CDS Spreads and Investor Sentiment During 2008 Global Financial Crisis

Transcription

1 CDS Spreads and Investor Sentiment During 2008 Global Financial Crisis Jeehye Lee *, Sol Kim and Yuenjung Park Abstract This paper examines whether investor sentiment can predict credit default swap (CDS) spread changes. Among several proxies for investor sentiment, change in equity put/call ratio performs best in predicting variations of CDS spread changes in both firm- and portfolio-level regressions, and especially its explanatory power is greater for investment-grade firms than for non investment-grade firms. More importantly, sentiment may be a critical factor to determine CDS spread changes during a financial crisis and best explains the differences in CDS spread in the group whose leverage ratio and stock volatility are highest. Keywords: Investor Sentiment, Credit Default Swap, Equity put/call ratio, Financial Crisis. JEL classifications: G01, G13, G14. * Graduate, College of Business, Hankuk University of Foreign Studies, 270, Imun-dong, Dongdaemun- Gu, Seoul, Korea. Tel: , Fax: , lovelyjeehye@naver.com. Professor, College of Business, Hankuk University of Foreign Studies, 270, Imun-dong, Dongdaemun- Gu, Seoul, Korea. Tel: , Fax: , solkim@hufs.ac.kr. Corresponding Author, Assistant Professor, Hallym University, 1 Hallymdaehak-gil, Chuncheon, Gangwon-do, Korea. Tel: , yjpark5830@gmail.com.

2 CDS Spreads and Investor Sentiment During 2008 Global Financial Crisis Abstract This paper examines whether investor sentiment can predict credit default swap (CDS) spread changes. Among several proxies for investor sentiment, change in equity put/call ratio performs best in predicting variations of CDS spread changes in both firm- and portfolio-level regressions, and especially its explanatory power is greater for investment-grade firms than for non investment-grade firms. More importantly, sentiment may be a critical factor to determine CDS spread changes during a financial crisis and best explains the differences in CDS spread in the group whose leverage ratio and stock volatility are highest. Keywords: Investor Sentiment, Credit Default Swap, Equity put/call ratio, Financial Crisis. JEL classifications: G01, G13, G14.

3 1. Introduction Recent literature investigating credit spread changes have focused on finding the systematic determinant to resolve the weak explanatory power of theoretical variables. 1 Avramov, Jostova, and Philipov (2007) conducted a linear time-series regression of differences in corporate bond credit spread using a structural model incorporating macroeconomic dummies and the Fama and French (1993) three factors (here after referred to as FF factors). They reported that although FF factors are significant in the total sample, the explanatory power of the FF variables are not observed when the corporate bonds are categorized by the three credit risk groups. Galil, Shapir, Amiram, and Ben-Zion (2014) proposed a model for CDS spread changes by analyzing the FF factors along with Pastor and Stambaugh (2003) the liquidity factor and the Chen, Roll and Ross (1986) five factors (henceforth, CRR factors). While the coefficients were significant only in the FF factor model, insignificant coefficients were observed in the CRR and FF factor models that added to the liquidity factor. These poor performances in preceding empirical studies of rational CDS spread determinants motivate us to undertake a study considering factors beyond traditional structural determinants. In this paper, we suggest a novel approach to corporate credit spread and investigate whether credit default swap (CDS) spread changes can be well predicted by investor sentiment proxies, utilizing a structural model that embeds theoretical factors as control variables. Thus we verify the hypothesis that investor sentiment has a role of common or systematic risk factors of CDS spread changes and explore which sentiment measurement can be the most effective determinant. 1 The theoretical model has shown a large gap from the historically observed credit spread. For example, Collin- Dufresne, Goldstein, and Martin (2001) revealed that the variables of the theoretical model explain only a limited portion of credit spread changes of bonds. In addition, using credit default swap (CDS) spreads as a proxy of credit risk, Ericsson, Jacobs and Oviedo-Helfenberger (2009) showed that theoretical variables have low explanatory power for credit spread changes. 1

4 In addition, more important hypothesis in this research is that investor sentiment explains CDS spread changes better in turmoil period. Tang and Yan (2010) empirically showed that the interaction between default risk of CDS spread and investor sentiment can be dependent on market states such as bullish or bearish market. Further, Stambaugh, Yu, and Yuan (2012) insisted that when sentiment exists widely in the market, the limitation in short sale contributes a major role to increase the degree of the existence of the mispriced asset. The authors also opined that a higher level of sentiment, in turn generally considered as the downturn, is associated with more overpricing. Based on the theory of Stambaugh et al. (2012) as well as the result of Tang and Yan (2010), we conjecture that in bad economic times overpricing boosted by the impediment to short selling hinders the hedging of CDSs by equity or equity options and moves up CDS spreads whereas in the stable period the effect of market-wide sentiment should not be strong due to the absence of roles of short selling restriction. Therefore, to examine the idea inferred from Stambaugh et al. (2012), we divide the sample period into two periods: the pre-crisis as a normal term and the global financial crisis period as a representative turmoil term. This paper restricts examination to the global financial crisis and do not include the post crisis period as the comparative sample. The reason is that asset prices in the U.S. market had been externally influenced by several events originated by the debt problem of the Greece in the post crisis period, which cannot be considered as purely normal or purely turmoil term. To test above hypotheses, we perform both firm- and portfolio-level regressions. While most other studies about credit risk only conducted regression analyses at the firm-level, a regression analysis of the portfolio level is carried out to alleviate the problem by idiosyncratic risk in firm-level regressions. We build the 5X5 portfolios using the leverage ratio and stock volatility of individual firms. 2 2 In detail, the process of constructing portfolios is conducted following Kim et al. (2013). We first compute the time series averages of volatility and leverage ratio for each corporation over the full sample period. Next, we set the upper and lower limits of five groups classified by the magnitude of the average leverage ratio and then, within a leverage ratio group, set the upper and lower limits of five groups classified by the magnitude of the average volatility. Each of 2

5 The main results of empirical tests are as follows. First, most of the sentiment proxies are economically significant factors to explain CDS spread changes. Among the sentiment measurements, change in Equity put/call ratio performs best in predicting CDS spread changes in both firm-level and portfolio-level regressions. Second, the results from the portfolio-level regressions show that changes in Equity put/call ratio shows outstanding explanatory power which is almost as triple high as that of structural model. In the detail observation in result of each portfolio, we find that sentiment can explain CDS spread best in the group whose leverage ratio and stock volatility are highest and vice versa. Also, the gap of ability to predict between the two polarized portfolios is markedly large. Finally, when we divide the sample period into the pre-crisis and crisis periods, sentiment models explain CDS spread changes much better in a turmoil period than in a stable period. 3 The rest of this paper is organized as follows. In section 2, the related background literature is outlined. In section 3, we describe the data used in empirical analysis. In section 4, we introduce the analytical models. Section 5 presents the results from the individual and portfolio regressions. Section 6 concludes with the empirical results and remarks. 2. Literature Review The investor sentiment has been applied to numerous theoretical and empirical financial considerations. Concerning stock asset pricing, Yu and Yuan (2011) tested how investor sentiment is related to the market s risk and return. They reported that the expected excess return in the stock market is high in low-sentiment the 285 firms are allocated to one of the 25 portfolios, classified by five volatility and five leverage ratio ranges. Finally, we calculate the cross-sectional averages of the CDS spread, volatility and leverage ratio for each portfolio, and generate the time series of each variable for the 25 separate portfolios. 3 Some studies found that the sample period categorized by regimes has economically important meaning in explaining CDS spreads. (Alexander and Kaeck (2008); Cesare and Guazzarotti (2010)) 3

6 periods, while variance has a weaker negative correlation with the expected risk premium in highsentiment periods, since the role of sentiment traders is larger during high-sentiment periods and sentiment traders can erode the risk-premium. Furthermore, investor sentiment shows outstanding results in options market. Han (2008) documented that investor sentiment proxies in the stock market were related to index option volatility when the risk-neutral skewness in the Standard and Poor s (S&P) 500 index option was measured according to Bakshi, Kapadia, and Madan (2003). These results cannot be explained by the rational option pricing model. However, for the credit risk, investor sentiment has not been significantly considered, with only few studies having addressed this variable. Tang and Yan (2010) incorporated the Conference Board Consumer Confidence (CBCC) Index as the proxy for changes of risk aversion that alters the market risk premium. The authors revealed that investor sentiment greatly predicts credit spread. However, their study focused more on the influence of market risk on CDS spread changes through the structural model, with sentiment not fully investigated. Further, general participants in derivatives market are large, often institutional, investors, rather than small investors, such as individuals. Considering only individual sentiment, such as the CBCC Index, may provide only weak evidence. Therefore, we strive in this paper to consider both individual s and institutions sentiments used in various derivatives markets. Chen and Wang (2010) incorporated investor sentiments in a model as the determinants of CDS spread from 2004 to They categorized diverse sentiments into two groups: market-wide and firm-specific sentiment derived from options market and reported that the investor sentiment explains CDS spread directly and indirectly by affecting stock market and options market. Saka et al. (2015) discover that only during the pre-announcement period, frequent days of large adverse shocks influencing simultaneously five Eurozone countries are detected by country-specific regressions of CDS spreads on systematic risk factors. They support De Grauwe (2011, 2012)'s fragility hypothesis and the underlying multiple-equilibria theory of the crisis. The latter adds that panic boosts exogenous shocks and can push an otherwise solvent country toward default in times of enormous economic adjustment. This result implies that there is additional factor beyond a strong association between sovereign credit 4

7 spreads and fundamentals to the recent Eurozone debt crisis. Without contradicting the role of fundamentals, multiple-equilibria theorists insist that a self-fulfilling dynamics driven by market sentiments of fear and panic has played critical role in the region pushing countries towards a worse equilibrium than is explained by fundamentals alone. To fill the gap in the CDS spread studies, this paper differs in several facets from preceding studies. First, we use various measurements for sentiments and examined which market sentiments best predict CDS, with the goal of exploring a representative proxy of sentiment. We assume that since many studies have described credit risk as being related to the risk of other markets, such as stock and options markets, their sentiments may have some explanatory power on CDS spread changes. Hence, to examine whether sentiments explain variation in CDS premia, we embed diverse sentiment proxies from different sources: that is, other sentiment proxies calculated with the stock market, especially the S&P 500 return, in addition to the individual sentiment proxies such as the Michigan Consumer Sentiment index and American Association Investment Index 4. Moreover, as Han (2008) indicated, individual sentiment is relatively unrelated to the derivatives market because individuals are not major participants in the market. Thus, we incorporate market-wide sentiments from the options market as well as the futures market. Second, Compared to research that has involved only linear regression of the CDS spread on the sentiment variables in Chen and Wang (2010), our study is based more on a structural model embedding theoretical factors as controlling variables. Moreover, the sample period of Chen and Wang (2010) did not include the global financial crisis, which is an extremely important term for CDS studies whereas we expand the research period from 2006 to 2009 to include the period of global financial crisis. Third, while Saka et al. (2015) explored the increasing role played by the pessimistic market sentiment that can trigger default for sovereign CDS spreads during the European crisis period, this paper focused the influence of sentiment beyond those of fundamental factors for determining U.S. corporate CDS spreads 4 These sentiment proxies are generally called as individual sentiment since the responders of the survey are individuals. 5

8 during the global crisis period. 3. Data 3-1. CDS We use CDS data of senior unsecured USD-denominated debt with a modified restructure clause and five years of maturity obtained from Markit. We remove firms from the utility and financial sectors in addition to firms with unknown ratings. 667 firms remain by matching stock price data from the Center for Research in Security Prices (CRSP) and accounting data from Compustat. Then, 285 firms with complete monthly observations for the 44 month period from January 2006 to August 2009 are selected. 5 To date, researches have addressed credit risk using with pure credit spread in level and changes. However, as many studies document, this method to measure credit risk has statistical problems. Greatex (2008) casted doubt on findings with CDS spread levels, rather than differences. Unit root test results indicate that CDS spread levels are mostly non-stationary, which means the data is statistically inappropriate to be used in time-regression analysis as a variable. Moreover, Das et al. (2007) and Duffie et al. (2007, 2009) revealed that corporate default rate has an exponential relationship with variables, so the variables have a linear relationship with natural logarithm differences of CDS, rather than with the pure differences of CDS spread. In fact, reduced form-model studies (e.g., Pan and Singleton (2008), which is another approach to the credit risk, have established that the model using logarithmic stochastic process can predict best. Thus, considering these reasons and following the Remolona et al. (2007) s methodology, we use the differences of CDS spread natural logarithm as the dependent variable. That is, the CDS spread changes are calculated as below: 5 We select monthly data along with the previous researches (Collin-Dufresne et al. (2001); Ericsson et al. (2009); Kim et al. (2013)) and we reports the results utilizing data on first trading day of each month in this paper. In addition, we confirm similar results when using data from the last trading of each month. 6

9 Δlog (CDS i,t )= log (CDS i,t ) log (CDS i,t 1 ) (1) where log (CDS i,t ) is the natural logarithm of CDS spread in month t for firm i Figure 1 depicts the time series of average CDS spread and average of CDS spread changes from January 2006 to August [Figure 1 placed here] 3-2. Investor Sentiment A. Michigan Consumer Sentiment index 6 (SentMi) The SentMi is a survey-based consumer confidence index that is conducted monthly by the Survey Research Center at the University of Michigan, based period of December Survey question primarily concerns three broad areas: personal finances, business conditions, and buying conditions of consumers statistically designed to be representative of all American households. 7 According to Cooper and Gubellini (2011), it is usually the indicator of the future statement of the overall economy. Thus, we assume SentMi is the sentiment for the business cycle. B. American Association of Individual Investor bull-bear spread (SentAA) SentAA is the survey-based investor sentiment conducted on a weekly basis. Individuals selected by ranks of the American Association Individual Investor (AAII) membership are questioned as to whether 6 As Tang and Yan (2010) insist that the correlation between Michigan Consumer Sentiment index and Conference Board Consumer Confidence Index is high, we report only the result of Michigan Consumer Sentiment index as one of individual sentiment. 7 Since it is monthly data and announced at the middle of the month as a prelim and the late of the month as a final index, we use one month lagged sentiment which is the closest result to the CDS data. 7

10 the stock market will be bullish, bearish, or neutral for the next 6 months and the percentage of each stock market statement is reported as the index. We utilize the bull-bear spread which is computed by AAII bullish percentage minus AAII bearish percentage as a proxy for sentiment. 8 Since Qiu and Welch (2006) insist that survey-based sentiment performs better than other formulated sentiment in stock market, we use SentAA as another individual s sentiment measurement in addition to SentMi. C. Baker and Wurgler (2007) (SentBW) Following the equation presented by Baker and Wurgler (2007), the SentBW is calculated by the weighted sum of closed-end fund discount detrended log turnover, number of initial public offerings (IPOs), the first-day return on IPOs, the dividend premium, and the equity share in new issues. 9 Since it measures sentiment in the stock market (Baker and Wurgler, 2007), we use this proxy as the stock market sentiment. D. Equity put/call ratio (SentEPCR) According to Chen, Lung, and Tay (2005), we assume that the equity put/call ratio represents a measure of option market sentiment. Chen, Lung, and Tay (2005) show that the equity put/call ratio is a credible measure for discerning good and bad news information embedded in option trades because it considers information that are reflected in both option trading volumes and premiums. They seek to establish the relationship between stock returns and option trading activities and identify a measure for discerning good news versus bad news information embedded in option trades. They 8 Since the SentAA is weekly data, not daily, the date does not match with CDS data which is daily-based one. Therefore, we have to convert weekly data to daily one, and when we do that, we match the CDS date with the week of SentAA which includes the CDS s date because we think that the CDS is affected at the time when it is announced. For example, SentAA announced at September 2 nd, 2010 lasts until September 8 th when the following value will be announced. Thus, the CDS data of between 2 nd to 8 th is matched with SentAA data announced at September 2 nd, Since it is monthly data and announced not at the first trading date, we use one month lagged sentiment which is the closest result to the CDS data. 8

11 consider a two-state world where informed traders have information about the probabilities of the up and down states. The informed traders are risk adverse, share the same log utility function, and have homogenous expectations about market movements and price changes. They are interested in understanding how the informed traders would trade to maximize their expected utilities if they choose to trade on their information in the option markets. Within this framework, informed traders cash flows are summarized as follows. Long call Long put Cash flow at time 0 -Q C * P C -Q P * P P Payoff at time T Prob(S=S U) = π U Q C * (S U X C) 0 Prob(S=S D) = π D 0 Q P * (X P S D) Where S stands for stock price and Q C(Q P) is the call (put) trading volume. Prob(S = S U) and Prob(S = S D) are the probabilities that the stock price will increase to S U and decrease to S D at some time T in the future, respectively. The magnitudes of price up-movement and down-movement are assumed to be the same. π U (π D ) denotes the probability that stock price increases (decreases). X C (X P) is the exercise price for call (put) option and P C (P P) refers to the call (put) premium. Given their private information and subject to an initial wealth constraint of W 0, informed traders will select optimal quantities of call and put to purchase to maximize their expected utilities. According their assumptions, it is found that the ratio of the probability of a price increase (π U ) to the probability of a price decrease (π D ) is equal to the ratio of the call trading value (Q CP C) to the put trading value (Q PP P). When the call put option trading VR, (Q CP C)/(Q PP P), is greater (smaller) than unity, stock return is more likely to be positive (negative). Thus, the VR is a credible measure for discerning good and bad news information embedded in option trades because it considers information reflected in both option trading volumes and premiums. Equation confirms that option volume alone does not fully reflect the market expectations and implies that stock returns should be positively related to VR. 9

12 Thus, we utilize the ratio of the Chicago Board Options Exchange (CBOE) total equity put to call trading volume as the representative measure of sentiment in the options market. This ratio is defined as the aggregate trading volume of put options to the aggregate trading volume of call options. E. Long-short S&P 500 futures (SentLS) Following studies that investigated Long-Short S&P 500 futures (Chen and Wang, 2010; Han, 2008), the SentLS is the net position of large speculators in S&P 500 futures which is released weekly by Commodity Futures Trading Commission. It is derived by the open interests of short large speculators divided by the sum of the open interests of long large speculators and the open interests of short large speculators. As it is derived from future trading data, we consider this sentiment proxy as the representative of sentiment in future market. Figure 2 illustrates the time series of changes of five sentiment measurements for the entire period from January 2006 to August It shows the relationship between the movement of differences in CDS spread and that of individual sentiment proxies. Among the five graphs, the option derived sentiment proxy shows the most identical movement with the CDS spread changes. [Figure 2 placed here] 3-3. Theoretical Factors A. Leverage ratio (Lev) According to a prior model (Merton, 1974), as Lev increases, the distance to default barrier becomes narrower. Thus, the default probability will be higher. We use the CRSP to gain market value of firm equity and Compustat to collect the quarterly book value of firm debt and preferred stock. Lev was calculated as follows: Leverage ratio(lev) = [book value(debt+preferred stock)] [book value(debt+preferred stock)+market value(equity)] (2) 10

13 B. Stock return volatility (Vol) Since a number of studies revealed a strong linkage between corporate bond spreads and stock return realized volatility or stock option implied volatility through empirical analysis, we implement stock return volatility as the substitute for firm value volatility while Collin-Dufresne et al. (2001) use S&P 500 index option implied volatility. A very recent literature document that the realized volatility can be better measure for the true latent volatility (see e.g. Fernandez-Perez, Fuertes and Miffre, 2016). By summing sufficiently finely sampled high-frequency returns, it is possible to construct ex-post realized volatility measures for the integrated latent volatilities that are asymptotically free of measurement error. From Anderden, Bollerslev, Diebold and Ebens (2001) and Anderden, Bollerslev, Diebold and Labys (2000a, 2000b, 2001a, 2001b), Andersen, Bollerslev, Diebold and Vega (2003), the realized volatilities are generally estimated using the intraday data, for example, 5-minute interval data. However, in our paper, it is difficult to use the intraday data for all individual stocks. Instead, we use exponentially weighed moving average (EWMA) volatilities as the alternative to the realized volatilities. Exponentially weighed average of daily squared returns over the previous three months is calculated with decay factor The reason why we assume decay factor 0.94 is that we follow Risk-metrics of JP Morgan assumption. The data source of stock price time series for each firm is CRSP. C. Risk-free rate (Rf) The 10-year maturity Treasury bond yield collected from the FRED dataset is used for the Rf. Longstaff and Schwartz (1995) reported a negative association between Rf and credit spread or default probability, since a high reinvestment interest rate will increase the future firm value Controlling Variables A. Term spreads (Term) The term structure of the yield curve is basically computed by deducting the two-year maturity treasury 11

14 yield from 10-year maturity treasury yield. Since Collin-Dufresne et al. (2001) discovered that term structure has the negative relation with credit spread measured by bond, we also use the term structure as the controlling variable. B. S&P 500 (SP) S&P 500 returns obtained from the CRSP are used as the proxy of business climate. It is well known that bull stock market represents a good business environment and increases the expected recovery rates of companies which lowers the CDS spreads. C. Smirk Collin-Dufresne et al. (2001) insisted that since negative jump in a corporation, which can be detected by option implied volatility smiles, can explain the probability of large negative jump, with the magnitude of a downward jump expected to result in an increase of credit spreads. In an empirical study, Zhang, Zhou, and Zhu (2009) reported that the volatility and jump risks of individual corporations using high-frequency equity prices predict a large portion of CDS premia. Likewise, we include smirk as the controlling variable. Smirk was calculated as follows. First, we select the options from standardized S&P 500 Index options of which a delta of puts and of calls exceed -0.5 and 0.5, respectively, and of which the remaining maturities are longer than a month. With the selected data and implied volatility curves from OptionMetrics, the volatility of the strike price is calculated by a linear regression of implied volatility on strike price: imvol(k) = a + bk (3) where K is the strike price and imvol( ) is an implied volatility function. The slope of smirk is eventually generated by the following formula: imvol(0.9s) - imvol(1.1s) (4) 12

15 where S is the current stock price. Panel A of Table 1 provides the variables defined in this study and their descriptions, including the specific data used in their estimation. The last column shows the expected results of the regression for differences of natural logarithm of CDS. In addition, for stationary check for the variables employed in the regression analysis, the unit root test was done before conducting the regression. We first generate the averages of CDS, Lev and Vol and test whether the data is stationary or non-stationary. The null hypothesis is that time series is non-stationary; H=0 means failing to reject the null hypothesis and H=1 means rejecting the null hypothesis. In Panel B and C of Table 1, we give the results with the test statistic including corresponding p-values and t-statistics for further information. Panel B represents the Augmented Dickey Fuller Test for the raw time series data of all variables. Except sentiment variables, SentAA and SentLS, the variables are non-stationary. Panel C indicates that the logarithm of CDS spread, theoretical variables and a battery of market sentiment proxies are all stationary in first differences. The statistics for the theoretical variables coincide with the result of Greatex (2008) that the variables suggested by structural model are stationary in the first differences. [Table 1 placed here] Panel A of Table 2 summarizes descriptive statistics of the log difference of CDS spreads and the explanatory variables in first difference used in the multiple regressions. For the firm-specific variables and CDS spread, cross-sectional averages are reported. The first four columns are for the basic statistics and the last three columns (skewness, kurtosis, and J-Q test) are for the information for the distribution. With the exception of the Rf, all variables were not skewed. However, several variables, especially three theoretical variables and SentLS, showed fatter tails than the normal distribution. As expected, these variables reject the J-Q test, of which null hypothesis is the data is normally distributed. Panel B of Table 2 shows the correlation of all independent variables. The maximum correlation is

16 between Rf and term spread, which is still not significant large correlation. This result is intuitive because term spread itself is rendered from Rf. The correlation coefficients among sentiment variables are relatively low, (The highest correlation coefficient is 0.28 between SentBW and SentMi) which means it is worthwhile to examining all sentiment measurements that we provide. [Table 2 placed here] 4. Analytical Framework 4-1. Structural model (Merton model) Since we utilize differences of logarithm CDS spread as the proxy for the credit risk in this study, we test the model of Ericsson et al. (2009) by replacing differences of CDS spread with this new proxy and investigate the validity of structural model. Employed regression equations are as follows: Δlog (CDS i )= a i + β lev Δlev i + ε i (5) Δlog (CDS i )= a i + β vol Δvol i + ε i (6) Δlog (CDS i ) = a i + β rf ΔRf + ε i (7) Δlog (CDS i )= a1 i + β1 lev Δlev i + β1 vol Δvol i + β1 rf ΔRf + ε1 i (8) The regression equations (5),(6) and (7) verify that each theoretical variable can explain the credit spread and the model (8) tests the explanatory power of theoretical variables in total. We report the average of 285 coefficients and the Mean Group t-statistic of Pesaran and Smith (1995), which is computed by dividing the mean of individual coefficients by the standard deviation of coefficient values scaling by squared root of number of firms. The Mean Group (MG) estimation has been employed in many other studies such as Collin-Dufresne et al. (2001), Ericsson et al. (2009), and Brun-Aguerre, Fuertes and Phylaktis (2012). 14

17 4-2. Sentiment Model A. Firm level To be consistent with previous studies, such as Collin-Dufresne et al. (2001), Ericsson et al. (2009) and Kim et al. (2013), we implement the following regression equations adding the sentiment variable to theirs: Δlog (CDS i )=a2 i + β2 sent ΔSent + ε2 i (9) Δlog (CDS i )=a3 i + β3 lev Δlev i + β3 vol Δvol i + β3 rf ΔRf + β3 sent ΔSent + ε3 i (10) Δlog (CDS i )=a4 i + β4 lev Δlev i + β4 vol Δvol i + β4 rf ΔRf + β4 sent ΔSent + β4 term Δterm +β4 sp Δsp + β4 smirk Δsmirk + ε4 i (11) To examine whether each sentiment can predict spread variations, we use univariate regression analysis for each sentiment as equation (9). Then we perform the regression of CDS spread on the three theoretical factors for each individual firm and sentiment variable i, as represented in equation (10), and averaged the 285 coefficients. To test the robustness, we also add the sentiment variable to the base regression of Collin- Dufresne et al. (2001) in equation (11). For the t-statistics, we divide the mean of each beta by the standard deviation of 285 firms, as adopted previously (Pesaran and Smith, 1995). B. Portfolio level We also conduct portfolio level regression to examine whether each sentiment can predict spread variations after the idiosyncratic risk is controlled (Kim et al., 2013). We first use univariate regression for each sentiment as equation (12). Then we perform the regression of CDS spread on the three theoretical factors suggested for each of the 25 portfolios and sentiment variable, as represented in equation (13). We also add the sentiment variable to the base independent variables of Collin-Dufresne et al. (2001) as equation (14). Δlog (CDS p )=a p + γ sent ΔSent + ε p (12) 15

18 Δlog (CDS p ) = a1 p + γ1 lev Δlev p + γ1 vol Δvol p + γ1 rf ΔRf + γ1 sent ΔSent + ε1 p (13) Δlog (CDS p ) = a2 p + γ2 lev Δlev p + γ2 vol Δvol p + γ2 rf ΔRf + γ2 sent ΔSent + γ2 term Δterm +γ2 sp Δsp + γ2 smirk Δsmirk + ε2 p (14) 5. Empirical Results 5-1. Structural Factors Table 3 presents the results of the regression analysis between variations of CDS spread and traditional structural variables according to Merton (1974) and validated by Ericsson et al. (2009). The first three columns display the coefficients of univariate regressions. Lev, Vol, and Rf are statistically significant in predicting CDS spread changes and the adjusted R 2 s are 12%, 5% and 7%, respectively. The last column of the table displays the results of multivariate regression incorporating all three factors. Based on the information from the adjusted R 2, the structural model predicts only about 20% of log CDS spread differences. This result agrees with the finding of Ericsson et al. (2009), who reported that these three determinants explained approximately 23% of the CDS spread difference. Our result of explanatory power of the variables is about 3% lower compared to Ericsson et al. (2009). The difference may reflect the method of differencing CDS spreads, that is, a statistical difference between naïve difference in CDS premia and difference in logarithm of CDS premia, and from the period of sample. [Table 3 placed here] 5-2. Investor Sentiment Factors A. Firm-level regression results Table 4 reports the result of basic regressions between each selected sentiment measurement and individual firm s CDS spread changes. Overall, sentiment proxies of diverse markets are significant. 16

19 ΔSentMi, ΔSentAA and ΔSentBW are negatively significant, while ΔSentEPCR and ΔSentLS are positively significant. Only ΔSentMi lose significance showing altering sign from negative to positive when the controlling variables are added to the regression. Among the valid factors, ΔSentEPCR displays a significantly positive effect on the variations in CDS spread changes, which indicates that larger changes in ratio of trading volume of put to trading volume of call is associated with larger changes in CDS spreads. Univariate regression indicates that ΔSentEPCR is a superior predictor than the volatility, a firm-specific factor; the adjusted R 2 s of volatility and ΔSentEPCR are 5.92% and 6.41%, respectively. Moreover, ΔSentEPCR enhances the explanatory power by 2% compared to that of the three-factor theoretical model. [Table 4 placed here] B. Firm-level regression results for investment grade firms vs. non-investment grade firms In order to examine whether sentiment affects CDS premium variations more precisely, CDS is categorized by ratings into two groups: investment-grade firms CDSs, ranging from AAA to BBB, and non-investment-grade firms CDSs, ranging below BBB, as previously described (Kim et al., 2013; Chen and Want, 2010). Table 5 presents the result of the regression model that examines the explanatory power of sentiment on the investment-grade firms CDSs for 44 months from January 2006 to August Except for ΔSentMi, sentiment proxies are significant and align with the expected signs. However, their magnitudes of coefficient and t-statistics markedly decrease compared to basic regression; most of the adjusted R 2 s are either negative or nearly zero in univariate regression, except for ΔSentMi. Only the coefficient resulting from ΔSentMi is statistically significant. Its adjusted R 2 is as high as the theoretical factors, but its coefficient shows opposite signs in the model with controlling variables. Overall, sentiment barely explains high-graded CDS spread changes. [Table 5 placed here] 17

20 Table 6 presents the result of the regression model that examines the explanatory power of sentiment on the non-investment-grade firm s CDSs for the entire sample period. Comparing the results of two divided CDS ratings, unlike the results from other studies (Greatrex, 2008; Chen and Wang, 2010), there is no evidence that theoretical variables can explain CDS premium changes for speculative firms better than for investment grade firms. However, one sentiment model performs better in predicting differences in speculative CDS by showing markedly higher R 2 in models on speculative CDSs than investment graded CDS. For the non-investment grade CDSs, the univariate regression of ΔSentEPCR predicts about 8% of CDS premium differences and is much better than results from the univariate regression on theoretical variables of Vol and Rf of which the adjusted R 2 is both only about 5%. 10 [Table 6 placed here] C. Portfolio-level regression results A 5x5 portfolio is constructed based on five leverage ratio and five stock return volatility as previously described (Kim et al., 2013). Table 7 presents summary statistics of each portfolio and the cross-sectional mean of correlation coefficients of all variables employed in models for portfolio. CDS spread grouped by rankings of leverage ratio and volatility is highly correlated (about 80%) to the portfolio leverage ratio. [Table 7 placed here] Table 8 reports the linear regression results for 25 portfolio matrix by quintiles based on leverage ratio 10 The univariate regression results on theoretical variables for non-investment grade firms are not reported in table 6 but we can provide them if requested. 18

21 and stock volatility, each with 44 monthly CDS spread quotes over the full sample period. Overall explanatory power is escalated than that in the firm level regression by approximately 13%, from about 20% to about 33% in the theoretical model. The coefficients of sentiment proxies in all sentiment models are statistically significant. Especially, the magnitude of coefficient is the biggest among five proxies and is economically and statistically significant in the model incorporating ΔSentEPCR. Moreover, the adjusted R 2 is increased by 3% over that of another model (Merton, 1974). Nevertheless, ΔSentMi is positively valid, with an opposite sign to the expected signs derived from the study of Tang and Yan (2010), which reports that the CBCC Index shows a similar move with ΔSentMi, and it is negatively correlated to CDS. This discrepancy reflects the regressions, which is divided by the period in the next tables. [Table 8 placed here] For portfolio level regression, the adjusted R 2 of individual portfolios are determined in Table 9. For every result from the regression of sentiment model at the portfolio level, the most stable portfolio in which CDS spread with the lowest Lev and Vol are assigned has the lowest adjusted R 2 and shows general ascending trend when the Lev and Vol grows. In other words, the most volatile portfolio having the highest leverage ratio and volatility is explained best by the models. The difference of the adjusted R 2 between the lowest and highest Lev-Vol portfolios ranges from 41.33% in the SentAA model to 47.20% in the Merton (1974) model. [Table 9 placed here] The average adjusted R 2 for each of the five Lev values and the five Vol of structural and extended models are determined including controlling variables in Table 10. Similar to Table 9, the explanatory power increases when the firm s Vol and Lev increase. More specifically, we examine the ability of models 19

22 to predict CDS spread difference by the Lev and Vol criteria. The steadily ascending pattern in the explanatory power by portfolio order is obtained by dividing by Lev, especially after controlling of the model only with the exception of results from ΔSentBW and ΔSentLS regressions. In the sentiment models, the gap between the adjusted R 2 of two polarized portfolios by Lev is bigger than this of two polarized portfolios by volatility. [Table 10 placed here] We have three main conclusions from portfolio level analyzing. First, when idiosyncratic risk is eliminated, the explanatory power increases, in agreement with Kim et al. (2013). ΔSentEPCR shows extremely excellent predicting power compared to the structural model. Finally, the 5X5 portfolio has a large adjusted R 2 gap between the lowest and highest Lev-Vol portfolio, with the portfolio being better organized under the Lev criterion, rather the Vol criterion. This is an intuitive result indicating that CDS spread changes that reflect default risk by the degree of the Lev and not by the degree of the firm s Vol, which embeds the credit risk and other risk factors including market and operational risks. This line of reasoning echoes Cesare and Guazzarotti (2010), who reported that Lev explains CDS spread variances better than Vol. D. Portfolio-level regression results for the pre-crisis period vs. the crisis period We separate the sample period into two parts: prior to the global financial crisis term from January 2006 to August 2007 and during the global financial crisis from September 2007 to August Table The global crisis period is somewhat inconsistently defined on the literature concerning the corporate CDS spread. For example, Coro, Dufour and Varotto (2013) designate the crisis subsample as April 2007 to July 2009 and Galil, Shapir, Amiran and Ben (2014) do as August 2007 to June Tang and Yang (2013) set the crisis period as July

23 reports the result of various models that have examined the relationship between each selected sentiment measurement and CDS spread changes grouped by five Lev and five Vol values during the pre-crisis 20- month period from January 2006 to August The explanatory power is lower compared to the aggregate period regressions. The structural model performs disappointingly, reporting an adjusted R 2 of only 5.9%, and other models embedding sentiment variables show no more differentiation, except for the regression on ΔSentEPCR. The SentEPCR model has an explanatory power of approximately 15%, which is almost three times higher than that of theoretically structured model. This is a highly significant and large magnitude of coefficients, which are also consistent with the expected signs. This result implicates ΔSentECPR as the best candidate among various sentiment measurements derived from different markets. On the other hand, other sentiment factors perform poorly to predict CDS spread changes. ΔSentMi and ΔSentAA shows opposite coefficient signs to the expected signs and ΔSentBW has very low explanatory power at 1.2%. [Table 11 placed here] Table 12 presents the results of regression models that study the relationship between each selected sentiment proxies and CDS spread changes grouped by five Lev values and five stock price Vol values during the 24-month crisis period from September 2007 to August While the overall explanatory power of variables is lower and the perplexing signs are observed in stable period, the result for the turmoil term convinces our hypothesis. All five sentiments are economically significant with consistent signs. However, unlike previous results in the aggregate period and stable period, the four factor model with to March 2009 and Kim, Park and Noh (2013) do as September 2007 to August Especially, for splitting subsamples, Kim et al. (2013) conduct volatility break test of Inclan and Tiao (1994) on unit recovery claim extracted from U.S. corporate CDS spreads. Since their division is based on econometric methodology, differently from the other literature, we define the crisis period according to Kim et al. (2013). 21

24 ΔSentAA predicts 48.77% of changes in CDS premium, which is the best prediction among the models including proxies for sentiment and is about 7.5% higher than the theoretical model. During the crisis period, even ΔSentBW, which is insignificant in another stage, displays a strongly significant negative coefficient. We conclude that sentiment models perform better in a crisis period rather than in a stable period. Our results are consistent with a prior explanation. Based on an idea espoused by Miller (1997), Stambaugh et al. (2012) insisted that when sentiment exists widely in the market, the limitation in shortsale has a major role in increasing the degree of the existence of the mispriced asset. The authors further opined that a higher level of sentiment, in turn generally considered as the downturn, fuels overpricing. On the other hand, due to the absence of roles of short selling in a bullish market, sentiment does not provide a convincing explanation in this period. Based on the argument of Stambaugh et al. (2012), we interpret that in bad economic states overpricing renders the hedging of CDSs by equity or equity options difficult and moves up CDS spreads. Thus, sentiment explains the CDS spreads better in a period of turmoil. [Table 12 placed here] There are two main findings. First, the most notable finding is that even though both stock return and the implied volatility of equity option are effective factors to explain CDS spread, ΔSentBW derived from stock market performs poorly in predicting CDS whereas ΔSentEPCR is superior among the several sentiment candidates, with the exception with predicting CDS spread differences for non-investment grade firms. This may indicate that the sentiment in options market predict CDS spread variation rather than sentiment in stock market, because the options market is also sensitive to the volatility of asset return just as CDS spread, but the stock market is sensitive to the direction of asset values. Moreover, it aligns with the result of other studies, which reveal the price discovery role of derivative market in stock price. Since low initial cost and high leverage is more prevalent in derivatives market, it responses to the information 22

25 efficiently. Therefore, in terms of information efficiency, the derivatives marker performs better than the stock market due to the feature. Second, sentiment explains CDS spread changes better in a period of turmoil. The multiple-equilibria model as theoretical background of Saka et al. (2015) can explain this result. According to the model, during the crisis period, markets may not always behave optimally and thus move without any major change in fundamental factors. Further, a self-fulfilling liquidity crisis may be caused by the decisions of panic-driven investors. Therefore, we interpret the result as that sentiment from panic-driven investors may lead liquidity contraction of CDS market during the financial crisis and thus the sentiment variables can explains CDS spread changes better during the global financial crisis period than in the normal period. 6. Conclusion We investigated whether investor sentiment can predict CDS premium changes using several proxies. We verify our hypothesis that investor sentiment can explain CDS spread changes. We explore which sentiment measurement can be the most effective determinant. In general, most sentiment proxies are valid to explain CDS spread changes and change in equity put/call ratio is the best predictor due to the characteristic of options market. In portfolio level regression analysis, overall explanatory power is increased compared to individual level regression analysis as Kim et al. (2013) documented. In addition, change in equity put/call ratio shows outstanding explanatory power, which is nearly three times greater than the structural model. When precisely observing the result of each portfolio, we find that sentiment explains differences in CDS spread best in the group whose Lev and Vol are highest and vice versa. Also, the gap of ability to explain between the two polarized portfolios is dramatically large, ranging from 41.33% to 47.20%. We divide the sample period into two stages: the pre-crisis period from January 2006 to August 2007 and the crisis period from September 2007 to August 2009, to test the explanatory power of sentiment by business cycle. We find that our suggested models explain CDS spread changes much better in a turmoil 23

26 period than in a stable period. The results align with the hypothesis that sentiment explains the mispriced CDS spreads better in turmoil period. Stambaugh et al. (2012) insisted that when sentiment exists widely in the market, the limitation in short sale contributes a major role to increase the degree of the existence of the mispriced asset. The period of higher level of sentiment, in turn generally considered as the downturn, the more overpricing presents. It can be interpreted as that in bad economic states overpricing renders the hedging of CDSs by equity or equity options difficult and moves up CDS spreads. Thus, sentiment explains CDS spreads better in turmoil period. On the other hand, due to the absence of roles of short selling impediment in bullish market, sentiment cannot explain well in the period. In addition, the results support the multiple-equilibria theory. In this theory, markets may not always take optimal behaviors during the crisis period and a self-fulfilling liquidity crisis may be led by panic and fear of investors without any massive change in fundamental factors. Therefore, sentiment from panicdriven investors may cause liquidity contraction of CDS market during the crisis period and the sentiment variables can explain CDS spread changes better during the global financial crisis period than in the normal period. 24