CDS Trading and Stock Price Crash Risk

Transcription

1 CDS Trading and Stock Price Crash Risk ABSTRACT When pessimistic investors suspect that corporate managers are withholding bad news, the market for credit derivatives provides them an alternative trading venue, which facilitates the revelation of negative information. Supporting this information transmission view, we find that stock price crashes are less frequent after the inception of credit default swaps (CDS) referencing the firm s debt. This crash-reduction effect is stronger when the managers of the firms actively manipulate the information in their corporate reports or when they can benefit from higher share prices. Our study offers novel evidence on how financial innovations in the debt market help improve price discovery in the equity market. Keywords: CDS trading; stock price crash risk; information spillover; earning management; compensation incentives JEL Classification: G10; G14; G32; M41

2 I. INTRODUCTION The credit default swap (CDS), a major financial innovation in the multi-trillion-dollar credit derivatives market, has a short but extraordinary history. A CDS is similar to an insurance contract wherein a buyer pays periodic insurance premiums to a seller for coverage of loss from adverse credit events such as bankruptcy. CDSs help standardize trading on credit risk, making it more feasible. Although at first glance a CDS may seem a side bet on a firm s underlying assets that has no effect on firm fundamentals, recent studies document that CDS trading has significant effects on financial markets, corporate policies, and the real economy. 1 Since the 2008 credit crisis, regulators, practitioners, and academics have debated the role of CDSs (e.g., Stulz 2010). CDS trading has often been blamed for the excessive risk-taking and the dramatic stock price crashes of many firms, such as Lehman Brothers, that occurred during the 2008 subprime financial crisis. George Soros, a prominent hedge fund manager, was reported as calling for an outright ban on CDS trading (Reuters, June 12, 2009). While CDSs have borne the brunt of the blame for the 2008 crisis, there is little factual evidence to date to support this allegation. In this study, we empirically examine how CDS trading affects the risk of stock price crashes. Previous studies of CDSs offer mixed predictions for the relationship between CDS trading and stock price crash risk. One stream of research argues that CDS-referenced firms might take on greater risk because the inception of CDS trading affects the creditor-borrower relationship. Specifically, CDS contracts allow the focal firm s creditors to hedge their credit risk, which increases their willingness to supply additional credit to the focal firm. The firm, in turn, is able to undertake more risky projects. Further, the lenders whose credit exposures are hedged by 1 See Augustin, Subrahmanyam, Tang, and Wang (2014, 2016) for a discussion of the CDS market s development and a review of the literature. 1

3 corresponding CDS contracts may not be as vigilant in monitoring borrowers, potentially allowing firms to engage in even more risk-taking. Bolton and Oehmke (2011) build a theoretical model of the empty creditor to formalize discussions on the creditor incentives associated with CDS trading. Empirically, Subrahmanyam, Tang, and Wang (2014) find that firms with CDS trading on their debts are more likely to be downgraded and to file for bankruptcy. Prior studies have shown that weak monitoring by lenders can be detrimental to shareholders (Diamond 1984; Besanko and Kanatas 1993; Kim and Zhang 2015). Hence, to the extent that CDS trading results in more risktaking, especially excessive risk-taking, one might expect a positive association between CDS trading and stock price crash risk. However, CDS trading may have an important informational role: it could mitigate managers attempts to hide bad news from various stakeholders, including shareholders. Managers often withhold information, especially negative news (Kothari, Shu, and Wysocki 2009). Withholding negative information can result in large stock price drops (crashes) when the news is eventually and cumulatively revealed (Jin and Myers 2006; Hutton, Marcus, and Tehranian 2009). Stulz (2010, p. 73) notes that one of the useful functions of CDS trading is revealing useful information about credit risk of the reference firms. Not surprisingly, due to the asymmetric payoff between debt and equity, CDS traders are more concerned about negative information (Acharya and Johnson 2007). Prior studies provide evidence that CDS spreads reveal a firm s private information (Zhang and Zhang 2013; Lee, Naranjo, and Sirmans 2016; Han, Subrahmanyam, and Zhou 2016) and lead the other markets in price discovery, especially when there is hidden negative information. Batta, Qiu, and Yu (2016) show that analyst forecasts become more accurate after CDS trading. Hence, to the extent that CDS trading facilitates the flow of information (especially negative 2

4 information that firms attempt to hide) into the equity market and the corresponding price discovery, one may expect CDS trading to be negatively associated with stock price crash risk. In sum, the risk-taking argument suggests a positive relationship between CDS trading and future stock price crashes, whereas the information revelation argument predicts a negative relationship. As noted above, there is significant academic and practitioner interest in how CDS trading contributes to the stability of financial markets. Hence, the relationship between CDS trading and stock price crash risk is an important empirical question that we address in this study. We use a long time-series dataset to identify the relationship between the inception of CDS trading and stock prices. The dataset includes transactions and quotes drawn from multiple leading CDS data sources, including the GFI Group and Markit, which other researchers have also used, including, among others, Subrahmanyam, Tang, and Wang (2014, 2017) and Li and Tang (2016). We construct three standard measures of stock price crash risk, including negative skewness, down-to-up volatility, and the probability of actual stock price crashes (e.g., Jin and Myers 2006; Hutton et al. 2009; Kim, Li, and Zhang 2011a, 2011b). Our first key finding is that the inception of CDS trading significantly reduces the future occurrence of extremely negative stock returns. This finding remains robust after controlling for the known determinants of stock price crash risk. The decrease in stock price crash risk is both statistically significant and economically meaningful for all three stock price crash risk measures. For example, CDS trading in the current year reduces the probability of experiencing a stock price crash in the following year by about 12 percent. 2 We consider the possibility that the timing of the inception of CDS trading for firms may not be randomly determined, which may pose endogeneity concerns. The first is that the expectation 2 12 percent represents the percentage change in the probability of experiencing an actual stock price crash relative to the unconditional mean. 3

5 of future crash risk may affect the probability of initiating CDS trading. The second concern is that the association between the initiation of CDS trading and stock price crash risk may be driven by some latent, possibly omitted, variables. We address both of these endogeneity issues using three additional tests: a difference-in-difference analysis, an instrumental variable approach, and placebo tests. We find that the negative effect of CDS trading on stock price crash risk is robust to these tests. Specifically, to address the first endogeneity concern about reverse causality, we perform a difference-in-difference analysis on the treatment and control groups. We distinguish between the two groups using propensity scores of the likelihood of introducing CDS trading. We find that CDS trading has a significantly negative influence on stock price crash risk. Further, we use the following as instrumental variables: the amount of foreign exchange (FX) derivatives used by a firm s lender for hedging purposes, standardized by the lender s total assets and the maximum percentage of CDS-referenced borrowers among all of a firm s lenders. Both the use of FX derivatives to hedge currency exposure and that of CDS contracts to hedge credit risk exposure reflect lenders preference for using financial derivatives to hedge their position, as opposed to addressing concerns about an individual (borrower) firm s stock price crash risk. In the first-stage regression, using the indicator of CDS trading as the dependent variable, we find a significant and positive effect for both the use of FX derivatives and CDS contracts on the initiation of CDS trading. In the second-stage regression, we find significant negative coefficients for the predicted indicator of CDS trading extracted from the first-stage regression, lending more credence to the possibility that CDS trading has a negative effect on stock price crash risk. To mitigate the second endogeneity concern about the possibility of latent omitted variables, we use placebo tests. If a negative relationship between CDS trading and stock price crash risk is 4

6 driven by some unobserved factors, we predict that we would observe this relationship regardless of the exact timing of the initiation of CDS trading. To test this, we incorrectly assign the timing of CDS inception. We find that the negative effect of CDS trading on stock price crash risk disappears when the CDS inception date is incorrectly assigned to one or two years before the actual date, strengthening our main findings. The negative effect of CDS trading on stock price crash risk suggests that information revelation, as opposed to risk-taking, is the dominant channel through which CDS affects the stability, in terms of downside risk, of a firm s stock price. The revelation of information, especially of private information that spills over into price discovery in the equity market, is likely to be more influential when a firm is trying to hide bad news. Hence, to provide further analysis of the information revelation channel, we examine whether CDS trading has a greater effect on reducing stock price crash risk when the firm has a greater tendency to hide bad news. Firms have multiple ways of hiding bad news. They can do so by being more opaque in their financial reporting (Hutton et al. 2011), less conservative in their financing accounting (Kim and Zhang 2015), or by opportunistically disclosing news via management forecasts (Hamm, Li, and Ng 2016). We conduct interaction analyses and find results that are consistent with an information revelation role for CDS trading. First, when firms try to hide bad news by being more opaque in their financial reporting, information revelation from the CDS market leads to timely price discovery in the equity market. This, in turn, reduces the stock price crash risk. Indeed, we show that the negative association between CDS trading and stock price crash risk is greater for firms with more opaque financial reporting. Second, firms choose the timing of when to recognize their losses, with conservative firms opting for more timely recognition. We find that the negative association between CDS trading and stock price crash risk is smaller for firms that are more 5

7 conservative in their financial reporting. Third, managers may attempt to boost share prices by making optimistic management forecasts. We find that the negative association between CDS trading and stock price crash risk is stronger for firms with more management forecasts. Importantly, the effect of CDSs on crash risk is beyond the direct effects of accounting conservatism or voluntary disclosure. Firms give different compensation packages to their managers, which have different implications for managers incentives to disclose information and affect stock prices. Kim, Li, and Zhang (2011b) find that equity-based compensation is positively associated with stock price crash risk because such compensation incentivizes managers to conceal bad news. Along this line, we find that the negative association between CDS trading and stock price crash risk is stronger for firms with high equity-based compensation. This result further highlights the importance of CDS trading as a source of information that can mitigate managers attempts to hide bad news. Our study contributes to the literature in the following ways. First, by demonstrating that CDS trading reduces the likelihood of future stock price crashes, our study adds to the literature that documents information spillover from CDS markets to related financial markets (e.g., Acharya and Johnson 2007; Lee, Naranjo, and Velioglu 2017). Our findings have important implications for investors, corporate executives, and policy makers. Second, this study advances the literature on the determinants of stock price crash risk. Our research provides new evidence that credit derivatives trading can mitigate stock price crash risk, over and above the other factors identified by previous studies. Third, adding to Martin and Roychowdhury (2015), the new findings documented in this paper further our understanding of the effects of CDS trading on the information environment. One implication of our findings is that CDS trading can act as a crossmarket mechanism that mitigates the agency problem of hiding bad news from public shareholders. 6

8 The rest of this paper is organized as follows. Section 2 presents an overview of the arguments linking CDS trading to stock price crash risk, while Section 3 describes the statistics of our sample data. Section 4 discusses our main empirical findings with regard to the effect of CDS trading on stock price crash risk and addresses the endogeneity concerns. Section 5 provides a comprehensive analysis of the channel through which information is revealed. Section 6 concludes the paper. II. RELATED LITERATURE AND HYPOTHESIS DEVELOPMENT In recent years, the externalities of CDS trading have been extensively examined. Nevertheless, no prior study has been able to link CDS trading to stock price crash risk. In this section, we discuss the potential connections between CDS trading and stock price crash risk from two key perspectives: risk-taking and information revelation. First, CDS trading can exacerbate risk-taking by the reference firms. As a CDS contract provides coverage for a lender s losses upon default, lenders are more willing to lend to a firm when CDS trading is available (Bolton and Oehmke 2011). The lender is also likely to allow referenced firms to take on more risk and to be less concerned about losses arising from agency problems. Ashcraft and Santos (2009) and Saretto and Tookes (2013) find that for risky firms, the cost of both debt and leverage increase after the onset of CDS trading. Chang, Chen, Wang, Zhang and Zhang (2017) find that when CDS trading is available, firms pursue riskier innovations, as CDS contracts enhance both the lender s risk tolerance and the borrower s risk-taking in the innovation process. Subrahmanyam, Tang, and Wang (2014) show that the initiation of CDS trading results in a larger lender base and an increase in bankruptcy risk due to a higher empty creditor problem and an increase in the probability of coordination failure in the face of financial distress. The incremental default probability due to CDS trading implies a higher probability of stock price crash risk in the equity market. Martin and Roychowdhury (2015) argue that CDS 7

9 trading reduces lenders incentives to continue monitoring borrowers. They find that borrowers accounting conservatism decreases for firms with traded CDSs, consistent with weaker monitoring by lenders. Kim and Zhang (2016) show that conservatism is negatively related to stock price crash. Furthermore, Hong, Ryou, and Srivastava (2017) argue that when corporate managers compensation is risk sensitive, they increase corporate risk-taking after the inception of CDS trading. Hence, the association of CDS contracts with more risk-taking by borrowers and weaker monitoring by lenders implies a higher stock price crash risk. Second, CDS trading may carry information relevant to other markets (Stulz 2010). As equity and CDSs are related securities that are linked through common firm fundamentals, according to structural models (Merton 1974), concurrent equity and CDS trading should generate spillover effects on each other and facilitate the capitalization of information into financial markets. Prior studies have found that because of the CDS markets concentrated structure (Atkeson, Eisfeldt, and Weill 2013), 3 insider information flows into financial markets through CDS trading, especially when a firm experiences adverse credit news and has a greater number of bank relationships (Acharya and Johnson 2007). Moreover, using a sample of short-sale-banned stocks in 2008, Ni and Pan (2011) find a cross-sectional predictability of CDS signals for future stock returns. Also, Blanco, Brennan, and Marsh (2005), Berndt and Ostrovnaya (2008), Qiu and Yu (2012), and Kryzanowski, Perrakis, and Zhong (2017) document that the CDS market leads the price of other related securities. Recently, Lee, Naranjo, and Sirmans (2016) provide evidence indicating that information flows between CDS and equity markets; they show that the relative past stock and CDS returns significantly affect the cross-section of stock return momentum. Further, Batta, Qiu, 3 Atkeson et al. (2013, Figure 2) show that as of the end of 2011, only 12 of the largest bank holding companies trading in derivatives had trading assets in excess of $5 billion. 8

10 and Yu (2016) show that CDS spreads contain information that is useful for both equity and credit rating analysts. They find that post-cds trading, the dispersion and error of earnings-per-share forecasts are reduced and that downgrades by both types of analysts become more frequent and timely before large negative earnings surprises. Hence, by facilitating the process of capitalizing bad news into equity price, CDS trading reduces a firm s ability to hoard negative information and results in the lower likelihood of an increase in stock price crash risk. Given the opposite predictions from above discussions, the relationship between CDS trading and stock price crash risk becomes an open empirical question. The CDS-risk-taking channel predicts a positive association, whereas the CDS-information-revelation channel predicts a negative one. We provide empirical evidence on those hypotheses, followed by further analyses of the dominant one. If CDS trading does impose an effect on stock price crash risk, through either the risk-taking or information-revelation channels, then the conventional determinants of stock crash risk such as financial reporting opacity, accounting conservatism, voluntary disclosure, and executive compensation incentives may be either enhanced or attenuated. To examine the moderating effects of the new determinant, we study the interaction effects between CDS trading and these determinants. III. DATA AND EMPIRICAL METHODS CDS Data We assemble a composite dataset to identify CDS trading using both CDS transaction data from CreditTrade and the GFI Group and CDS quotes from the Markit Group. 4 The actual CDS transactions reflect a CDS s agreed-on price between counterparties, whereas CDS quotes show 4 Similar data are used by Subrahmanyam et al. (2014) and Li and Tang (2015), among others. 9

11 the CDS contracts sell-side offering price. Due to the limited number of transactions in the CDS market, CDS quotes are used to provide continuously complementary information about the focal firms. Hence, the combination of transactions and quotes provides a full picture of CDS activities and reveals information about the focal firms. We focus on the single-name CDS contracts in the U.S. Specifically, we use CreditTrade data, which covers the period from June 1997 to March 2006; GFI Group data for the period from January 2002 to April 2009; and Markit data, which covers the period from August 2001 to December Therefore, our composite dataset covers continuous CDS activities from 1997 to The overlapping time periods allow us to validate the data quality for each source. In our baseline analysis, we use information about the inception of CDS trading or CDS quotes to assess the changes in the stock price crash risk with the onset of CDS contracts. [Please insert TABLE 1 about here] Stock Price Crash Risk Measures We use the negative conditional firm-specific skewness of weekly returns (NCSKEW) as a primary proxy for firm-specific crash risk (Hutton et al. 2009; Kim and Zhang 2011a, 2011b; Kim, Li, and Li 2014). As alternative measures, we adopt the down-to-up volatility (DUVOL) and actual probability of stock price crashes (CRASH) to check the robustness of our results. The stock crash risk proxies are constructed using stock returns from CRSP. To avoid a lookahead bias and ensure that our analysis of stock price crash risk only considers the financial data available to investors, we follow Kim et al. (2011a, 2011b), among others, and use weekly returns for the 12-month period ending three months after the firm s fiscal year-end. Then, for all the firms, 10

12 we regress the weekly stock returns for a year on the value-weighted market return in the current week, two weeks forward, and two weeks back, as follows: where r it, r r r r r r. i, t i 1, i m, t 2, i m, t 1 3, i m, t 2 4, i m, t 1 5, i m, t 2 i, t is the stock return for firm i in week t, r mt, (1) is the return of CRSP s value-weighted market index in week t, and it is an error term. We use Equation (1) to break down the total return into systematic and firm-specific components after introducing the lead and lag returns to account for non-synchronous trading. The natural logarithm of one plus the residual in Equation (1), log (1+ it, ), proxies for the firm-specific weekly return for firm i in week t ( W it, ). We calculate the NCSKEW by taking the negative of the third moment of the firm-specific weekly returns, W it,, for each sample year divided by the standard deviation of the firm-specific weekly returns raised to the third power. Specifically, we calculate NCSKEW for each firm i in year t as follows: 3/2. (2) 3/2 3 2 NCSKEWi, t n( n 1) Wi, t / ( n 1)( n 2) Wi, t where W it, is as defined above and n is the number of weekly return observations in year t. A higher negatively skewed return distribution (i.e., a higher value for NCSKEW) indicates a higher crash risk. The first alternative stock price crash proxy, DUVOL, is calculated as the natural logarithm of the standard deviation of the weekly stock returns, W it,, during the weeks in which W, is lower than its annual mean (down weeks) over the standard deviation of the weekly stock returns, W it,, it 11

13 during the weeks in which W it, is higher than its annual mean (up weeks). Specifically, DUVOL for each firm i in year t is calculated as follows: 2 2 DUVOLi, t log ( n u 1) Wi, t / ( nd 1) Wi, t DOWN UP. (3) where n u is the number of up weeks and n d the number of down weeks. A higher value for DUVOL indicates a higher crash risk. The second alternative proxy for stock price crashes, CRASH, is an indicator with a value equal to one if a firm experiences a stock price crash(es) in a year and zero otherwise. A stock price crash is defined as an extremely negative weekly stock return that is below the mean of the firm-specific weekly returns in a fiscal year by a standard deviation of 3.2. This standard deviation indicates approximately 0.1 percent in a normal distribution. Descriptive statistics We extract the equity information from CRSP, the firm s fundamental information from COMPUSTAT, analysts forecast data from I/B/E/S, loan data from Dealscan, corporate bond data from Mergent FISD, and institutional holding data from Thomson Reuter s 13f dataset. After merging these datasets and removing missing observations, we have 761 firms with CDS contracts for the period. As shown in Table 1, the majority of the CDS contracts were initiated in 2000 and [Please insert TABLE 2 about here] In contrast to the whole sample, CDS firms, on average, are of larger size and have higher leverage, higher profitability, and a greater mean of firm-specific weekly stock returns with a lower standard deviation. The percentage of firms experiencing actual stock price crashes is greater 12

14 during financial crises (e.g., the bursting of the dot-com bubble in 2002 and the subprime financial crisis in 2009). Moreover, the CDS-referenced firms in our sample period experience, on average, a lower percentage of actual stock price crashes than do non-cds firms. Nonetheless, the means of NCSKEW and DUVOL of the CDS-referenced firms are greater than those for the whole sample. To isolate the effect of other known determinants of firm-specific crash risk (Jin and Myers 2006; Hutton et al. 2009; Kim et al. 2011a, 2011b, 2015; among others), we control for firm-specific stock returns, the firm-specific standard deviation, size, the market-to-book ratio, leverage, profitability, and the opacity of the financial reports. Detailed descriptions of the variables are available in the appendix. IV. MAIN EMPIRICAL RESULTS This section reports our empirical findings on the effect of CDS trading on individual firms stock price crash risk. First, we present our baseline panel regression results. We then address endogeneity concerns related to the selection of CDS trading and possible omitted variables. CDS Trading and Stock Price Crashes To examine the effect of CDS trading on stock price crash risk, we create a multivariate regression model that links our crash risk measures in year t to the indicator of CDS trading in year t-1 and to a set of control variables in year t-1: Stock Crash Risk t = α 0 + α 1 CDS Active t 1 + β i ControlVariable i,t 1 + ε t. (4) where Stock Crash Risk t refers to the stock price crash risk measures adopted in our analysis. In particular, we use the negative skewness of the firm-specific weekly return (NCSKEWt) as the primary measure and the down-to-up volatility (DUVOLt) and the actual probability of stock price m i=1 13

15 crashes in a firm-year (CRASHt) as alternative measures of firm-specific stock price crash risk. CDS_Active t 1 is an indicator variable that is equal to one for any year after the inception of CDS trading, zero otherwise. We define the inception of CDS trading as the first time either CDS transactions or quotes are observed in our combined dataset. 5 Equation (4) is estimated using an ordinary least squares (OLS) approach for NCSKEWt and DUVOLt and a logit regression for CRASHt, due to its truncated distribution. The set of control variables includes DTURNt-1, SIGMAt-1, RETt-1, SIZEt-1, MBt-1, LEVt-1 ROAt- 1, and ACCMt-1. We use these variables to isolate the effect of well-documented determinants of stock price crash risk (e.g., Chen, Hong, and Stein 2001; Hutton et al. 2009; Kim et al. 2011a, 2011b). DTURNt-1 is the average monthly share turnover in year t-1 minus that for year t-2 and a proxy for differences of opinion among investors. We expect the coefficient for DTURNt-1 to be positive, because Chen et al. (2001) show that the existence of heterogeneous opinions is positively related to the probability of experiencing extremely negative stock returns in the future. SIGMAt-1 and RETt-1 are, respectively, the standard deviation and the arithmetic mean of firm-specific weekly returns in year t-1. According to the evidence documented in Chen et al. (2001), i.e., that stocks with higher past volatility or returns are more likely to experience price crashes in year t, we expect both SIGMAt-1 and RETt-1 to have positive coefficients. SIZEt-1 and MBt-1 are the logarithm of a firm s total assets and the market value of equity divided by the book value of equity in year t-1, respectively. As stocks with a large size and a high market-to-book ratio are more likely to experience future stock price crashes (e.g., Chen et al. 2001; Hutton et al. 2009; Kim et al. 2001), we expect positive coefficients for SIZEt-1 and MBt-1. LEVt-1 is the total long-term debt divided by 5 We treat the observed CDS quoted spread as an indicator of the inception of CDS trading because it reflects firmspecific information from the perspective of CDS dealers (sell side), such as banks and hedge funds, which are informed traders. The fluctuation of CDS quotes disseminates the information regardless of whether transactions occur. 14

16 total assets. Hutton et al. (2009), Kim et al. (2001), and Callen and Fang (2013) show that financial leverage is negatively related to stock price crash risk. We also control for profitability, which is reflected by ROAt-1, which is the ratio of income before extraordinary items divided by the total assets in year t-1. The effect of profitability in year t-1 on stock price crash risk is inconclusive. Specifically, Hutton et al. (2009) and Kim et al. (2011a, 2011b) find that the return on equity (ROE) and the return on assets (ROA) are negatively related to stock price crash risk, whereas Callen and Fang (2013) and Kim et al. (2014) find that ROE and ROA are positively related to crash risk, respectively. Moreover, because Hutton et al. (2009) find that the opacity of financial reports is significantly positively related to stock price crash risk, we control for it using ACCMt-1, the threeyear moving sum of the absolute discretionary accruals. In addition, as Chen et al. (2001) report that firms with a high stock price crash risk in year t-1 are likely to also have a high crash risk in year t, we control for the corresponding lagged stock price crash risk measures. [Please insert TABLE 3 about here] The multivariate regression results for Equation (4), using negative skewness under various scenarios, are reported in Table 3. We report t-values calculated using robust standard errors, corrected for firm clustering (Petersen, 2009). Our results using the full model (Model (3)) show a significantly negative coefficient for CDS_activet-1, indicating that CDS trading in year t-1 alleviates the stock crash risk in year t, supporting the information-revelation channel. To alleviate concerns about bias caused by fixed effects and to check the robustness of this finding, we examine the effect of CDS trading on stock price crash risk using restricted models that exclude firmspecific characteristics (Model (1)) and fixed effects (Model (2)). As shown in Table 3, we still find a significant and negative relationship between the indicator for CDS trading and the proxies for future stock price crash risk. After substituting the firm-fixed effects for industry-fixed effects 15

17 and a CDS_Firm dummy, we find that the negative relationship between CDS_Activet-1 and NCSKEWt is significant and robust, lending further credence to the argument that revealed information plays a key role. In addition, we divide our whole sample period into several subsample periods that correspond to the subprime financial crisis, which began at the end of and the CDS Big Bang protocols in 2009; 7 we find consistent results using multivariate regressions. 8 Our results are also robust to removing firms with zero leverage. [Please insert TABLE 4 about Here] We perform the same multivariate analysis using DUVOLt-1 and CRASHt-1 as alternative stock price crash risk measures and report the results in Table 4. As CRASHt-1 is an indication of the occurrence of actual stock price crashes in a firm-year, its distribution is truncated. Thus, we conduct Tobit regressions for CRASHt-1. The results are consistent with the evidence we document using NCSKEWt-1; we find a relationship between CDS trading in year t-1 and the stock price crash risk in year t, reinforcing the evidence for a negative relationship between the inception of CDS trading and stock price crash risk. [Please insert TABLE 5 about here] Additionally, the liquidity of financial instruments, including adverse selection, search frictions, and inventory costs, directly affects information revelation in financial markets. For instance, illiquidity increases the difficulty of capitalizing information into current prices because of high search frictions and inventory costs. In contrast to the equity market, the CDS market is relatively 6 The subprime financial crisis lasted from 2007 to 2010 (see, e.g., essays/subprime_mortgage_crisis). We define the pre-crisis period as and the post-crisis period as The International Swap and Derivative Association (ISDA) implemented a series of protocols in April 2009 to standardize CDS contracts, which had a significant effect on the CDS market. 8 The results of these multivariate regressions are reported in Table A3 of the Internet Appendix. 16

18 illiquid. Tang and Yan (2007) report that liquidity is priced in the CDS market. Therefore, if the information-revelation argument is true, we expect the negative effect of CDS trading on future stock price crashes to be more pronounced for firms with liquid CDS contracts. To test this conjecture, we use the number of distinct dealers that provide quotes for a firm in a fiscal year as a proxy for CDSs endogenous liquidity (Qiu and Yu 2012), 9 and the total number of actual CDS trades on a single-name CDS contract in a fiscal year as a proxy for actual liquidity. Specifically, to mitigate the effect of extreme values, we define CDS_Quotes as the logarithm of one plus the number of distinct dealers and CDS_Trades as the logarithm of one plus the aggregated number of actual trades. As reported in Table 5, we find significant and negative coefficients for CDS_Quotes and CDS_Trades. This finding supports our conjecture and lends further strength to the argument for CDS trading having a negative effect on future firm-specific stock price crash risk. To evaluate the economic significance of the effect of CDS trading on stock price crash risk, we estimate the marginally expected decrease in the probability of a crash as a function of the occurrence of CDS trading, with all the other variables at their sample mean. Numerically, the marginal effect is 10.1 percent for NCSKEWt-1 and 5.9 percent for DUVOLt-1, as reported in the unrestricted models in Tables 3 and 4. In contrast to the means of the unconditional NCSKEWt-1 and DUVOLt-1, which are about 1.4 percent and 1.1 percent, respectively, the marginal effect of CDS trading on the likelihood of an increase in the stock price crash risk is economically 9 In their dataset, Markit reports the number of distinct dealers that provide daily quotes for a firm. We use the average number of distinct dealers for a firm in a fiscal year as a proxy for the liquidity of CDSs in a fiscal year. 17

19 significant. Furthermore, we find that CDS trading in the current year reduces the percentage of firm-years that experience actual stock price crashes in the following year by about 12 percent. 10 The coefficients for the control variables, as expected, are generally consistent with previous studies. Specifically, we find significantly positive coefficients for DTURN, SIGMA, RET, SIZE, and MB in year t-1, consistent with prior research (e.g., Chen et al. 2001; Hutton et al. 2009; Kim et al. 2011a, 2011b). We also find a positive relationship between the profitability measure, ROA, and crash risk, consistent with Callen and Fang (2013) and Kim et al. (2014). Moreover, consistent with Hutton et al. (2009), we find a positive coefficient for the opacity of financial reports, measured by ACCMt-1, suggesting that firms with high financial report opacity experience a severe stock price crash risk. To summarize, the results in Table 3 provide strong evidence in support of the informationrevelation argument, i.e., that CDS trading alleviates future firm-specific crash risk, after controlling for the opacity of financial reports (Hutton et al. 2009), investor heterogeneity (Chen et al. 2001), and other known determinants of crash risk. This result is robust to the use of two alternative proxies for stock price crash and to the restricted models that remove fixed effects and firm-specific characteristics. Endogeneity Analysis The previous section demonstrates a significant statistical relationship between the inception of CDS activities and a decrease in the likelihood of a firm experiencing extremely negative stock returns one year later. However, to establish the causality of CDS trading on the decrease in the stock price crash risk, we must consider the possibility of reverse causality. In this case, one would 10 As reported in Model (4) in Table 5, the coefficient of CDS_Active is Because the mean of CRASH is 0.206, the percentage change in CRASH after the introduction of CDS trading equals / percent. 18

20 predict that CDS trading might be initiated in firms for which the expected stock price crash risk is high and that omitted variables could drive both the likelihood of initiating CDS contracts and future stock price crashes. To address these concerns, we adopt a difference-in-difference analysis, placebo tests, and an instrumental variables approach to control for potential endogeneity and examine the robustness of our results. Propensity Score Matching The propensity score matching (PSM) approach, first proposed by Rosenbaum and Rubin (1983), is a statistical matching technique that accounts for the covariates that predict the likelihood of receiving the treatment. It reduces the bias caused by the possibility that the difference between the two groups may depend on the features that affect whether a member received treatment rather than the effect of the treatment per se. One of PSM s implicit assumptions is that the potential outcomes are independent of the treatment assignment, conditional on the propensity score, which is strong and untestable. Nonetheless, according to a survey conducted by Roberts and Whited (2012), the PSM approach is one of the most frequently used techniques for addressing endogeneity concerns. Therefore, we use PSM to control for the possibility that an increase in the likelihood of initiating CDS trading is caused by expected future credit risk. We match the treated and control groups based on a five-year event window. Specifically, we define a treated firm as one that initiates CDS trading in the third year (year t) of a five-year window. In other words, there are no CDS activities in the first and second years (year t-2 and t-1) and after the inception of CDS trading in the third year (year t), CDS activities are continued in the fourth and fifth years (year t+1 and t+2). We then select a corresponding control firm from the group of no-cds firms that do not engage in CDS trading during that five-year window. The 19

21 control firm is matched to a treated firm in year t in the five-year window if it is in the same twodigit SIC industry and if it has the closest propensity score for the initiation of CDS trading in year t. We compute the propensity score for the inception of CDS trading in year t using the following multivariate regression: m CDS Active t = α 0 + γ i (i th Determinants t 1 ) + ε t. (5) i=1 where CDS Active t is an indicator that equals one if there are CDS activities in year t and zero otherwise. Following Subrahmanyam et al. (2014), we consider several firm characteristics as determinants of the inception of CDS trading, including the logarithm of firm size (SIZE), financial leverage (LEV), the return on total assets (ROA), the market-to-book ratio (MB), the ratio of total sales over total assets (SALE), the ratio of earnings before interest and tax over total assets (EBIT), the ratio of working capital over total assets (WCAP), the ratio of retained earnings over total assets (RE ), the ratio of capital expenditure over total assets (CAPX ), and the ratio of tangible assets over total assets (PPE) in year t-1. We also incorporate equity market characteristics, including the means and standard deviations of the weekly returns and the de-trended average monthly stock turnover (DTURN) for the previous year. Furthermore, as Cao, Jin, Pearson, and Tang (2016) find that firms with equity options are more likely to initiate CDS trading, we also include an option trading indicator term (Option_Firm) as a determinant of the initiation of CDS trading, that equals one for firms with option trading in our sample period, zero otherwise. Subrahmanyam et al. (2014) and Shan, Tang, and Yan (2016) show that a lending relationship is positively related to the likelihood of trading CDSs. To identify lending relationships, we merge our CDS dataset with Mergent FISD and Thomas Reuter s Dealscan to identify bond and syndicated loan information, 20

22 respectively. To indicate the availability of lenders, we use a lender dummy (Lender_dummy), that equals one if there are lenders for a firm in a year and zero otherwise. In addition, we also control for year- and industry-fixed effects in our regression Equation (5). The probit regression results are reported in Table A4 of the Internet Appendix. By using this matching procedure, we construct a control group that is similar to the treated group in terms of industry and the likelihood of initiating CDS trading in year t, but CDS trading only occurs in the treated group. Thus, the change in stock price crash risk in the control group is equivalent to the change in crash risk that would have occurred during the event window had the treated group not received the treatment. Consequently, the difference between the change in the stock price crash risk in year t+1 of the five-year window for the treated group and that for the control group reflects the causal effect of CDS trading on stock price crash risk. [Please insert FIGURE 1 about here] First, in Panel A of Figure 1, we plot the means of the NCSKEW values for treated firms around the inception of CDS trading. At one year (year 1) after the inception of CDS trading, we observe a decrease in the average stock price crash risk and find that the stock price crash risk continues to decline in the following year. When we compare the change in the stock price crash risk for the treated firms with that for the control firms (the difference in the NCSKEW values between these two groups, as exhibited in Panel B of Figure 1), a decline in the stock price crash risk is observed after the inception of CDS trading. [Please insert TABLE 6 about here] Next, we use a multivariate regression model to rigorously analyze the relationship between the stock price crash risk and CDS trading. We define Treatedt-1 as an indicator that equals one for the 21

23 observations in the treated group and zero for those in the control group. We introduce Treatedt-1 and the interaction of Treatedt-1 and CDS_Activet-1 into regression Equation (4) to conduct a multivariate analysis of the treated and control firms. The results of Model (1), given in Table 6, show a significantly negative coefficient for the interaction term. This suggests that the stock price crash risk decreases in the treated group compared to the control group, which further supports the argument that information revelation plays a key role. We perform a similar difference-indifference analysis using the alternative stock price crash risk measures and find consistent results. 11 Placebo Tests It is possible that the negative relationship between the CDS trading indicator and the stock price crash risk measures is driven by some unobserved factors that drive both actions simultaneously, but in opposite directions. If this is true, then we expect to observe the comovement of CDS trading and stock price crash risk, regardless of the exact timing of the initiation of the CDS trading. We adopt a placebo test to address this concern. We incorrectly assign the timing of the treatment (initiation of CDS trading) to one or two years before the actual event. If the negative relationship between CDS trading and stock price crash risk is driven by a predetermined trend or unobserved variable, we would expect to observe a similar effect when using the wrong date for CDS trading. However, if the decrease in the stock price crash risk is driven by the inception of CDS trading, this negative relationship should disappear when we assign the wrong date to CDS trading. The second and third columns in Table 6 report the regression results for Models (2) and 11 The detailed results are reported in Table A5 in the Internet Appendix. 22

24 (3), which incorrectly assign the CDS trading date to one and two years before the actual events, respectively. We find that the coefficients for the interaction term are both negative but not significant at the conventional level, which alleviates the concern that unobserved variables are responsible for our results and lends credence to the negative effect of CDS trading on the future firm-specific stock price crash risk. In addition, when we use DUVOL as an alternative measure of stock price crash risk in the placebo tests, as shown in Table A5 of the Internet Appendix, the results are consistent with those given in Table 6. Instrumental Variable Analysis Although the placebo test described in the previous section rules out the concern that a predetermined trend or unobserved variables are driving our results, there is still potentially a reverse effect of the expected stock price crash risk in year t+1 on the initiation of CDS trading in year t. Specifically, investors who are expecting a high stock price crash risk in the future may proactively hedge their position. As CDS contracts are designed to transfer credit risk between two counterparties, they are an ideal financial instrument for hedging. Thus, there is the possibility that the expectation of a high stock price crash risk in the future increases the likelihood of CDS trading. This would also result in a negative relationship between stock price crash risk and current CDS trading. In this section, we address this potential endogeneity concern using an instrumental variable regression approach. We chose instrumental variables that directly affect the initiation of CDS trading but not the stock price crash risk in the equity market. These instrumental variables could only affect the stock price crash risk through the channel of CDS trading. Specifically, we use the Lender_FX_Hedge and Borrower_CDS_Ratio as instrumental variables to perform a two-stage regression. Lender_FX_Hedge is the amount of foreign exchange (FX) derivatives used by a firm s lender for 23

25 hedging purposes, scaled by the lender s total assets; it reflects the hedging preference of the lenders. Minton, Stulz, and Williamson (2009) find that banks that use interest rates, foreign exchange, equity, and commodity derivatives are more likely to use CDS contracts, indicating that their hedging preferences extend across a variety of financial markets, including the credit derivative market. Therefore, we conjecture that the amount of FX derivatives used for hedging should be positively related to the likelihood of initiating CDS trading. Furthermore, as the hedging preference of a firm s lender is unlikely to directly affect firm-specific stock price crash risk in the equity market, Lender_FX_Hedge also satisfies the exclusion condition that there is no direct effect of Lender_FX_Hedge on stock price crash risk. To calculate Lender_FX_Hedge, we first link our dataset with the Mergent Fixed Income Securities Database (FISD) and Thomas Reuter s Dealscan to extract the issuance information about bonds and syndicated loans. We extract all the bond underwriters and leading lenders of syndicated loans over the previous five years. To calculate the ratio of the amount of FX derivatives over total assets, we collect data on the use of FX derivatives and the fundamental lenders data from the Call Reports at the Federal Deposit Insurance Corporation (FDIC). As there could be more than one lender for a firm in a year, we define the maximum ratio of the FX derivatives for all of an individual firm s lenders as Lender_FX_Hedge. 12 Our second instrumental variable, Borrower_CDS_Ratio, reflects the lender s preference for the use of CDS from the perspective of borrowers CDS-referenced status. A bank can lend to many firms simultaneously. If a bank prefers or is willing to use a CDS contract to hedge its credit risk position, we expect to observe more CDS-referenced borrowers on its list. Thus, if a high 12 We use the maximum ratio of FX derivatives among all the lenders, but not the means, because a lender s strong hedging preference is enough to motivate the inception of CDS trading, while the mean of the ratio of FX derivatives reflects the average willingness to participate in CDS trading. 24

26 percentage of a bank s borrowers are CDS-referenced borrowers, excluding the firm under examination, then there is a high probability of CDS trading in the future. Similar to the argument made for the hedging preference reflected by Lender_FX_Hedge, we argue that the ratio of CDSreferenced borrowers to total borrowers for a lender is less likely to directly affect the stock price crash risk in the equity market, satisfying the exclusion condition. First, we use the lenderborrower relationship from bonds and loans extracted from FISD and Dealscan to identify all of a given lender s borrowers in the past five years. Then we examine the status of CDS trading, including both CDS transactions and CDS quotes for each borrower in the current year, and after removing the firm of interest, we calculate the percentage of borrowers that are CDS referenced for each lender in every year. Because a firm could have multiple lenders, we use the maximum ratio of CDS-referenced borrowers as Borrower_CDS_Ratio. 13 [Please insert TABLE 7 about here] We use the two instrumental variables described above to conduct a two-stage regression. In the first stage, we conduct a probit regression to examine the effect of the introduction of CDS trading using the regression model given in Equation (5). Relative to the CDS prediction model based on the propensity score, we incorporate more detailed lenders characteristics into the firststage regression, including the lenders maximum size (Lender_Sizet-1) and the total number of lenders (Lender_Numt-1). As Shan et al. (2014b) document a non-linear relationship between a bank s size and the probability of CDS trading, we also control for the square of the maximum size of the lender Lender_Sizet-1 2. Table A6 of the Internet Appendix reports the results of the firststage probit regressions. As expected, we find significant and positive coefficients for both 13 We use the maximum ratio of CDS-referenced borrowers to all borrowers, rather than the mean, because a lender s strong hedging preference is enough to motivate the inception of CDS trading, whereas the mean of the ratio of FX derivatives reflects the average willingness to participate in CDS trading. 25