003    0.006  0.024      0.006  Uti  0.012  0.011  0.015    0.046  0.012  0.012  Median  0.008  0.011  0.011  0.024  0.034  0.033  0.013    (a) Full sample (Jan2005–Dec2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.000  0.009  0.011  0.067  0.053  0.011  CG  0.007    0.013  0.031  0.039  0.044  0.031  CS  0.005    0.006  0.017  0.019  0.033  0.017  Fin  0.013    0.022  0.024    0.000  0.017  HC      0.013  0.026  0.029  0.017  0.021  Ind  0.009    0.014      0.001  0.009  OG    0.017  0.009  0.029    0.050  0.023  Tec      0.009  0.003  0.028  0.048  0.019  Tel  0.003    0.006  0.024      0.006  Uti  0.012  0.011  0.015    0.046  0.012  0.012  Median  0.008  0.011  0.011  0.024  0.034  0.033  0.013    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.003  0.008  0.018  0.007  0.008  0.008  CG  0.005    0.012  0.009  0.015  0.033  0.012  CS  0.002    0.002  0.002  0.005  0.023  0.002  Fin  0.006    0.006  0.004      0.006  HC      0.004  0.016  0.044  0.009  0.012  Ind  0.013    0.011      0.011  0.011  OG    0.012  0.009  0.007    0.009  0.009  Tec      0.005  0.000  0.021    0.005  Tel  0.005    0.005  0.007      0.005  Uti  0.018  0.016  0.006    0.007  0.007  0.007  Median  0.005  0.012  0.006  0.007  0.011  0.009  0.008    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.003  0.008  0.018  0.007  0.008  0.008  CG  0.005    0.012  0.009  0.015  0.033  0.012  CS  0.002    0.002  0.002  0.005  0.023  0.002  Fin  0.006    0.006  0.004      0.006  HC      0.004  0.016  0.044  0.009  0.012  Ind  0.013    0.011      0.011  0.011  OG    0.012  0.009  0.007    0.009  0.009  Tec      0.005  0.000  0.021    0.005  Tel  0.005    0.005  0.007      0.005  Uti  0.018  0.016  0.006    0.007  0.007  0.007  Median  0.005  0.012  0.006  0.007  0.011  0.009  0.008    Table III Median adjusted long-term mean of default intensity in different ratings and sectors This table reports the median of the adjusted long-term mean of the default intensities (c0 plus the long-term mean of Zt under measure P, αP/κP) of different ratings and sectors. Panel (a) is based on the full sample (Jan2005–Dec2011) estimates, Panel (b) is based the first half of the sample (Jan2005–Jun2008) estimates. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) Full sample (Jan2005–Dec2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.000  0.009  0.011  0.067  0.053  0.011  CG  0.007    0.013  0.031  0.039  0.044  0.031  CS  0.005    0.006  0.017  0.019  0.033  0.017  Fin  0.013    0.022  0.024    0.000  0.017  HC      0.013  0.026  0.029  0.017  0.021  Ind  0.009    0.014      0.001  0.009  OG    0.017  0.009  0.029    0.050  0.023  Tec      0.009  0.003  0.028  0.048  0.019  Tel  0.003    0.006  0.024      0.006  Uti  0.012  0.011  0.015    0.046  0.012  0.012  Median  0.008  0.011  0.011  0.024  0.034  0.033  0.013    (a) Full sample (Jan2005–Dec2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.000  0.009  0.011  0.067  0.053  0.011  CG  0.007    0.013  0.031  0.039  0.044  0.031  CS  0.005    0.006  0.017  0.019  0.033  0.017  Fin  0.013    0.022  0.024    0.000  0.017  HC      0.013  0.026  0.029  0.017  0.021  Ind  0.009    0.014      0.001  0.009  OG    0.017  0.009  0.029    0.050  0.023  Tec      0.009  0.003  0.028  0.048  0.019  Tel  0.003    0.006  0.024      0.006  Uti  0.012  0.011  0.015    0.046  0.012  0.012  Median  0.008  0.011  0.011  0.024  0.034  0.033  0.013    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.003  0.008  0.018  0.007  0.008  0.008  CG  0.005    0.012  0.009  0.015  0.033  0.012  CS  0.002    0.002  0.002  0.005  0.023  0.002  Fin  0.006    0.006  0.004      0.006  HC      0.004  0.016  0.044  0.009  0.012  Ind  0.013    0.011      0.011  0.011  OG    0.012  0.009  0.007    0.009  0.009  Tec      0.005  0.000  0.021    0.005  Tel  0.005    0.005  0.007      0.005  Uti  0.018  0.016  0.006    0.007  0.007  0.007  Median  0.005  0.012  0.006  0.007  0.011  0.009  0.008    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.003  0.008  0.018  0.007  0.008  0.008  CG  0.005    0.012  0.009  0.015  0.033  0.012  CS  0.002    0.002  0.002  0.005  0.023  0.002  Fin  0.006    0.006  0.004      0.006  HC      0.004  0.016  0.044  0.009  0.012  Ind  0.013    0.011      0.011  0.011  OG    0.012  0.009  0.007    0.009  0.009  Tec      0.005  0.000  0.021    0.005  Tel  0.005    0.005  0.007      0.005  Uti  0.018  0.016  0.006    0.007  0.007  0.007  Median  0.005  0.012  0.006  0.007  0.011  0.009  0.008    Panel (a) of Table IV presents the median of the recovery rates across ratings and sectors. For ratings, the estimated recovery rates range from 41% (CCC) to 93% (A). Except for the A rating, we find a monotone relation between the median recovery rates and the ratings, i.e., the higher the rating the larger the median recovery. This is consistent with the intuition that the firms with higher ratings typically have larger recovery rates. The deviation of the median recovery rates among different sectors is smaller, as the range is from 46% (CG) to 68% (Ind). Again, the interpretation of these estimated recovery rates comes with a caveat that although the relative ranking among the recovery rates is accurate, the absolute level of the recovery rates is somewhat overestimated. According to the correction formula in the Online Appendix, the recovery rates of 41% (CCC), 93% (A), 46% (CG), and 68% (Ind) should be 33% (CCC), 76% (A), 37% (CG), and 55% (Ind), respectively. Table IV Median recovery rate in different ratings and sectors This table reports the median of the recovery rates in different ratings and sectors. Panel (a) is based on the full sample (Jan2005–Oct2011) estimates, Panel (b) is based the first half of the sample (Jan2005–Jun2008) estimates. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) Full sample (Jan2005–Oct2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.925  0.709  0.421  0.471  0.467  0.471  CG  0.880    0.794  0.460  0.410  0.416  0.460  CS  0.869    0.755  0.501  0.456  0.412  0.501  Fin  0.571    0.499  0.416    0.427  0.463  HC      0.831  0.547  0.513  0.436  0.530  Ind  0.888    0.684      0.408  0.684  OG    0.931  0.532  0.468    0.401  0.500  Tec      0.726  0.598  0.389  0.401  0.499  Tel  0.871    0.647  0.472      0.647  Uti  0.905  0.857  0.543    0.451  0.272  0.543  Median  0.876  0.925  0.697  0.470  0.454  0.412  0.572    (a) Full sample (Jan2005–Oct2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.925  0.709  0.421  0.471  0.467  0.471  CG  0.880    0.794  0.460  0.410  0.416  0.460  CS  0.869    0.755  0.501  0.456  0.412  0.501  Fin  0.571    0.499  0.416    0.427  0.463  HC      0.831  0.547  0.513  0.436  0.530  Ind  0.888    0.684      0.408  0.684  OG    0.931  0.532  0.468    0.401  0.500  Tec      0.726  0.598  0.389  0.401  0.499  Tel  0.871    0.647  0.472      0.647  Uti  0.905  0.857  0.543    0.451  0.272  0.543  Median  0.876  0.925  0.697  0.470  0.454  0.412  0.572    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.940  0.919  0.503  0.459  0.508  0.508  CG  0.959    0.958  0.556  0.277  0.456  0.556  CS  0.899    0.853  0.457  0.392  0.471  0.471  Fin  0.837    0.778  0.505    0.514  0.646  HC      0.900  0.598  0.671  0.392  0.634  Ind  0.958    0.849      0.522  0.849  OG    0.973  0.947  0.389    0.225  0.668  Tec      0.945  0.737  0.338  0.497  0.617  Tel  0.896    0.607  0.309      0.607  Uti  0.975  0.972  0.872    0.012  0.147  0.872  Median  0.928  0.972  0.886  0.504  0.365  0.471  0.845    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.940  0.919  0.503  0.459  0.508  0.508  CG  0.959    0.958  0.556  0.277  0.456  0.556  CS  0.899    0.853  0.457  0.392  0.471  0.471  Fin  0.837    0.778  0.505    0.514  0.646  HC      0.900  0.598  0.671  0.392  0.634  Ind  0.958    0.849      0.522  0.849  OG    0.973  0.947  0.389    0.225  0.668  Tec      0.945  0.737  0.338  0.497  0.617  Tel  0.896    0.607  0.309      0.607  Uti  0.975  0.972  0.872    0.012  0.147  0.872  Median  0.928  0.972  0.886  0.504  0.365  0.471  0.845    Table IV Median recovery rate in different ratings and sectors This table reports the median of the recovery rates in different ratings and sectors. Panel (a) is based on the full sample (Jan2005–Oct2011) estimates, Panel (b) is based the first half of the sample (Jan2005–Jun2008) estimates. The last column (the last row), except the last cell, is the median over each row (column). The last cell at the right-lower conner is the median over all firms. (a) Full sample (Jan2005–Oct2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.925  0.709  0.421  0.471  0.467  0.471  CG  0.880    0.794  0.460  0.410  0.416  0.460  CS  0.869    0.755  0.501  0.456  0.412  0.501  Fin  0.571    0.499  0.416    0.427  0.463  HC      0.831  0.547  0.513  0.436  0.530  Ind  0.888    0.684      0.408  0.684  OG    0.931  0.532  0.468    0.401  0.500  Tec      0.726  0.598  0.389  0.401  0.499  Tel  0.871    0.647  0.472      0.647  Uti  0.905  0.857  0.543    0.451  0.272  0.543  Median  0.876  0.925  0.697  0.470  0.454  0.412  0.572    (a) Full sample (Jan2005–Oct2011)     AA  A  BBB  BB  B  CCC  Median  BM    0.925  0.709  0.421  0.471  0.467  0.471  CG  0.880    0.794  0.460  0.410  0.416  0.460  CS  0.869    0.755  0.501  0.456  0.412  0.501  Fin  0.571    0.499  0.416    0.427  0.463  HC      0.831  0.547  0.513  0.436  0.530  Ind  0.888    0.684      0.408  0.684  OG    0.931  0.532  0.468    0.401  0.500  Tec      0.726  0.598  0.389  0.401  0.499  Tel  0.871    0.647  0.472      0.647  Uti  0.905  0.857  0.543    0.451  0.272  0.543  Median  0.876  0.925  0.697  0.470  0.454  0.412  0.572    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.940  0.919  0.503  0.459  0.508  0.508  CG  0.959    0.958  0.556  0.277  0.456  0.556  CS  0.899    0.853  0.457  0.392  0.471  0.471  Fin  0.837    0.778  0.505    0.514  0.646  HC      0.900  0.598  0.671  0.392  0.634  Ind  0.958    0.849      0.522  0.849  OG    0.973  0.947  0.389    0.225  0.668  Tec      0.945  0.737  0.338  0.497  0.617  Tel  0.896    0.607  0.309      0.607  Uti  0.975  0.972  0.872    0.012  0.147  0.872  Median  0.928  0.972  0.886  0.504  0.365  0.471  0.845    (b) First subsample (Jan2005–Jun2008)     AA  A  BBB  BB  B  CCC  Median  BM    0.940  0.919  0.503  0.459  0.508  0.508  CG  0.959    0.958  0.556  0.277  0.456  0.556  CS  0.899    0.853  0.457  0.392  0.471  0.471  Fin  0.837    0.778  0.505    0.514  0.646  HC      0.900  0.598  0.671  0.392  0.634  Ind  0.958    0.849      0.522  0.849  OG    0.973  0.947  0.389    0.225  0.668  Tec      0.945  0.737  0.338  0.497  0.617  Tel  0.896    0.607  0.309      0.607  Uti  0.975  0.972  0.872    0.012  0.147  0.872  Median  0.928  0.972  0.886  0.504  0.365  0.471  0.845    Since our main focus, which is to develop a trading strategy, does not require an unbiased estimate of the recovery rate, and because our model fits the CDS upfront data reasonably well, we proceed using these parameter estimates without correction. 5. Statistical Arbitrage Strategy In this section, we first briefly review the definitions of statistical arbitrage and explain how CDS contracts are traded. We then discuss the design and implementation of the strategy. 5.1 Definitions of Statistical Arbitrage Statistical arbitrage is typically referred to as trading strategies that rely on mathematical modeling techniques seeking profit opportunities from pricing inefficiencies (see, e.g., Whistler, 2004; Pole, 2007; Avellaneda and Lee, 2010). In the academic community, statistical arbitrage is defined more rigorously and often employed to develop statistical tools for testing market efficiency (Bondarenko, 2003; Hogan et al., 2004; Jarrow et al., 2012). For example, Hogan et al. (2004) define a statistical arbitrage to satisfy four conditions: (i) it is a zero initial cost self-financing strategy,9 (ii) it has positive expected profits in the limit as time goes to infinity, (iii) the probability of a loss converges to zero, and (iv) if loss the probability is nonzero in finite time, a time-average variance converges to zero. Notice that the academic definition does not conflict with the practical one. The strategy we develop is consistent with the both definitions. First, the practical implications of our strategy make it potentially useful for hedge funds and investment banks engaging in quantitative trading in the CDS market, while we admit that the technical details involved in actual trading is not our focus and depends on an idiosyncratic set of constraints faced by the arbitrageur (Fleckenstein, Longstaff, and Lustig, 2014); second, the previously developed statistical arbitrage test can be applied to the profits generated by our strategy to examine the extent of any mispricing in the CDS market. 5.2 Trading CDS Contracts As mentioned before, after the “CDS Big Bang” the quoting convention for CDS contracts changed from quoting par spreads10 to quoting upfronts. With this quoting convention, the upfronts can be directly regarded as the price of the contract. The capital gain before transaction costs (CGb) of a trade is simply the notional amount (NA) times the positive or negative difference between the starting and ending upfronts of the trade depending on whether the trade is initially a long or a short position:   CGt+Δt=NAtIt(Upfrontt+Δt−Upfrontt), (15) where Δt is the duration of the trade; It is an indicator variable that equals 1 if the trade is initiated in a long position and −1 if it is initiated in a short position. The return of a trade is given by   rt=CGtIMt−Δt, (16) where IMt−Δt is the initial margin posted at time t−Δt. In this section, we abstract from transaction costs and focus primarily on the idea of the strategy itself while leaving the detailed discussion of transaction costs to Section 7. 5.3 Design and Implementation The idea behind our trading strategy is to exploit the predictability in the pricing residuals produced by an arbitrage-free term structure model. Economically significant predictability in the pricing residuals in government bonds and LIBOR/Swap markets has been documented in Sercu and Wu (1997) and Bali, Heidari, and Wu (2009). Here we develop a trading strategy similar to Bali, Heidari, and Wu (2009) to explore mispricings in the CDS market. We rely on our estimated CDS pricing model to extract the state variable of the individual default risk (under the risk-neutral measure) and produce pricing residuals. We construct market-neutral portfolios of CDS contracts that are immune to both the first- and second-order changes in the state variable. Then, we long (short) under (over) valued hedged portfolios. Here are the details of the strategy. We consider the second-order expansion of the upfronts pricing function U(Zt,τ) around the state variable Zt with the following first- and second-order derivatives H1(Zt,τ)=∂U(Z,τ)∂Z|Z=Zt and H2(Zt,τ)=∂2U(Z,τ)∂Z2|Z=Zt. The closed-form formulae of H1 and H2 are presented in Appendix C. Specifically, we have   U(Zt+Δt,τ)=U(Zt,τ)+H1(Zt,τ)(Zt+Δt−Zt)+12H2(Zt,τ)(Zt+Δt−Zt)2+O((Zt+Δt−Zt)3). (17) We assume for Δt= 1 week,11 U(Z^t+Δt,τ) can approximate U(Z^t+Δt,τ−Δt) well, i.e.,   U(Zt+Δt,τ−Δt)≈U(Zt+Δt,τ). (18) Then, by the above approximation and ignoring high order terms, (17) can be rewritten as   U(Zt+Δt,τ−Δt)≈U(Zt,τ)+H1(Zt,τ)(Zt+Δt−Zt)+12H2(Zt,τ)(Zt+Δt−Zt)2. (19) Denote any maturity in ( τ=1,2,3,5,7,10,15,20, and 30) by τ0, and the two adjacent maturities of τ0 by τ1 and τ2. To see how τ1 and τ2 are chosen, let us look at a few examples: if τ0=1, then τ1=2, and τ2=3; if τ0=7, then τ1=5, and τ2=10; and if τ0=30, then τ1=15, and τ2=20. Therefore, given Δt, at time t for a unit NA of a CDS contract with maturity of τ0, we employ two other CDS contracts with maturities of τ1 and τ2 to form a hedged portfolio, whose value is immune to the variation of Zt up to the second order but subject to the variation of the pricing residuals. To see this, let us recall our measurement Equation (11) in Section 3.2.1 for τ0, τ1, and τ2:   Upfrontstτi⃗︸3×1=U(Zt,τi⃗)︸3×1+εtτi⃗︸3×1, (20) where τi⃗=[τ0,τ1,τ2]⊺. Denote the portfolio weight by M(Zt,τi⃗)︸3×1. We fix the first element in M(Zt,τi⃗) at one for any τi⃗. So the future value of the portfolio at time t+Δt is given by   M(Zt,τi⃗)⊺Upfrontst+Δtτi⃗−Δt=M(Zt,τi⃗)⊺U(Zt+Δt,τi⃗−Δt)+M(Zt,τi⃗)⊺εt+Δtτi⃗−Δt≈M(Zt,τi⃗)⊺U(Zt,τi⃗)+M(Zt,τi⃗)⊺[H1(Zt,τi⃗)(Zt+Δt−Zt)+12H2(Zt,τi⃗)(Zt+Δt−Zt)2]+M(Zt,τi⃗)⊺εt+Δtτi⃗. (21) If the second and third elements of M(Zt,τi⃗) are specified such that   M(Zt,τi⃗)⊺H1(Zt,τi⃗)=0,M(Zt,τi⃗)⊺H2(Zt,τi⃗)=0, (22) i.e.,   M(Zt,τi⃗)⊺=[1,m1(Zt,τi⃗),m2(Zt,τi⃗)],m1(Zt,τi⃗)=H2(Zt,τ0)H1(Zt,τ2)−H1(Zt,τ0)H2(Zt,τ2)H1(Zt,τ1)H2(Zt,τ2)−H2(Zt,τ1)H1(Zt,τ2),m2(Zt,τi⃗)=H2(Zt,τ0)H1(Zt,τ1)−H1(Zt,τ0)H2(Zt,τ1)H1(Zt,τ2)H2(Zt,τ1)−H2(Zt,τ2)H1(Zt,τ1), (23) we have the hedged portfolio. That is   M(Zt,τi⃗)⊺Upfrontst+Δtτi⃗−Δt≈M(Zt,τi⃗)⊺U(Zt,τ)+M(Zt,τi⃗)⊺εt+Δtτi⃗=M(Zt,τi⃗)⊺Upfrontstτi⃗+M(Zt,τi⃗)⊺Δεt+Δtτi⃗, (24) where Δεt+Δtτi⃗=εt+Δtτi⃗−εtτi⃗. For each τi⃗, we can either long or short this hedged portfolio. The long (1) and short (−1) indicators are summarized in a 9×1 vector, It. Given It and M(Zt,τi⃗), the portfolio weight vector W(Zt,I,τ⃗) for all the nine CDS contracts is given by   W(Zt,It,τ⃗)︸9×1=[I1,tM⃗(Zt,τ1⃗),I2,tM⃗(Zt,τ2⃗),⋯,I9,tM⃗(Zt,τ9⃗)]︸9×9=1︸9×1, (25) where τ⃗=[1,2,3,5,7,10,15,20,30]⊺, Ii,t is the ith element in It, M⃗(Zt,τi⃗) is a 9 × 1 vector in which the elements corresponding to τi⃗ in τ⃗ are M(Zt,τi⃗), and others are zero,12 and 1 is the vector of ones. Therefore, the expected capital gain after transaction costs at time t from holding the hedged portfolio of the whole term structure of CDS contracts for a period of Δt, given It, W(Zt,It,τ⃗), and the forecast of Δεt+Δtτi⃗, Et(Δεt+Δtτ⃗), is   Et(CGt+Δt)=W(Zt,It,τ⃗)⊺Et(Upfrontst+Δtτ⃗−Δt−Upfrontstτ⃗)≈W(Zt,It,τ⃗)⊺Et(Δεt+Δtτ⃗). (26) We use the negative exponential moving average to forecast Δεt+Δtτ⃗, specifically,   Et(Δεt+Δtτ⃗)=−Σj=0n(1−2n+1)j2n+1Δεt−jΔtτ⃗,n=4. (27) This is an exponential moving average of the past 5 weeks of Δεtτ⃗. The negative exponential moving average captures three characteristics of Δεtτ⃗: (i) the conditional mean is close to zero; (ii) the speed of mean reverting is large; and (iii) the recent lags are informative in forecasting next period’s value. Since we have nine maturities, at each point in time t, It can be chosen from Σi=199!i!(9−i)!=511 different vectors of −1 and 1 combinations. Denote the entire set of these 511 vectors by I. At each point in time t, the portfolio weight W(Zt,It,τ⃗) is chosen over I to maximize Et(CGt+Δta):   W˜t={W(Zt,I˜t,τ⃗)09×1if W(Zt,I˜t,τ⃗)⊺Et(Δεt+Δtτ⃗)>0if W(Zt,I˜t,τ⃗)⊺Et(Δεt+Δtτ⃗)≤0, (28) where13 I˜t maximizes (26), i.e.,   I˜t=arg⁡max⁡It∈I{W(Zt,It,τ⃗)⊺Et(Δεt+Δtτ⃗)}. (29) A real-life example of this weight calculation can be found in the Online Appendix. The realized capital gain CGt+Δt is given by   CGt+Δt=st[W˜t⊺( Upfrontst+Δtτ⃗−Δt−Upfrontstτ⃗)], (30) where st is a scale factor and defined as   st={1L¯W˜t−Δt⊺1−S̲W˜t−Δt⊺1if 0≤W˜t−Δt⊺1≤L¯ or −S̲≤W˜t−Δt⊺1≤0if W˜t−Δt⊺1>L¯if W˜t−Δt⊺1<−S̲, and L¯ ( S̲) is the cap on the net NA for a net long (short) position. Once the model is estimated, Zt can be backed out by fitting the estimated model to the term structure of the observed Upfronts. Given the estimated model parameters and Zt, H1(Zt,τ), H2(Zt,τ), εt can be calculated for any τ and t, which then allow us to implement the above optimization procedure. In the main test, we assume that we invest $1 M as an initial deposit, i.e., IM0= $1 M which covers the initial margin for counterparty risk and collaterals for compensating the loss in case of an underlying default. With this $1M, we are allowed to short with no more than the initial deposit, $1M, and long with no more than three times the initial deposit, 3×$1M = $3M. That is L¯=3×$1M = $3 M and S̲= $1M. During the sample period, we add cash to restore the account to $1 M whenever the total money in the account is less than $1M, which ensures no margin calls. We never withdraw cash from the account during the sample period, implying that if we are consistently making profit then the returns will be lower for a given capital gain. 6. Evidence and the Economic Contents of Mispricings In this section, we first show evidence of the existence and persistence CDS market mispricings in terms of positive trading strategy returns. Then, we explore the economic contents of the mispricings by regressing the trading strategy returns onto various economic variables. 6.1 Trading Strategy Returns as a Mispricing Measure It is natural to use the returns generated from the trading strategy as a measure of mispricing in the sense of statistical arbitrage. If the strategy consistently generates positive returns, then the mispricing is persistent. We focus only on the out-of-sample performance of the strategy. We include the in-sample performance for comparison purposes in the next subsection and Section 7. S