Noise Trading in the Credit Default Swap Market Felix Irresberger∗ TU Dortmund University Gregor N.F. Wei߆ University of Leipzig February 1, 2016 ABSTRACT We show that noise trading affects the cross-section and time-series of credit default swap (CDS) spreads. We employ internet search volume data to proxy for an exogenous shock to the pessimistic sentiment of households about credit risk at the market level. Consistent with the conjecture of significant noise trading existing in the CDS market, we find our proxy of household sentiment to be a significant driver of CDS spreads. Our results are in line with theories that emphasize the role noise trading plays in providing market liquidity with higher levels of sentiment being associated with lower average CDS spreads. Keywords: Credit default swaps, noise trading, household sentiment, market liquidity. JEL Classification Numbers: G21, C58, G01. ∗ Otto-Hahn-Str. 6, D-44227 Dortmund, Germany, telephone: +49 231 755 8212, e-mail: [email protected]. † Corresponding author: Grimmaische Str. 12, D-04107 Leipzig, Germany, telephone: +49 341 97 33820, e-mail: [email protected]. 1 Introduction Economic intuition suggests that noise trading should have a weaker impact (if any) on the prices of derivatives than on the prices of the underlying securities. As institutional (and thus presumably better informed) investors dominate derivatives markets, prices should not be influenced as strongly by investors’ attention and sentiment as it is well documented for stocks (see, e.g., Baker and Wurgler, 2006, 2007; Tetlock, 2007; Da et al., 2011, 2015). This view has been challenged by empirical evidence by Han (2008) who finds that investor sentiment about the stock markets significantly affects the prices of stock options. Perhaps even more surprising is the finding of Hilscher et al. (2015) that informed traders are primarily active in the equity market rather than the market for credit default swaps (CDS) with stock returns leading returns on CDS contracts. Despite these initial results, however, no study has so far analyzed the role of noise trading for the pricing of credit default swaps. In this paper, we study how noise trading affects the cross-section and time-series of CDS spreads. We construct a proxy for households’ pessimism about overall credit risk in the economy (“default sentiment”) and show that it is significantly drives CDS premia. Surprisingly, default sentiment is negatively correlated with CDS spreads. Our proxy for default sentiment does not only explain a significant portion of the cross-sectional and time series variation in CDS premia, but it also complements rather than substitutes competing macroeconomic covariates. We then test and confirm the hypothesis that sentiment shocks influence CDS spreads via the market liquidity channel where uninformed traders overreact to pessimistic sentiment, thereby boosting market liquidity and lowering CDS spreads. We use weekly internet search volumes for queries on Google that are related to pessimism concerning default risk to build our proxy for household sentiment at the market-level. A key advantage of using search volume data from Google is that it provides us with the opportunity to construct a proxy of household pessimism that should be exogenous to institutional investors in the CDS market. Although it is quite possible that traders in the CDS market also use Google searches to gain information on expected default probabilities of individual firms, these biasing 1 influences should be smoothed out by the aggregated search volumes of millions of Google users. We then regress the CDS spreads of 228 firms over a period of 351 weeks from 2004 to Q3:2010 on our sentiment proxy and various control variables and find the following key results: First, higher sentiment lowers average CDS spreads even when controlling for various proxies of average credit risk in the economy. Second, the correlation between default sentiment and CDS spreads in our regressions is highly sensitive to the inclusion of proxies for CDS market liquidity. We test two different channels through which sentiment might affect CDS spreads. The first hypothesis that we test then postulates that shocks in household pessimism reflect actual increases in the default risk of firms in the economy. In other words, the pessimistic sentiment of households turns out to be a true information signal to which both informed and uninformed traders react. There are (at least) two reasons for why household sentiment could reflect actual overall credit risk. One reason could be that households are able to observe pervasive increases in default risk in the economy, a fact, which is then revealed by their concerns expressed via their Google searches. Alternatively, it could simply be that both household sentiment and overall credit risk are jointly driven by macroeconomic covariates. In both cases, household sentiment (reflecting an increase in default risk) should unambiguously be positively related to CDS spreads as both informed and uninformed traders buy protection against genuinely higher credit risk. The second hypothesis, the one which we empirically confirm in our study, is concerned with the increasing effect of noise trading on market liquidity. In case the concerns of households regarding overall credit risk are not justified by fundamentals, uninformed traders could still (over) react to the pessimistic signal while informed traders take short positions. In contrast to the first scenario described above, the effect of sentiment on CDS spreads then becomes ambiguous. As shown by Baker and Stein (2004) in their model, the overreaction of uninformed traders can have a positive effect on market liquidity in such a situation. Consequently, the increasing effect of sentiment on CDS spreads might be offset by the side-effect caused by an increased liquidity in the CDS market (and thus a higher liquidity of the individual CDS contract). All our empirical findings support this second explanation. Our proxy for default sentiment 2 significantly drives changes in CDS spreads in both time series and panel regressions. Increases in default sentiment are associated with decreases in CDS spreads. This finding is robust to the inclusion of various idiosyncratic and macroeconomic variables with one important exception: the negative correlation between our sentiment proxy and CDS spreads immediately becomes insignificant as soon as we include a proxy for CDS market liquidity. Analyzing this result in more detail, we find our proxy of default sentiment to be strongly negatively correlated with the bid-ask spread of a CDS market index. Moreover, default sentiment explains a significant portion of the variation in individual CDS contracts’ liquidity, especially in times of higher limits to arbitrage as observed during the financial crisis. Finally, we find this liquidity-induced negative relation between sentiment and CDS spreads to be limited to sentiment shocks that are closely related to default risk as proxies of households’ sentiment towards the general economic outlook are expectedly positively correlated with CDS spreads. The negative relation between default sentiment and CDS spreads, however, remains unchanged in its significance. The results that we find in this study are important for several reasons. First, analyzing the impact of noise trading on CDS spreads helps explain the puzzle why observable firm-specific and macroeconomic covariates fail to explain the variation in CDS spreads despite their usefulness for predicting default.1 Second, from a risk management perspective, understanding the role sentiment plays in the pricing of credit risky securities is important for trading and hedging CDS contracts, especially in times of higher limits to arbitrage. Next, the results we document are also of significant value to regulators for understanding the mechanics driving the systemic importance of the CDS market. Finally, and perhaps most importantly, our results add a wholly new argument to the current discussion of the unclear role liquidity and liquidity risk play in the pricing of CDS contracts (see Tang and Yan, 2014; Bongaerts et al., 2011; Junge and Trolle, 2015). Our study is related to several studies in the literature on asset pricing that have addressed the question whether investor attention and investor sentiment are priced in the cross-section of stock returns. While traditional models of asset pricing assume an instantaneous incorporation of news 1 See, e.g., the studies of Shumway (2001), Duffie et al. (2007), Chava et al. (2011), and Galil et al. (2014). As a solution for this puzzle, Doshi et al. (2013) propose a discrete-time no-arbitrage model with observable covariates. 3 into asset prices, in reality, asset prices could be influenced by limits on attention due to scarce cognitive resources of investors (see, e.g., Kahneman, 1973; Sims, 2003; Peng and Xiong, 2006). Quite similarly, asset prices could also be influenced by investor sentiment as irrational noise traders and limits on arbitrage will lead to mispricings and excess volatility (see De Long et al., 1990). Although earlier studies do not find attention to be generally priced in stock returns (see, e.g., Cutler et al., 1989; Kogan et al., 2006, 2009), there now seems to be a consensus in the literature that increases in both investor attention and investor sentiment can lead to large movements in asset prices at least in the short-run (see, e.g., Baker and Wurgler, 2006; Tetlock, 2007; Da et al., 2011, 2015). In contrast, empirical evidence on noise trading in derivatives markets is much less clear than for equity markets. In one of the first studies on this subject, Figlewski (1989) finds standard arbitrage in the market for stock index options to be heavily exposed to risk and transaction costs. In a related study, Figlewski and Green (1999) point out in their empirical study that imperfect pricing models can induce considerable risk exposure to option writers. More recently, Ofek et al. (2004) find empirical evidence of limited arbitrage in options market. However, explicit tests for the presence of investor sentiment in options markets are scarce with the work of Han (2008) being a recent exception. All of these studies, however, have so far concentrated on stock options neglecting the possibility that corporate default and the prices of credit risky securitites are as well sensitive to changes in investor or creditor sentiment. Our study is also related to a large body of literature that tries to explain the variation of credit spreads and credit default swap premia (and their changes) in the cross-section and time series. Although early studies in this strand of literature have questioned the explanatory power of observable covariates for credit spreads,2 recent studies by Ericsson et al. (2009), Doshi et al. (2013), and Meine et al. (2015b) show that macroeconomic and firm-specific information can explain a significant portion of the variation in CDS spreads over time and across firms. To the best our knowledge, however, only the study by Hilscher et al. (2015) has investigated the role of noise 2 See, e.g., Collin-Dufresne et al. (2001); Campbell and Taksler (2003); Cremers et al. (2008). 4 trading in CDS markets so far. We complement their work by building on the rich asset pricing literature on investor sentiment and identify market-level sentiment as a previously neglected covariate in regressions of CDS premia. The rest of this paper is organized as follows. In Section 2, we quickly review the literature and build our two main hypotheses on the expected effect of default sentiment on CDS spreads. Section 3 presents our empirical strategy. The construction of our sample and our empirical results are discussed in Section 4. Section 5 concludes. 2 Theoretical background and hypothesis building In this section, we first derive and discuss the two main hypotheses that we test in out study. Then, we shortly comment on the usefulness of internet search volume data for measuring sentiment. 2.1 Default sentiment and the pricing of credit derivatives In theory, market-level sentiment could have both a positive and negative effect on prices of credit derivatives. On the one hand, market-level sentiment is likely to have a significant impact on credit risk as it reflects consumer and creditor behaviour.3 More precisely, search volumes at the market-level for terms like “credit risk”, “default”, or “bankruptcy” reveal concerns of investors and creditors about losses from their investments and loans. These concerns could be caused by earlier spikes in default probabilities making our proxy for market-level sentiment an indicator of contemporaneous market-wide increases in credit risk. Alternatively, creditors could search for credit-related terms simply out of fear of future losses. However, increases in market-level sentiment should in both cases coincide with a decline in private consumption and investments and thus ultimately a worsening of the overall state of the economy. 3 In this respect, our Google-based proxy for market-level sentiment is similar to, e.g., the University of Michigan Consumer Sentiment Index. 5 In fact, there exist at least two channels through which the sentiment of households at the market-level could influence default risk. First, higher market-level sentiment and worsening macroeconomic conditions should influence corporate default via average recovery rates in the economy. As shown in the model of Shleifer and Vishny (1992) and confirmed empirically by Acharya et al. (2007), firms that default amidst a time of poor macroeconomic conditions have lower asset values at liquidation and thus lower recovery rates (see, e.g., Duffie and Singleton, 1997, 1999). Market-level sentiment, on the other hand, should capture declines in the overall state of the economy better than measures that are only available at low frequencies (e.g., GDP growth rates) or indirect proxies (e.g., option implied volatilities in the form of the VIX, or mutual fund flows). Second, corporate default itself is closely associated with macroneconomic conditions. Several studies have empirically tested and confirmed that default probabilities are in large part determined by macroeconomic covariates like, e.g., the quarterly growth in industrial production, interest rate levels, or stock market indexes (see, e.g., Blume and Keim, 1991; McDonald and Van de Gucht, 1999; Hillegeist et al., 2004; Pesaran et al., 2006).4 Alternatively, just like equity markets (see Shleifer and Vishny, 1997; Baker and Wurgler, 2006, 2007), credit derivative markets could be influenced by limits to investor attention and limits to arbitrage. As CDS markets are dominated by institutional investors, CDS prices should suffer to a lesser extent from mispricings caused by noise traders and limits to arbitrage. However, limits to arbitrage could just as well exist in credit derivative markets with sentiment driving (at least in part) the relative demand for speculative investments in selected CDS contracts. This reasoning is in line with a large number of studies on investor behavioral biases in options markets. 5. Finally, CDS premia could also depend directly on macroeconomic covariates as shown in the study of Doshi et al. (2013). They propose a discrete-time no-arbitrage model which allows for a closed-form solution for CDS prices based on observable macroeconomic and firm-specific co4 In the study by Duffie et al. (2007), only the three-month Treasury bill rate and the trailing one-year return on the S&P 500 index are identified as significant covariates for explaining estimated default. Recent studies on CDS pricing have adopted this set of macroeconomic covariates (see Ericsson et al., 2009; Doshi et al., 2013). 5 See, for example, Figlewski (1989), Stein (1989), Figlewski and Green (1999), Poteshman (2001), Poteshman and Serbin (2003), Bollen and E. (2004), Mahani and Poteshman (2008) and Ofek et al. (2004) 6 variates. Although they test the significance of various macroeconomic indices (e.g., the consumer price index (CPI), producer price index (PPI), real GDP, etc.) in their empirical study, the potential impact of a direct measure of investor and creditor sentiment on CDS premia remains unexplored. Consistent with what we refer to as the default risk hypothesis, sentiment could thus be positively related to CDS premia either because of the correlation of our sentiment measure with macroeconomic covariates and recovery rates, the price impact of limits to arbitrage and noise traders, or because sentiment genuinely reveals increases in overall default risk: H1 : Increases in default sentiment lead to higher CDS spreads (default risk hypothesis). On the other hand, default sentiment could lower CDS spreads by increasing CDS market liquidity to such an extent that the price impact of noise trading is off-set. Let us assume that both informed and uninformed investors trade in the CDS market, and that inattentive traders (e.g., insurers, pension funds, but also banks) are short-sales-constrained. As shown by the model of Baker and Stein (2004), sentiment could cause market liquidity to increase causing an ambiguous impact on asset prices. In the context of the CDS market, their reasoning reads as follows. Uninformed traders could overreact to unjustified, pessimistic signals about overall default risk in the economy and be tempted to buy credit protection. Informed traders would take short positions. When informed investors trade on a subsequent date, irrational investors will also believe their information to be better than that of the informed traders and thus will the market more liquid even in the face of such informed trading. In case the inattentive traders receive positive signals about default risk, however, their short-sales-constraint will keep them out of the market. As a consequence, the overreactions of uninformed traders could have a positive effect on market liquidity and default sentiment could lower CDS spreads to some extent as it increases overall market liquidity (and thus the liquidity of CDS contract’s with higher liquidity risk). Considering the importance of liquidity for CDS pricing (see, e.g., Tang and Yan, 2014), it is safe to assume that such a liquidity effect would more offset the positive price impact of noise trading. For this reasoning to be true, of course, both of the above made assumptions should be valid. However, both assumptions appear to be reasonable in the context of the CDS market. 7 Hilscher et al. (2015) confirm in their empirical study that the proportion of inattentive investors in the CDS markets in fact seems to be larger than in the equity market. Moreover, especially inattentive traders that only enter the CDS market to hedge their credit exposure are likely to be short-sales-constrained (either voluntarily or due to regulatory rules). This conjecture is also underlined by Mengle (2007) and Bongaerts et al. (2011) who note that long-term investors usually sell credit protection while commercial banks only appear to use CDS contracts for hedging. In summary, we formulate the following hypothesis that explains the correlation between default sentiment and CDS spreads via the market liquidity channel: H2 : Increases in default sentiment lead to higher CDS market liquidity and thus lower CDS spreads (market liquidity hypothesis). We test both hypotheses by using proxies of default sentiment based on internet search volume data. 2.2 Measuring sentiment Measuring investor sentiment in financial markets has been a relatively challenging mission to accomplish so far. Many studies propose the use of aggregated informationen and opinions on the market from written publication such as newsletters, column articles, or financial forums’ message boards (see, e.g., Tetlock, 2007; Tetlock et al., 2008; Antweiler and Frank, 2004) and define implicit measures that reflect the general market sentiment. 6 Other approaches include surveying, analyzing trading activity, extreme returns, and measuring valuation errors where prices deviate from rational pricing models (see, e.g., Barber and Odean, 2008; Han, 2008). In this paper, we make use of data on internet search volume in the United States to study the impact of household sentiment on prices in the credit default swap market. More precise, we employ information on the time evolution of specific search queries by internet users related to credit risk and the default 6 The effect of news coverage is an important factor for financial markets. For example Dang et al. (forthcoming) use a huge database on firm news around the world to explain co-movements in stock returns and stock liquidity. 8 of corporations. Also, in our robustness checks, we control for the effect of individual investor attention to underlying equities and market-wide household sentiment on CDS spreads. The use of search volume data yields several advantages over other common methods to derive measures of investor sentiment. For example, data on internet search volume is available on a daily and weekly basis and is easily accessible. This would not be viable with, e.g., survey data, which is hard to obtain and not scalable and thus, is limited to a fraction of the population. 7 Also, search volume data is especially insightful for the population in the U.S. because of its high percentage of active internet users.8 In addition to this, the relative search volume on specific phrases is an indicator of interest for a broader audience and represents the active information retrieval of households. Measures based on written publications (that are, e.g., derived by using linguistic techniques (see Tetlock et al., 2008) reflect rather passive information procurement of investors. While such proxies for investor sentiment may be able to capture market sentiment, as perceived by the authors of the texts, it is flawed in that way that there is no information on the actual interest and attention of investors to such outlets. With search volume data, however, we are able to infer the importance of specific topics to households and investors. The usefulness of such data to detect trends and forecast economic indicators has been well documented in the recent literature. One of the first successful applications of internet search volume data from the Google Trends analytics tool can be found in Ginsberg et al. (2009) who predict influenza epidemics. Choi and Varian (2011) also provide possible applications when analyzing internet search volume data. In the finance literature, the most prominent and most related studies are given in Da et al. (2011, 2015). 9 Google Trends lets its user type in search phrases and returns information on the time evolution of search phrases for a given time period, relative to the number 7 Also note that survey data may contain biased information, which is not the case with anonymous search behavior of internet users (see Singer, 2002). 8 As of 2013, over 82% of the population in the United States use the internet for informational use (see http://data.worldbank.org/indicator/IT.NET.USER.P2). From that percentage of the population, around 72% use Google’s search engine as their primary source of information (see Hitwise, 2009). 9 For other examples see, e.g., Kristoufek (2013); Preis et al. (2013); Hamid and Heiden (2014), or Ding and Hou (2013), who illustrate the usefulness of Google Trends when studying portfolio selection, volatility, Value at Risk forecasting, and stock liquidity. 9 of all searches with Google’s search engine. 10 Further, we are able to restrict the output to cer- tain geographic areas, in our case the United States. In our empirical study, we employ the Google Trends analytics tool to proxy for the sentiment of households concerning increases in market-wide average default risk. 3 Empirical strategy In this section, we introduce our main variables of interest and give some motivation on their relation to credit default swap spreads. Specifically, we construct a measure of default sentiment and related variables based on internet search volume data obtained from Google Trends. Afterwards, we define and explain known determinants of CDS spreads from the previous literature. Finally, we present our econometric design to test our main hypotheses. 3.1 Default sentiment index We intend to capture the specific sentiment of households to the credit and lending business and the default of corporations. When measuring the attention to a narrow topic with internet search volume data the question arises which set of search phrases is most appropriate. To accomplish this task, we build a fundamental set of search phrases and bi-grams that reflect the circumstance of a firm or credit to default. This approach yields more focused information on internet users’ motivation to search for specific themes rather than general phrases like “recession”, which are hard to interpret. In order to decide for the most suitable search terms, and to find additional ones, we use Google Trends and its “related search terms” feature to see which phrases are actually searched for and to what extent. Our goal is to create a market-level index that reflects the sentiment of households and creditors towards failures or defaults of corporations and the credit business. We do this by combining similar expressions for business entities and creditor activities with phrases describing default, failure, or bankruptcy. As our basis, we follow Da et al. (2015) and make 10 The resulting time-series is then scaled to a maximum of 100. 10 use of the Harvard dictionary and filter all words that have the tags “Econ@” or “ECON” (743 words). First, we screen our initial list for synonyms of the phrase “firm”, such as “business”, “company”, or “corporation”. Second, we select from the list all terms associated with the credit and lending business. Finally, we manually search the list for unambiguous negative phrases that describe the failure of corporation or financial products. However, we only have lists of single phrases (1-grams) but would also like to receive information on 2-grams or 3-grams. Information on only “bankruptcy” is too broad and needs to be associated with a meaning, in this case “business bankruptcy”. Therefore, we perform any combination of the words on each list, such as “corporate bond” or “corporate bond default”. We add to our list of words every combination that is sensible (e.g., not “broker accrue” or “accrue dealer”). Combinations of the synonyms for firm or financial products are typed into Google Trends and compared using the following system: For each default word, we compare its respective combination with a firm word, e.g., “business bankruptcy and “firm bankruptcy” via Google Trends to gain insights on which word is actually used. This makes sense since “business default”, and “company default”, etc. express the same circumstance of a default only with a synonym for the business entities. The 2-gram with the highest average search volume between 01/2004 and 09/2010 of a certain group of words is then chosen. If the search volume of a 2-gram has lower average search volume than the top phrase but exceeds the search volume of the most searched phrase at least once over the time period, it is also included in future considerations. Doing so, we can guarantee that the search volume of a word is relevant enough. If none of the search terms has sufficient search volume to be shown in the Google Trends tool, the 2-gram is omitted. Similarly, we proceed with the combinations of our credit words and default words. This approach could suffer from a simple issue: One could expect that people search for “personal bankruptcy” and so on that do not arise from the methodology used. Therefore, we look for related search terms, which also include 3grams and more, to mitigate the concern that we omitted relevant combinations. We perform this comparison for the business and default words and the credit and default words and arrive at our initial list of search terms. 11 Then, we type into Google Trends every search term and collect every related search terms. We then proceed by removing all duplicates from our set of search phrases. After that, we download all search volume data and exclude from our list all terms with insufficient search volume data (e.g., no or only monthly data). Also, we exclude every related search terms that are clearly not expressing the expectation of a business or credit-like default, e.g., “motors liquidation company”. In total, we end up with a list of thirty search phrases that have weekly search volume for the time period from 01/2004 to 10/2010. Following Da et al. (2015), we winsorize each weekly time series at the 5% level. To mitigate concerns about seasonality and heteroskedasticity, we regress the individual time series on month dummies, take the residual, and then divide by the standard deviation of the time series. Our final Default Sentiment Index is the average of all thirty time-series. 3.1.1 Retail investor attention To complement our proxy of default sentiment, we also employ measures of investor attention in our analysis. In their pioneering study, Da et al. (2011) find that retail investor attention predicts higher returns and over time leads to price reversals in equity prices. Since it is possible that there are spillover effects from the equity market to the credit markets (see Hilscher et al., 2015), we cannot rule out that individual investor attention on specific firms or the presence of noise traders in the equity market does not affect CDS spreads. Therefore, we control for equity attention by employing the search volume index (SVI) for stock ticker symbols from Google Trends. 11 We download all SVI data on stock ticker symbols in our sample for the period from 01/2004 to 09/2010 and restrict the output to the United States. In cases where the ticker symbol may have ambiguous meaning (e.g., “CAT” or “GPS”) and thus, a unusually high search volume, we exclude the ticker from our sample. We omit all ticker symbols that yield only monthly or no search volume data. Then, we calculate abnormal search volume as introduced in Da et al. (2011) by taking the 11 Using stock ticker symbols is less ambiguous than, e.g., a company’s name since internet users might search for a certain product or the company for other reasons than information on the equity. 12 log differences of SVI and the median SVI over the previous eight weeks.12 3.1.2 Household sentiment - FEARS To account for sentiment on the household level in our analyses, we use the Financial and Economic Attitudes Revealed by Search (FEARS) index proposed by Da et al. (2015). It is composed of thirty economics related search phrases such as “gold prices” or “social security card” and reflects the general negative sentiment among the population. The log changes of the respective SVI values are winsorized at the 5% level and are adjusted for seasonality and possible heteroskedasticity. FEARS is then calculated as the average of the respective thirty time-series. 13 In this regard, higher values of FEARS indicate a higher level of negative sentiment. 3.2 Additional determinants of CDS spreads In addition to our proxies of creditor attention and sentiment, we shortly comment on additional variables that should in theory drive default probabilities and thus prices of default-sensitive securities. All variable definitions and data sources are provided in Appendix I. The set of potential determinants of CDS spreads can broadly be classified into variables that are derived from structural models of credit risk and variables that relate to the trading and hedging of CDS contracts. In the first group, theoretical determinants are extracted from Merton’s 1974 classical model of credit risk and subsequent extensions (see, e.g., Longstaff and Schwartz, 1995; Leland and Toft, 1996; Duffie and Lando, 2000). All of these models center around a firm’s leverage as the key determinant of a firm’s probability to default. Consequently, leverage (or equivalently firm value) should be a major driver of CDS spreads with increases in leverage driving up premia for credit protection via a CDS. Two common problems associated with a firm’s usual leverage ratio are that it cannot be directly measured at higher frequencies and that does not capture off-balance sheet items. We therefore follow Christie (1982) and Collin-Dufresne et al. (2001) 12 Note that we lose value for the first eight weeks when including this variable in future regression analyses. Note that we download weekly search volume data for the list of thirty words and adjust the winsorized time series using month dummies. 13 13 and employ a company’s firm value instead of the leverage ratio and proxy a firm’s change in firm value by using weekly arithmetic stock returns. To compute our variable Firm value, we use daily stock prices taken from Datastream. A second potential determinant of default risk that emanates from most structural models is a firm’s asset volatility. In these models, an increase in asset volatility leads to an increase in default risk and should thus be reflected in a higher CDS spread of the respective firm. In our empirical study, we follow Ericsson et al. (2009) and include a proxy for individual equity volatility. The volatility is calculated with GARCH(1,1)-models for conditional variances using a rolling window of 52 weeks.14 In addition to asset volatilities, CDS spreads should also in theory be significantly influenced by changes in the risk-free rate. As the risk-free rate increases, the firm’s asset value process should move away from the default barrier leading to lower CDS spreads. We employ the two-year U.S. Treasury Benchmark yield, as our variable Risk-free interest rate. The second set of theoretical determinants comprises variables that are only indirectly connected to a firm’s default risk. For example, studies by Tang and Yan (2013, 2014); Bongaerts et al. (2011); Lesplingart et al. (2012), and Junge and Trolle (2015) all document a premium in CDS spreads for the contracts’ illiquidity. We account for a possible influence of illiquidity on CDS spreads in our empirical study by performing separate regressions on bid and ask quotes of CDS contracts in unreported (successful) robustness checks (see also Ericsson et al., 2009; Meine et al., 2015b). In addition to liquidity, CDS spreads could also be influenced by the contract’s joint tail risk with the CDS market. As Meine et al. (2015b) show in their study for European banks, protection sellers receive a premium for bearing the risk of extreme upward co-movements in default risk. While it is not clear that CDS of non-financial firms carry a similar tail risk premium, we neverthe14 Various studies on the drivers of credit and CDS spreads (see, e.g., Collin-Dufresne et al., 2001; Cremers et al., 2008; Alexander and Kaeck, 2008; Junge and Trolle, 2015) employ volatilities extracted from individual equity options. Sometimes, values of the VIX implied volatility index to measure the market’s expectation of future stock market volatility are used as well. Alternatively, one could also have used historical volatilities as it is done, e.g., by Cremers et al. (2008). 14 less include a variable Tail risk in our regressions that is built by estimating the upper tail dependence between individual CDS spreads and a relevant CDS market index. The CDS tail risk estimates are computed using the Dynamic Asymmetric Copula (DAC) model of Christoffersen et al. (2012) and Christoffersen et al. (2013) based on the skewed t-copula with both the copula correlations and the degrees of freedom being time-varying. Complementing these potential drivers of CDS spreads, there is also ample evidence in the empirical literature (see, e.g., Longstaff et al., 2005; Zhang et al., 2009) that credit risk premia are sensitive to changes in the business climate in which a firm operates and additional state variables. We thus include end-of-week values of the S&P 500 index in our regressions to account for general stock market momentum. We expect increases in the S&P 500 to mirror decreases in overall default risk as well as increases in recovery rates. Positive index changes should thus be negatively correlated with CDS spreads in our sample. Finally, we also employ the slope of the yield curve as a further explanatory variable in our regression analyses. Here, we use the yield curve slope as an indicator for the country-wide future economic activity of a bank’s home country. Our variable Slope is defined as a country’s respective 10-year minus 2-year government bond benchmark yields. Data on government bond yields are taken from Datastream. In theory, spot rates converge to their long-term counterparts, thereby increasing the risk-neutral drift of the asset value process making default less likely to occur (see Longstaff et al., 2005). Nevertheless, also monetary policy measures may be reflected in the slope coefficient. Hence, we expect the direction of the effect of the yield curve slope on CDS spreads to be unrestricted. 3.3 Econometric design We now briefly present the empirical approach to test our hypotheses on the relation between default sentiment and CDS spreads. We first test whether default sentiment drives the time-series and cross-section of CDS spreads. Therefore, we perform a series of time-series regressions of CDS spreads on independent variables including our proxy for default sentiment and control vari15 ables such as the three Merton-type variables proposed by structural approaches. For each CDS contract, we run regressions in first differences of the following type: ΔCDSt = α + β × ΔDefault Sentimentt + γ × ΔMerton-typet + Θ × ΔControlst + εt , where CDSt is the spread of a CDS mid-quote, Default Sentimentt is the sentiment measures constructed from Google Trends data and Merton-typet and Controlst are various control variables that have been identified as significant drivers of CDS spreads in the theoretical and empirical literature (see Section 3.2). We calculate corresponding t-statistics as outlined in Collin-Dufresne et al. (2001) and Ericsson et al. (2009) and estimate all regressions in first differences to remove the non-stationarity in our time-series (see also Meine et al., 2015a,b).15 Further, we perform panel regressions in first differences, similar to the regression equation above, that also account for crosssectional effects and clustering on the firm level. In most of our regressions, we include firm-fixed effects and also employ year dummies to capture unobserved heterogeneity over time. By employing fixed effects, we mitigate the bias that might arise from omitted variables that jointly determine our measure of default sentiment and other covariates. Further, it is very unlikely that overall levels of search volume on generic phrases such as “loan default” are driven by CDS market participants, e.g., as a result from downward CDS market trends.16 We therefore believe that our regressions of (differences in) CDS spreads on default sentiment do not suffer from reverse causality. To better understand the channel through which households’ sentiment might affect CDS prices, and therefore test the hypothesis that sentiment increases CDS market liquidity, we also estimate additional regressions in which we include several measures of individual or market-level CDS liquidity to test whether default sentiment is in fact partially replacable by liquidity effects or whether it accounts for unobserved effects that determine CDS spreads (credit risk hypothesis). 15 As Ericsson et al. (2009) point out, it is often harder to explain differences than levels, which is why we opt for this specification. 16 For example, comparing the search volume on the phrase “loan default” with a more frequent word such as “entrepreneurship” reveals that a relatively high search frequency is needed to generate the displayed relation. 16 Similar to our other regressions, we run panel regressions in first differences of the following form: ΔCDSit = αi + μt + β × ΔDefault Sentimentt + γ × ΔLiquidityit + Θ × ΔControlsit + εit , where αi and μt are firm-fixed effects and year dummies, respectively.17 In further analyses, we also include interaction terms between default sentiment and our liquidity proxies. 4 Empirical analysis In this section, we discuss the central features of our data sample and present our empirical results. 4.1 Sample construction and descriptive statistics Our data set consists of all single-name CDS time series for U.S. firms that have a traded CDS contract on their debt available in the Credit Market Analysis (CMA) database via Thomson Reuters Financial Datastream. 18 For these firms, we download CDS data with available daily mid, bid, and ask quotes. We use CDS contracts with a maturity of five years, as these are considered the most liquid, and those that refer to senior-debt issues denominated in U.S. dollar. We start with the full set of U.S. firms that have a traded CDS contract and are listed on an U.S. stock market.19 Afterwards, we match the CDS symbols from Thomson Reuters with the respective equity symbol to obtain stock price data later on. 20 Our final sample consists of 228 CDS contracts from 19 different industries over 351 weeks which leaves us with a maximum of 80,028 firmweek observations. For increased transparency, all firm names and their ticker symbols are listed in Appendix IV. Finally, for each firm i in our sample, we then calculate a corresponding CDS 17 In our set of robustness checks, we also include more granular time effects such as week dummies. The CMA database has several advantages over other providers such as reliability and reporting of only sufficiently often quoted CDS contracts (see Mayordomo et al., 2014). 19 We further exclude U.S. sovereign debt issues. 20 We follow Meine et al. (2015b) and extract from each CDS identifier the firm name and add a prefix (“U:”) to get the identifier of the firm’s equity. Whenever we find an ambiguous match or do not find significant variation in the time series of CDS quotes, we delete it from our sample. 18 17 market index by equally-weighting all individual CDS contracts in our sample excluding the CDS of firm i. Appendix I shows data sources, definitions of variables used in this study, and respective descriptive statistics. Average CDS spreads for the full sample period are 152.44 basis points (bps hereafter). The minimum is 1 bps and the maximum is 12,577 bps indicating that the sample is characterized by strong cross-sectional variation. Mean stock returns are around 0.19% and the range is from -84.7% and +295.12%. Looking at the spread of the bid- and ask-quotes also reveals a wide range between the minimum and maximum value and a high standard deviation of 31.26 bps. The market’s bid-ask-spreads, however, are slightly narrower. We now present empirical evidence on the interplay of our variables of interest. Figure 1 shows the time evolution of CDS spreads, CDS bid-ask-spreads, and the Default Sentiment Index (DSI) from 2004 to Q3 2010. The grey-shaded areas present the 20%- and 80%-quantile of CDS spreads (upper panel) and CDS bid-ask-spreads (lower panel). Insert Figure 1 about here. We observe that the mean CDS spread is very close to the 80% quantile almost all the time, which hints at the presence of few firms with extremely high spreads in our sample. The evolution of CDS spreads is relatively flat until the beginning of 2008, when mean values spike at around 200 bps. In the end of 2008 and beginning of 2009, however, CDS spreads are even higher with 80% quantiles being above 600 bps before dropping to an average of 200 bps after the crisis. The mean bid-ask-spread of CDS contracts for the full sample is around 11 bps, but the spreads we observe before the crisis period are constantly below 10 bps indicating a highly liquid CDS market. In late 2007, the market dried up a little with increasing bid-ask-spreads until 2009, when we find extremely high mean values of CDS bid-ask-spreads, possibly driven by few outliers. Additionally, we plot the time evolution of the weekly Default Sentiment Index (DSI). The index starts at relatively high values but has a downward trend until the beginning of 2007. Throughout the pre-crisis period, weekly default sentiment is more volatile while CDS (bid-ask-)spreads 18 are low and stable during this time. Starting in 2007, we observe an upward trend of the index during the crisis, but it reaches its maximum after the crisis in 2010, when bid-ask-spreads and CDS premia were low again. 4.2 Does default sentiment drive CDS spreads? In this section, we present our benchmark results from time-series and panel regressions in first differences of CDS spreads and control variables to see whether default sentiment significantly drives CDS premia. 4.2.1 Time-series regressions The results from the time-series regressions are given in Table I. Insert Table I about here. First, we run time-series regressions for every firm including the first differences of firm value, interest rate, and equity volatility as explanatory variables. These are the three determinants proposed by the structural approach derived from the work of Merton (1974). The average coefficients are reported in columns (1) to (3). Here, higher firm value of the reference entity is associated with a lower spread of its credit default swap. This is the opposite of what is suggested by theoretical models and found, e.g., by Ericsson et al. (2009). The difference in our study is the weekly frequency of the data as opposed to monthly values in Ericsson et al. (2009). However, we find that the coefficient of firm value remains negative in all of our regressions.21 Similarly, the average coefficient of the interest rate variable is highly significant with a negative sign. Higher asset (equity) volatility, on the other hand, is significantly, positively correlated with CDS premia, which confirms results found in previous studies as well (see, e.g., Ericsson et al., 2009; Meine et al., 2015b). Column (4) shows results from regressions including all three variables at once and confirms our 21 This incompability with theory is also found in Meine et al. (2015b) and Meine et al. (2015a) using quarterly data. 19 previous results. However, the average adjusted R 2 is only 8%, which underlines the limitation of the structural approach to explain CDS spreads in a comprehensive way. In column (5) and (6), we include in our regressions the proxy for default sentiment as our only explanatory variable and along with the Merton-type variables, respectively. The resulting coefficient in (5) is -1.069 and statistically significant at the 1% level. According to these timeseries regressions, our proxy for default sentiment is negatively related to CDS premia and enters the analysis with a highly significant coefficient. As sentiment is not supposed to influence a market dominated by institutional traders, we ask the question whether the significance of our sentiment measure is due to confounding effects caused by the absence of variables that proxy for the general business climate in capital markets. To shed more light on the predictive power of the DSI, we thus include our macroeconomic variables introduced in Section 3.2. Higher values of option implied market volatility (VIX) are associated with higher CDS premia, which is similar to the effect of the equity volatility of the reference entity. Better business climate, represented by higher values of the S&P 500 benchmark index, and a higher slope of the yield curve are negatively related to CDS spreads with the coefficients being statistically significant at the 1% level. Including these variables in columns (8) and (9) weakens the statistical significance of the DSI slightly but does not alter the result found in previous regressions. Interestingly, the average adjusted R 2 increase from 8% to around 15% and 18%, respectively, indicating that these macroeconomic controls posses considerable explanatory power. In column (10), we report the coefficients from timeseries regressions involving the CDS tail beta of the individual default swap. The positive relation with CDS spreads is also found in Meine et al. (2015b) and the influence of DSI is qualitatively the same. In summary, our sentiment proxy remains a significant driver of CDS spreads in all regression specifications. 4.2.2 Panel regressions Next, we turn to panel regressions of first differences in credit default swap spreads on first differences in default sentiment and control variables. This time, we are able to include firm-fixed 20 effects (firm FE) and year dummies in our model to account for unobserved heterogeneity across firms and time. The results are shown in Table II with t-statistics corrected for clustering at the firm level. Insert Table II about here. As before, including the variables firm value and interest rate reveals the same negative and significant relation to CDS spreads as in the time-series regressions. The third Merton-type control variable, equity volatility, however, enters the regressions with an insignificant coefficient, although it is positive as predicted. Including all three Merton-type variables and the DSI in one regression does not change our conclusions from the analysis before. Our proxy for default sentiment significantly lowers CDS spreads. However, we notice that the coefficient of the DSI in our panel regressions does not drop as much as in the time-series regressions when we include additional variables in our setup. We repeat the last four regressions from Table I and report the findings of the panel regressions in column (6) to (9). The additional control variables are, again, highly statistically significant with the respective coefficient carrying the expected sign. Most notably, our proxy of default sentiment remains highly relevant in all of the regressions. Concluding these first analyses, we notice a highly significant negative relation between CDS spreads and the default sentiment index based on internet search volume data. This index represents the procurement by households in the U.S. of information on bankruptcy and credit related topics and thus, indicates worries concerning rises in default risk. Naturally, one would not expect that household sentiment could spill over to credit markets, which are dominated by institutional investors rather than individual traders and thus, prices should not be affected by it. The fact that we find such highly significant correlation raises the question on the possible channels through which the market-wide sentiment could influence the prices in the CDS market. One of these possible channels identified in the theoretical framework in section 2 is summarized in our market liquidity hypothesis, which explains the reduction in CDS spreads when default sentiment is high. We further explore this hypothesis in the next section. 21 4.3 Default sentiment and the market liquidity hypothesis Our next set of regressions analyzes the interplay of CDS prices and (market) liquidity in the credit default swap market. If the channel through which default sentiment influences CDS spreads is the CDS market’s liquidity, we expect that the DSI and liquidity are interchangeable when explaining CDS premia. To test the market liquidity hypothesis, we include in our regressions several variables that account for liquidity effects in the CDS market along with our measure of default sentiment and other controls. The liquidity proxies include a contracts Bid-Ask-Spread BASabs , which is the end-of-week absolute CDS bid-ask spread, calculated as daily CDS ask quote minus CDS bid price and denoted in bps, the Sector Bid-Ask-Spread BASIabs calculated as the cross-sectional average of absolute bid-ask spreads of all firms within an industry, and the Market Bid-Ask-Spread BAS M abs calculated as the cross-sectional average of relative bid-ask spreads of all sample firms excluding the respective firm. CDSILLIQ is the end-of-week value of the CDS market illiquidity measure proposed by Junge and Trolle (2015). Table III shows results from panel regressions that include first differences of the three Merton-type variables, the default sentiment proxy, and one of the liquidity proxies as explanatory variables at a time. Insert Table III about here. First, we employ the individual bid-ask-spread of a CDS contract along with the DSI. The former of the two variables has a statistically significant positive coefficient in our regression while the latter enters the regression with a negative sign. Different from our previous analysis, we find that the DSI is only significant at the 10% level indicating that the individual (il)liquidity of a CDS contract partially replaces the explanatory power of default sentiment. 22 Next, we use the cross-sectional average of (absolute) bid-ask spreads within an industry and include it along with default sentiment in our regression. Using the sector bid-ask-spread, we are able to see whether liquidity effects on CDS spreads that are specific to a certain industry interact 22 The positive relation of illiquidity and CDS spreads is also documented by, e.g., Tang and Yan (2014). 22 with the DSI. Similar to the individual bid-ask-spread, we find a significant and positive relation to CDS premia. The relevance of default sentiment, however, is not reduced in this case. Including both the CDS market bid-ask-spread and default sentiment in our regressions paints an interesting picture of the relation between the two variables: While market illiquidity and CDS spreads have a strong positive relation with a statistically significant coefficent at the 1% level, the significance of default sentiment vanishes. This supports the rationale of the market liquidity hypothesis with sentiment appearing to drive aggregate CDS market liquidity. The presence of high default sentiment could be associated with a more liquid CDS market and thus, reduces the average spread of CDS contracts as liquidity commonalities (see Meine et al., 2015a) and liquidity risk (see Junge and Trolle, 2015) directly impact CDS spreads. Using the end-of-week value of CDSILLIQ immensely reduces our sample size compared to the other three models. The results are nevertheless similar: liquidity and CDS spreads and default sentiment and CDS spreads are negatively related at a very high significance level. For the three Merton-type variables, we find the exact same signs for the coefficients as in our baseline panel regressions. Again, firm value and interest rate are highly significant while volatility is not relevant for explaining CDS spreads in our sample. We repeat these four analyses including year dummies to capture time effects. However, the results reported in column (5) to (8) are qualitatively and quantitatively very similar to the results from the regressions in (1) to (4). 4.4 Exploring the market liquidity hypothesis In our regressions including the proxies for liquidity in the CDS market, we notice that the Default Sentiment Index may be partially replacable by these variables when explaining CDS premia. As our next step, we analyze the dynamics of the relation between CDS (market) liquidity and default sentiment. First, we filter out the sentiment-caused component from the liquidity proxies by running time-series regressions (for each firm) in first differences of the original liquidity I variables BASabs , BASM abs , and BASabs on first differences of the Default Sentiment Index. From each regressions, we extract the respective residual and include these along with the DSI in further 23 panel regressions. The results of these analyses are given in Table IV. Insert Table IV about here. The residuals of the liquidity measures represent the portion of liquidity effects that are not explained by default sentiment from outside of the credit default swap market. Including only the liquidity residuals and default sentiment along with fixed effects and year dummies yields the following result: the DSI coefficient enters all regressions with a negative sign with the coefficients being significant at the 1% level. Thus, the effect induced by default sentiment remains relevant for the pricing of credit default swaps. The effect of (il)liquidity residuals on CDS spreads is also highly statistically significant with the well known positive sign of the coefficients. We observe that the change in the CDS market bid-ask-spread has the strongest effect on CDS premia with a coefficient of 2.326. If CDS liquidity and default sentiment were completely unrelated (which is what we would expect intuitively), the significance of our default sentiment variable would remain an unsolved puzzle. However, we do find that there is a strong correlation between household level default sentiment and aggregate liquidity in the CDS market. Assuming that DSI indeed drives CDS market liquidity, 23 we conclude that even if we remove the sentiment from the bid-ask-spreads by taking the residuals, we find a significant relation to CDS spreads. This is an indicator for the fact that CDS market liquidity is not entirely driven by sentiment measures but may be influenced by it in certain times. We conclude that liquidity may influence CDS prices on its own, but the magnitude may be higher in times of higher default sentiment. These observations motivate a more detailed analysis into the interaction of default sentiment with our liquidity proxies. Table V shows panel regressions (in first differences) of CDS spreads on interaction terms of CDS liquidity proxies and the default sentiment variable. Insert Table V about here. 23 It is very unlikely that such sentiment is created from institutional investors alone, since it is necessary to generate a high search volume for these rather generic terms to be displayed by Google Trends. Also, we compare our search phrases to very general word combinations to assess the magnitude of the respective search volume. 24 We interact first differences of default sentiment and individual or market bid-ask-spreads and include these along with the two variables in respective regression models. Most strikingly, we observe that the DSI is insignificant in the absence of market illiquidity. Neither of the coefficients of DSI in model (1) to (6) shows any relevant statistical significance. The main effects of individual and market bid-ask-spreads, however, are highly statistically significant on the 1% level when explaining CDS premia. When there is zero sentiment, illiquidity of CDS contracts will still increase their premia, which is in line with the findings of Bongaerts et al. (2011); Junge and Trolle (2015). This also complements our findings of the analyses including the residuals of the regressions of bid-ask-spreads on default sentiment. We find that all coefficients of interaction terms enter the regressions with a negative sign that is statistically significant at least at the 10% level. Obviously, when default sentiment is high, the effect of the bid-ask-spreads on CDS premia is reduced. In terms of liquidity, we observe that default sentiment increases the effect of liquidity (since it reduces the effect of illiquidity) and thus find support for our hypothesis on the channel of the effects of such sentiment on CDS prices. This relation is even more pronounced for the market bid-ask-spreads and the DSI. 4.5 Are some CDS contracts more sensitive to default sentiment than others? Looking at the significant interaction of default sentiment and CDS liquidity, we investigate whether the sensitivity of the market participants towards default sentiment plays a role in determining CDS prices. We argue that when firms are more sensitive to such specific sentiment, it should be reflected in their CDS prices. As a standard measure of sensitivity or systematic risk, we employ the betas and R2 s of the following regressions: BASit = α + βDSI × Default Sentimentt + εit , 25 where BASabs is the end-of-week Bid-Ask-Spread of a CDS contract. Figure 2 shows plots of the resulting average R2 s for each industry sector in our sample.24 Insert Figure 2 about here. In total, we have 18 average R2 s for different sectors with at least four firms.25 In most cases, the average R2 is highest in the crisis years 2008 and 2009. However, we find several differences in the sensitivity of bid-ask-spreads to default sentiment across sectors. Industrial goods and services is the sector with the most firms in our sample. Its mean R 2 ranges from approximately 2% in 2004 to around 7.5% in 2006 before the crisis. The maximum value is about 15% in 2009. Similar patterns can be found in the technology, chemicals, retail and construction and material sector. Compared to other sectors, insurance companies’ R 2 s from regressions of CDS bid-askspreads on default sentiment appear to be less volatile over time. Values in the three years before the crisis are slightly above 5% and peak around 15% in 2008 before dropping to pre-crisis levels in 2009. One interpretation is that parts of the insurance sector have been involved more heavily with the credit markets than other sectors and thus, could be more exposed to default sentiment. Interestingly, we find a different pattern in the healthcare sector, where the highest influence of default sentiment on CDS illiquidity can be found in 2006. CDS contracts on reference entities in the sectors basic resources, personal and household goods, and automobile and parts are influenced by default sentiment during the two crisis years with R 2 s surpassing values of 15%. Goodness-of-fit is very low in 2007 and similar to other sectors in 2006 with values of about 10%. Finally, we turn to the firms within the financial services, real estate, and banking sector. These types of firms are expected to suffer the most from credit and business defaults and thus, should be particularly prone to default sentiment. What we observe in the average R 2 s of the 17 firms in the real estate sector is a monotonic increase in the ability of default sentiment to explain CDS illiquidity starting at 2.5% and reaching 7% in 2007. In 2008 and 2009 we see a jump in these numbers to 15% and 17.5%, respectively, which are among the highest values across all sectors. We calculate the R2 for each year from 2004 to 2010 (only up to Q4 in 2010 due to data availability) and every issuer in our sample. 25 The telecommunications is excluded as only one firm in our sample is from that sector. 24 26 This is only surpassed by the value of 20% in 2008 in the banking sector, which may be explained by the fact that the crisis originated in that sector and thus, credit defaults were associated with banks more heavily than with other firms. However, we only find such a high value in 2008, while in all the other years, bank CDS bid-ask-spreads were not as affected by default sentiment as other sectors. The pattern for the financial services sector is very similar to the real estate firms, but on a slightly lower level. We conclude that the impact of default sentiment based on search volume of credit related phrases has different effects across industries and may also vary over time. Next, we take a closer look at average CDS spreads for firms that are most sensitive to changes in market-level values of liquidity and default sentiment. To determine the sensitivity of each CDS contract, we perform regressions of the following type: i × MLIQt + εit , BASit = α + βLIQ-MLIQ i × DSIt + εit , BASit = α + βLIQ-DSI i × MCDSt + εit , BASit = α + βLIQ-MCDS where MLIQt is the average value of all individual bid-ask-spreads (market illiquidity), DSI t is the Default Sentiment Index, and MCDSt is the average CDS spreads in week t. For each reference i i , βLIQ-DSI , and entity i, we run simple OLS regressions to obtain the coefficient estimates β LIQ-MLIQ i , which represent the systematic risk stemming from the sensitivity of an individual bidβLIQ-MCDS ask-spread to market illiquidity, default sentiment, and average CDS spreads. Afterwards, we split our sample into quintiles of these betas and calculate the weekly timeseries of average CDS spreads for each quintile. In Figure 3, we plot the time evolution of the results. Insert Figure 3 about here. Comparing the plots for all three betas, we notice that they paint an almost identical picture. Indeed, the pairwise correlations of the beta vectors are about 98% and 99%. This leaves us with 27 the conclusion that the same group of firms that are sensitive to market illiquidity in their bid-askspread, are also sensitive to changes in default sentiment and average market CDS spreads. We also see how the level of CDS spreads differs across the different quintiles of the betas introduced above. Generally, the higher the beta for a firm, the higher are the average CDS spreads. The difference across the first four quintiles of betas is only marginal compared to the fifth quintile. CDS contracts with the highest sensitivity of bid-ask-spreads to market-level values of illiquidity, sentiment, and CDS spreads have much higher average CDS spreads. The CDS spreads are relatively constant from 2004 to 2007 and increase during 2008. Extremely high peaks of CDS spreads can be found shortly before and after 2009, in the midst of the financial crisis. In a nutshell, we find that systematic risk stemming from default sentiment has very similar effects to systematic liquidity risk that has been identified in the literature before (see Bongaerts et al., 2011; Tang and Yan, 2014; Junge and Trolle, 2015). 4.6 Credit risk hypothesis and equity market spillover The evidence found so far speaks in favor of the market liquidity hypothesis, i.e., default sentiment from outside the CDS market causes inattentive traders in the CDS to overreact therby increasing CDS market liquidity and reducing CDS spreads. However, we can not let go of the fact that the DSI most likely captures more than just CDS market liquidity. Therefore, we test whether the fraction of default sentiment that does not account for liquidity effects is able to, at least in parts, support the validity of the credit risk hypothesis as well. To do so, we obtain residuals from simple OLS regressions of the DSI on individual, sector-wide, and rather market-level bid-askspreads (in first differences). We employ these residuals as regressors in our panel regressions along with the respective liquidity proxy. Results are shown in Table VI. Insert Table VI about here. Columns (1) to (3) show estimation results that only include the liquidity proxy, the residuals, and 28 firm-fixed effects and year dummies whereas models (4) to (6) also employ the various control variables from previous analyses. First of all, we find a strong, positive relation of bid-ask-spreads and CDS spreads on both the individual and market-wide level. This is underlined by the significance level of 1% in almost each of the regressions. Looking at the residuals from the regressions of DSI on individual bid-askspreads, we observe that it is negatively related to CDS spreads and slightly statistically significant on the 10% level. We find that the fraction of default sentiment that does not account for liquidity effects actually reduces average CDS spreads as well. If this part of the DSI was to increase CDS spreads, it could have supported a weaker form of the credit risk hypothesis. However, we find that this hypothesis is rejected again by our regression results. Testing the robustness of this significance by including our set of known determinants of CDS spreads leads to the vanishing of this slight significance. A qualitatively and quantitatively similar result is obtained when we employ the sector bid-askspreads. Although we do find the same negative relation of the DSI and market bid-ask-spread residual, we do not find any significance, which again supports the rejection of the credit risk hypothesis in this setup. Da et al. (2011, 2015) have shown that prices in equity markets are prone to noise trading and household-level sentiment using measures of retail investor attention based on Google Trends data (see section 3.1.1 and 3.1.2). If such proxies of investor sentiment have a profound influence on equity markets, this could, in theory, spill over to other markets as well. This leading relation of movements in equity markets on CDS markets has recently been documented by Hilscher et al. (2015). We therefore investigate the question, whether default sentiment may influence the CDS markets via those spill overs from equities. Moreover, we test whether the known sentiment and attention measures also affect CDS prices in the way they drive stock prices. Our motivation to choose the set of phrases was to capture the information demand of creditors and households on credit related topics. We argue that such behavior is not simply captured by obtaining search volume on single words, which might be out of context. By combining several words we are able 29 to bring specific meaning to the searches of internet users. 26 To test the validity of our constructed measure, we include the FEARS index introduced in Da et al. (2015), which is based on single generic search terms that might have a negative connotation. In Table VII, we show results from panel regressions including the abnormal search volume on a reference entity’s stock ticker symbol as introduced in Da et al. (2011) (Panel A) and the FEARS index from Da et al. (2015) (Panel B). Insert Table VII about here. In column (1), we include the abnormal search volume on ticker symbols (ASVI), the DSI, and our control variables in the regression model. The individual investor attention is positively related to CDS spreads. One could imagine that stocks that gain higher attention by retail investors, are the ones that experience more negative news coverage. Moreover, attention to individual ticker symbols is more likely to reflect actual changes in fundamentals in contrast to indiscriminate searches on default risk. Consequently, attention to individual reference entities appears to reflect (perceived or actual) increases in credit risk and thus increases in CDS spreads.27 More importantly, however, the estimate for the DSI coefficient remains negative and significant at the 10% level in this regression. Next, we also employ an interaction term of DSI and ASVI that accounts for the interplay between default sentiment and individual retail investor attention. This variable, however, is insignificant in all of the regressions in Panel A and indicates that the two main effects are isolated. In the other three models, (3), (4), and (5), we use the three bid-ask-spreads along with the other covariates. As in our main results, the default sentiment still has a negative influence on the level of CDS spreads, but is weakened and not significant at the usual statistical levels thereby again confirming the market liquidity hypothesis. The effect of ASVI on CDS premia remains positive and significant, which indicates only little interaction of (market) liquidity and investor attention, 26 For example, a search on “credit” could represent a variety of incentives whereas “credit default” provides a more nuanced interpretation. 27 Investor attention could be directed at a firm’s CDS or its stock, in which case the effect of investor attention could still spill over to the CDS market (see Hilscher et al., 2015). 30 compared to the strong relation between default sentiment and liquidity. We now turn to Panel B of Table VII in which we show regression results using the FEARS index. Higher values of FEARS indicate more negative sentiment among households and expresse their negative expectations of the general outlook on the economy. FEARS represents economic related views of the population but is not tied to specific themes that are necessarily close to credit markets. First, we notice that DSI is again significant with a negative sign but the effect is weaker when we include our liquidity variables. The FEARS index, however, is strongly positively related to CDS spreads. Higher negative sentiment in the overall economic household environment is associated with higher average CDS permia. This relation could, again, arise from the documented spillover effects from the equity markets and also funds, which are influenced by the household sentiment measure (see Da et al., 2015; Hilscher et al., 2015). 5 Conclusion In this paper, we show that sentiment affects the cross-section and time-series of CDS spreads. We employ internet search volume data to proxy for an exogenous shock to the pessimistic sentiment of households about credit risk at the market level. As our first result, we document the existence of significant noise trading in the CDS market as we find our proxy of household sentiment to be a statistically significant and economically important driver of CDS spreads. In particular, CDS spreads and sentiment are negatively correlated with buyers of credit protection earning a premium when market-wide sentiment is higher. After carefully testing the robustness of this finding, we test and confirm the hypothesis that sentiment affects CDS spreads via the market liquidity channel. Consistent with theoretical models that view spikes in market liquidity as an indicator of the presence of noise traders in the market, we find higher household sentiment concerning increases in default risk to be significantly positively correlated with changes in CDS market liquidity. Through liquidity commonalities and liquidity risk, the higher CDS market liquidity then translates into lower CDS spreads. The results we find 31 thus contribute significantly to the current discussion of the unclear role liquidity risk plays in the pricing of CDS contracts. 32 Appendix 33 34 0.02 11.25 11.28 11.25 ASVI BASabs BASIabs BASM abs 27.95 Volatility -1.08 (×10−3 ) 2.79 (Risk-free) Interest rate FEARS 0.19 Firm value -0.02 (×10−9 ) 152.55 CDS DSI Mean Variable 8.77 16.97 31.26 0.58 0.53 1.00 35.52 1.55 6.14 342.97 St. Dev. 4.11 1.00 0.00 -4.42 -2.31 -2.12 0.08 0.42 -84.7 1.00 Min. 78.91 987.80 3,201.00 4.62 1.96 2.15 2,369.14 5.25 295.12 12,577.58 Max. 80,028 79,677 80,020 62,192 351 351 79,264 351 79,570 80,020 Obs. Cross-sectional average of relative bid-ask spreads of all sample firms. Averages are calculated excluding the current firm and are reported in bps. Cross-sectional average of absolute bid-ask spreads of all firms within an industry. ICB supersector industry classifications are obtained from DS. Averages are calculated excluding the current firm and are reported in bps. End-of-week absolute CDS bid-ask spread, calculated as daily CDS ask quote minus CDS bid price and denoted in bps. Weekly abnormal search volume (SVI) of a firm’s ticker symbol in Google Trends. log [GSVIt ] − log [median(GSVIt−1 , . . . , GSVIt−8 )]. Weekly FEARS-index introduced in Da et al. (2015). FEARS is the average of the log changes in search volume of thirty economics related phrases, adjusted for seasonality and heteroskedasticity. The weekly DSI is the average of the Google Search Volume Indices (GSVI) for thirty default-related search phrases (see AppendixIII), winzorized at the 5% level and adjusted for heteroskedasticity and seasonality. Weekly equity return volatility, denoted in %. Volatility is calculated with GARCH(1,1)-models for conditional variances using a rolling window of 52 weeks. End-of-week two-year U.S. Treasury Benchmark yield, denoted in %. Weekly return on a firm’s equity, denoted in %. End-of-week CDS mid quote (in levels), denoted in basis points (bps). Definition CMA, EST CMA, EST CMA GT, EST GT, EST GT, EST DS, EST DS DS, EST CMA Source The table presents summary statistics, definitions, and data sources for all dependent and independent variables that are used in our empirical study. The data sources are Credit Market Analysis (CMA), Thomson Reuters Datastream (DS), and Google Trends (GT). EST indicates that the variable is estimated or computed based on data from the respective data source(s). Our sample encompasses 228 financial and non-financial companies for the period from January 2004 to September 2010. Appendix I: Variable definitions, descriptive statistics, and data sources. 35 20.72 1205.47 0.12 1.25 S&P 500 CDS tail beta Slope 9.10 CDSILLIQ VIX Mean Variable 1.01 0.11 181.63 11.17 13.49 St. Dev. -0.19 0.00 712.87 9.89 0.76 Min. 2.86 0.77 1562.47 74.26 78.58 Max. 351 79,787 351 351 211 Obs. Weekly U.S. 10-year minus 2-year Treasury Benchmark yields. Proxy of a CDS contract’s tail risk together with a CDS market index as proposed by Meine et al. (2015b) and estimated by using the Dynamic Asymmetric Copula (DAC) model of Christoffersen et al. (2012). Weekly values of the S&P 500 index. Weekly values of the Chicago Board Options Exchange Market Volatility Index measuring the implied volatility of S&P 500 index options. End-of-week values of the CDSILLIQ measure proposed by Junge and Trolle (2015). Definition DS, EST DS, EST DS DS http://sfi.epfl.ch/trolle Source The table presents summary statistics, definitions, and data sources for all dependent and independent variables that are used in our empirical study. The data sources are Credit Market Analysis (CMA) and Thomson Reuters Datastream (DS). EST indicates that the variable is estimated or computed based on data from the respective data source(s). Our sample encompasses 228 financial and non-financial companies for the period from January 2004 to September 2010. Appendix II: Variable definitions, descriptive statistics, and data sources (continued). Appendix III: Default Sentiment Index word list after bankruptcy bank bankruptcy bank default bankruptcy and credit bankruptcy credit cards bankruptcy credit counseling bankruptcy credit report bankruptcy filing bankruptcy filings bankruptcy laws bankruptcy loans business bankruptcy company bankruptcy credit after bankruptcy credit bankruptcy credit card bankruptcy credit card debt credit card default credit counseling credit crunch 36 credit default debt bankruptcy debt consolidation debt settlement file bankruptcy filing for bankruptcy interest default liquidation sale loan default personal bankruptcy Appendix IV: Sample firms and ticker symbols Name Ticker Name Ticker 3M COMPANY ABBOTT LABORATORIES ACE LTD ADVANCED MICRO DEVC INC AES CORP AETNA INC AIR PRODUCTS & CHEMS INC AK STEEL CORP ALCOA INC ALLERGAN INC ALLIANT EN CORP ALLSTATE CORP ALTRIA GROUP INC AMERICAN AXLE & MNFG INC AMERICAN ELEC PWR CO INC AMERICAN EXPRESS CO AMERICAN INTL.GROUP AMERICAN TOWER CORP AMERISOURCEBERGEN CORP ANADARKO PETROLEUM CORP ANHEUSER-BUSCH COS. INC APACHE CORP ARCHER-DANIELS-MIDL. CO ARROW ELECTRONICS INC ASHLAND INC AT&T CORP AUTOZONE INC AVALONBAY COMMNS. INC AVNET INC AVON PRODUCTS BAKER HUGHES INC BANK OF AMERICA CORP BAXTER INTERNATIONAL INC BEAZER HOMES USA INC BERKSHIRE HATH.INC BEST BUY CO INC BORGWARNER INC BOSTON PROPERTIES LP BOSTON SCIENTIFIC CORP BRISTOL-MYERS SQUIBB CO INTERNATIONAL PAPER CO INTERPUBLIC GP.COS. INC FEDEX CORP FELCOR LODGING LP FIRST INDUSTRIAL LP FIRSTENERGY CORP FORD MOTOR CO GANNETT CO INC GAP INC GATX FINANCIAL CORP GENERAL DYNAMICS CORP GENERAL MILLS INC GOLDMAN SACHS GP INC GOODYEAR TIRE & RUB. CO DANAHER CORP DARDEN RESTAURANTS INC DEAN FOODS CO DEERE & CO DELTA AIR LINES INC DEVELOPERS DIVR.REAL MMM ABT ACE AMD AES AET APD AKS AA AGN LNT ALL MO AXL AEP AXP AIG AMT ABC APC BUD APA ADM ARW ASH T AZO AVB AVT AVP BHI BAC BAX BZH BRKA BBY BWA BXP BSX BMY IP IPG FDX FCH FR FE F GCI GPS GMT GD GIS GS GT DHR DRI DF DE DAL DDR BRUNSWICK CORP CABLEVISION SYS CORP CAMERON INTL CORP CAMPBELL SOUP CO CAPITAL ONE BANK CARDINAL HEALTH INC CARNIVAL CORP CATERPILLAR INC CENTERPOINT ENERGY INC CHESAPEAKE ENERGY CORP CHUBB CORP CIGNA CORP CITIGROUP INC CMS ENERGY CORP COCA-COLA CO COLGATE-PALMOLIVE CO COMMERCIAL METALS CO COMPUTER SCIENCES CORP CONAGRA FOODS INC CONOCOPHILLIPS CONSOLIDATED EDISON INC COOPER TIRE & RUBBER CORNING INC CSX CORP CUMMINS INC CVS CAREMARK CORP CYTEC INDUSTRIES INC HALLIBURTON CO HARTFORD FINL.SVS.GP HEALTH CARE REIT INC HEALTHCARE REAL.TST. INC HERSHEY FOODS CORP HIGHWOODS REALTY LP HOME DEPOT INC HONEYWELL INTL.INC HOSPITALITY PROPS. TST HUMANA INC INTL.BUS.MCHS.CORP ILLINOIS TOOL WORKS INC INTERNATIONAL GAME TECH PFIZER INC PITNEY BOWES INC POLYONE CORP PPG INDUSTRIES INC PRAXAIR INC PROCTER & GAMBLE CO PROLOGIS TRUST PULTE HOMES INC NABORS INDUSTRIES INC NAVISTAR INTL.CORP NEWELL RUBBERMAID INC NEWMONT MINING CORP NIKE INCO NORDSTROM INC NORFOLK SOUTHERN CORP NORTHEAST UTILITIES NORTHROP GRUMMAN CORP NUCOR CORP NVR INCORPORATED OCCIDENTAL PETROLEUM CORP BC CVC CAM CPB COF CAH CCL CAT CNP CHK CB CI C CMS KO CL CMC CSC CAG COP ED CTB GLW CSX CMI CVS CYT HAL HIG HCN HR HSY HIW HD HON HPT HUM IBM ITW IGT PFE PBI POL PPG PX PG PLD PHM NBR NAV NWL NEM NKE JWN NSC NU NOC NUE NVR OXY 37 Appendix I: Sample firms and ticker symbols (continued) Name Ticker Name Ticker DEVON ENERGY CORP DIAMOND OFFSHORE DRL DOMINION RESOURCES INC DOVER CORP DOW CHEMICAL CO DR HORTON INC DU PONT E.I. DE NEMO URS DUKE ENRGY CAROLINAS DUKE REALTY LP EASTMAN CHEMICAL CO EATON CORP ELI LILLY & CO EMERSON ELECTRIC CO ENTERGY CORP ENTERPRISE PRDS.PTNS LP EVEREST RE GROUP EXELON CORP EXXON MOBIL CORP PEPCO HOLDINGS INC PEPSICO INC METLIFE INC MGIC INVESTMENT CORP MOHAWK INDUSTRIES INC JC PENNEY CO JOHNSON & JOHNSON JOHNSON CONTROLS INC JPMORGAN CHASE & CO KB HOME KELLOGG CO KIMBERLY-CLARK CORP KIMCO REALTY CORP KINDER MORGAN ENERGY PRNS KOHLS CORP KROGER CO XCEL ENERGY INC XEROX CORP XL CAPITAL LTD YUM! BRANDS INC VALERO ENERGY CORP VF CORP VORNADO REALTY LP VULCAN MATERIALS WAL-MART STORES INC WALT DISNEY CO/THE WEATHERFORD INTL.INC WELLPOINT INC WELLS FARGO & CO WEYERHAEUSER CO WHIRLPOOL CORP WILLIAMS COS INC/THE WISCONSIN ENERGY CORP WNGRTN.REALTY INVRS. TARGET CORP TECO ENERGY INC DVN DO D DOV DOW DHI DD DUK DRE EMN ETN LLY EMR ETR EPD RE EXC XOM POM PEP MET MTG MHK JCP JNJ JCI JPM KBH K KMB KIM KMP KSS KR XEL XRX XL YUM VLO VFC VNO VMC WMT DIS WFT WLP WFC WY WHR WMB WEC WRI TGT TE OFFICE DEPOT INC OLIN CORP OMNICOM GROUP ONEOK INC OWENS-ILLINOIS INC LENNAR CORP LINCOLN NATIONAL CORP LOCKHEED MARTIN CORP LOUISIANA-PACIFIC CORP MARATHON OIL CORP MARSH & MCLENNAN COS INC MARTIN MARIETTA MATS INC MASCO CORP MBIA INC MCDONALD’S CORP MCKESSON CORP MDC HOLDINGS INC MEADWESTVACO CORP MEDTRONIC INC MERCK & CO INC TENET HEALTHCARE CORP TESORO CORPORATION TEXTRON INC TIME WARNER INC TJX COMPANIES INC/ THE TOLL BROTHERS INC TRANSOCEAN INC TYSON FOODS INC UNION PACIFIC CORP UNISYS CORP UNITED PARCEL SER. INC UNITED RENTALS INC UNITED TECHNOLOGIES CORP UNITEDHEALTH GROUP INC UNIVERSAL HEALTH SVS INC UNUM GROUP RADIAN GROUP INC RADIOSHACK CORP RAYTHEON CO REPUBLIC SERVICES INC RITE AID CORP ROYAL CRBN.CRUISES LTD RPM INTERNATIONAL INC RYDER SYSTEM INC RYLAND GROUP INC SAFEWAY INC SEALED AIR CORP SEMPRA ENERGY SHERWIN-WILLIAMS CO SIMON PROPERTY GROUP INC SOUTHWEST AIRLINES STANDARD PACIFIC CORP STARWOOD HTLS.& RSTS WWD SUPERVALU INC ODP OLN OMC OKE OI LEN LNC LMT LPX MRO MMC MLM MAS MBI MCD MCK MDC MWV MDT MRK THC TSO TXT TWX TJX TOL RIG TSN UNP UIS UPS URI UTX UNH UHS UNM RDN RSH RTN RSG RAD RCL RPM R RYL SWY SEE SRE SHW SPG LUV SPF HOT SVU 38 References ACHARYA , V. V., S. T. B HARATH , AND A. S RINIVASAN (2007): “Does Industry wide Distress Affect Defaulted Firms? Evidence from Creditor Recoveries,” Journal of Financial Economics, 85, 787–821. A LEXANDER , C. AND A. K AECK (2008): “Regime dependent determinants of credit default swap spreads,” Journal of Banking an Finance, 32, 1008–1021. A NTWEILER , W. AND M. Z. F RANK (2004): “Is all that talk just noise? The information content of Internet stock message boards,” Journal of Finance, 59, 1259–1294. BAKER , M. AND J. C. S TEIN (2004): “Market liquidity as a sentiment indicator,” Journal of Financial Markets, 7, 271–299. BAKER , M. AND J. W URGLER (2006): “Investor Sentiment and the Cross-Section of Stock Returns,” Journal of Finance, 61, 1645–1680. ——— (2007): “Investor Sentiment in the Stock Market,” Journal of Economic Perspectives, 21, 129–151. BARBER , B. M. AND T. O DEAN (2008): “All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors,” Review of Financial Studies, 21, 785–818. B LUME , M. AND D. K EIM (1991): “Realized returns and volatility of low-grade bonds: 19771989,” Journal of Finance, 46, 49–74. B OLLEN , N. P. AND W. R. E. (2004): “Does Net Buying Pressure Affect the Shape of Implied Volatility Functions?” Journal of Finance, 59, 711–753. B ONGAERTS , D., F. DE J ONG , AND J. D RIESSEN (2011): “Derivative pricing with liquidity risk: theory and evidence from the credit default swap market,” Journal of Finance, 66, 203–240. C AMPBELL , J. AND G. TAKSLER (2003): “Equity Volatility and Corporate Bond Yields,” Journal of Finance, 58, 2321–2349. C HAVA , S., C. S TEFANESCU , AND S. T URNBULL (2011): “Modeling the loss distribution,” Management Science, 57, 1267–1287. C HOI , H. AND H. VARIAN (2011): “Predicting the present with Google Trends,” Working Paper. C HRISTIE , A. A. (1982): “The stochastic behavior of common stock variances. Value, leverage and interest rate effects,” Journal of Financial Economics, 10, 407–432. C HRISTOFFERSEN , P., V. E RRUNZA , K. JACOBS , AND H. L ANGLOIS (2012): “Is the Potential for International Diversification Disappearing? A Dynamic Copula Approach,” Review of Financial Studies, 25, 3711–3751. 39 C HRISTOFFERSEN , P., K. JACOBS , X. J IN , in Corporate Credit,” Working Paper. AND H. L ANGLOIS (2013): “Dynamic Dependence C OLLIN -D UFRESNE , P., R. S. G OLDSTEIN , AND J. S. M ARTIN (2001): “The Determinants of Credit Spread Changes,” Journal of Finance, 56(6), 2177–2207. C REMERS , M., J. D RIESSEN , P. M AENHOUT, AND D. W EINBAUM (2008): “Individual StockOption Prices and Credit Spreads,” Journal of Banking and Finance, 32(12), 2706–2715. C UTLER , D. M., J. M. P OTERBA , AND L. H. S UMMERS (1989): “What moves stock prices?” Journal of Portfolio Management, 15, 4–12. DA , Z., J. E NGELBERG , 1461–1499. AND P. G AO (2011): “In Search of Attention,” Journal of Finance, 66, ——— (2015): “The Sum of All FEARS: Investor Sentiment and Asset Prices,” Review of Financial Studies, forthcoming. DANG , T. L., F. M OSHIRIAN , AND B. Z HANG (forthcoming): “Commonality in News around the World,” Journal of Financial Economics. D E L ONG , J. B., A. S HLEIFER , L. H. S UMMERS , AND R. J. WALDMANN (1990): “Noise Trader Risk in Financial Markets,” Journal of Political Economy, 98, 703–738. D ING , R. AND W. H OU (2013): “Retail Investor Attention and Stock Liquidity,” Working Paper. D OSHI , H., J. E RICSSON , K. JACOBS , AND S. M. T URNBULL (2013): “Pricing Credit Default Swaps with Observable Covariates,” Review of Financial Studies, 26, 2049–2094. D UFFIE , D. AND D. L ANDO (2000): “Term Structure of Credit Spreads with Incomplete Accounting Information,” Econometrica, 69, 633–664. D UFFIE , D., L. S AITA , AND K. WANG (2007): “Multi-period corporate default prediction with stochastic covariates,” Journal of Financial Economics, 83, 635–665. D UFFIE , D. AND K. J. S INGLETON (1997): “An Econometric Model of the Term Structure of Interest Rate Swap Yields,” Journal of Finance, 52, 1287–1323. ——— (1999): “Modeling Term Structures of Defaultable Bonds,” Review of Financial Studies, 12, 687–720. E RICSSON , J., K. JACOBS , AND R. OVIEDO (2009): “The Determinants of Credit Default Swap Premia,” Journal of Financial and Quantitative Analysis, 44(1), 109–132. F IGLEWSKI , S. (1989): “Option Arbitrage in Imperfect Markets,” Journal of Finance, 44, 1289– 1311. F IGLEWSKI , S. AND T. C. G REEN (1999): “Market Risk and Model Risk for a Financial Institution Writing Options,” Journal of Finance, 54, 1465–1499. 40 G ALIL , K., O. M. S HAPIR , D. A MIRAM , AND U. B EN -Z ION (2014): “The determinants of CDS spreads,” Journal of Banking and Finance, 41, 271 – 282. G INSBERG , J., M. M. M OHEBBI , R. S. PATEL , L. B RAMMER , M. S. S MOLINSKI , AND L. B RILLIANT (2009): “Detecting influenza epidemics using search engine query data,” Nature, 457, 1012–1015. H AMID , A. AND M. H EIDEN (2014): “Forecasting Volatility with Empirical Similarity and Google Trends,” Working paper. H AN , B. (2008): “Investor Sentiment and Option Prices,” Review of Financial Studies, 21, 387– 414. H ILLEGEIST, S., E. K EATING , D. C RAM , AND K. L UNDSTEDT (2004): “Assessing the Probability of Bankruptcy,” Review of Accounting Studies, 9, 5–34. H ILSCHER , J., J. P OLLET, AND M. W ILSON (2015): “Are Credit Default Swaps a Sideshow? Evidence that Information Flows from Equity to CDS Markets,” Journal of Financial and Quantitative Analysis, 50, 543–567. H ITWISE (2009): “Google Hovering at 72 Percent of U.S. Searches in March 2009,” (March 26, 2014). J UNGE , B. AND A. B. T ROLLE (2015): “Liquidity Risk in Credit Default Swap Markets,” Working paper. K AHNEMAN , D. (1973): Attention and Effort, Prentice Hall. KOGAN , L., S. A. ROSS , J. WANG , AND M. M. W ESTERFIELD (2006): “The Price Impact and Survival of Irrational Traders,” Journal of Finance, 61, 195–229. ——— (2009): “Market Selection,” Working Paper MIT. K RISTOUFEK , L. (2013): “Can Google Trends search queries contribute to risk diversification?” Scientific Reports, 3, 2713. L ELAND , H. AND K. TOFT (1996): “Optimal Capital Structure, Endogenous Bankruptcy and the Term Structure of Credit Spreads,” Journal of Finance, 51, 987–1019. L ESPLINGART, C., C. M AJOIS , AND M. P ETITJEAN (2012): “Liquidity and CDS premiums on European companies around the Subprime crisis,” Review of Derivatives Research, 15, 257–281. L ONGSTAFF , F. AND E. S CHWARTZ (1995): “A Simple Approach to Valuing Risky Fixed and Floating Debt,” Journal of Finance, 50, 789–819. L ONGSTAFF , F. A., S. M ITHAL , AND E. N EIS (2005): “Corporate Yield Spreads: Default Risk Or Liquidity? New Evidence From The Credit Default Swap Market,” Journal of Finance, 60(5), 2213–2253. 41 M AHANI , R. S. AND A. M. P OTESHMAN (2008): “Overreaction to Stock Market News and Misevaluation of Stock Prices by Unsophisticated Investors: Evidence from the Option Market,” Journal of Empirical Finance, 15, 635–655. M AYORDOMO , S., J. I. P E ÑA , AND E. S. S CHWARTZ (2014): “Are All Credit Default Swap Databases Equal?” European Financial Management, 20, 677–713. M C D ONALD , C. AND L. VAN DE G UCHT (1999): “High-yield bond default and call risks,” Review of Economics and Statistics, 81, 409–419. M EINE , C., H. S UPPER , AND G. W EISS (2015a): “Do CDS spreads move with commonality in liquidity?” Review of Derivatives Research, forthcoming. ——— (2015b): “Is Tail Risk Priced in Credit Default Swap Premia?” Review of Finance, forthcoming. M ENGLE , D. (2007): “Credit derivatives: An overview.” Working paper. M ERTON , R. C. (1974): “On the pricing of corporate debt: The risk structure of interest rates,” Journal of Finance, 29, 449–479. O FEK , E., M. R ICHARDSON , AND R. F. W HITELAW (2004): “Limited Arbitrage and Short Sales Restrictions: Evidence from the Options Markets,” Journal of Financial Economics, 74, 305– 342. P ENG , L. AND W. X IONG (2006): “Investor attention, overconfidence and category learning,” Journal of Financial Economics, 80, 563–602. P ESARAN , M., T. S CHUERMANN , B.-J. T REUTLER , AND S. W EINER (2006): “Macroeconomic dynamics and credit risk: a global perspective,” Journal of Money, Credit, and Banking, 38, 1211–1262. P OTESHMAN , A. M. (2001): “Underreaction, Overreaction, and Increasing Misreaction to Information in the Options Market,” Journal of Finance, 56, 851–876. P OTESHMAN , A. M. AND V. S ERBIN (2003): “Clearly Irrational Financial Market Behavior: Evidence from the Early Exercise of Exchange Traded Stock Options,” Journal of Finance, 58, 37–70. P REIS , T., H. S. M OAT, AND H. E. S TANLEY (2013): “Quantifying Trading Behavior in Financial Markets Using Google Trends,” Scientific Reports, 3, 1648. S HLEIFER , A. AND R. W. V ISHNY (1992): “Liquidation Values and Debt Capacity: A Market Equilibrium Approach,” Journal of Finance, 47, 1343–1366. ——— (1997): “The limits of arbitrage,” Journal of Finance, 52, 33–55. S HUMWAY, T. (2001): “Forecasting bankruptcy more accurately: A simple hazard model,” Journal of Business, 74, 101–124. 42 S IMS , C. A. (2003): “Implications of rational inattention,” Journal of Monetary Economics, 50, 665–690. S INGER , E. (2002): “The Use of Incentives to Reduce Nonresponse in Household Surveys,” Working Paper, University of Michigan. S TEIN , J. (1989): “Overreactions in the Options Market,” Journal of Finance, 44, 1011–1022. TANG , D. Y. AND H. YAN (2013): “What moves CDS spreads?” Working Paper. ——— (2014): “Liquidity and Credit Default Swap Spreads,” Working Paper. T ETLOCK , P. C. (2007): “Giving Content to Investor Sentiment: The Role of Media in the Stock Market,” Journal of Finance, 62, 1139–1168. T ETLOCK , P. C., M. S AAR -T SECHANSKY, AND S. M ACSKASSY (2008): “More Than Words: Quantifying Language to Measure Firms Fundamentals,” Journal of Finance, 63, 1437–1467. Z HANG , B., H. Z HOU , AND H. Z HU (2009): “Explaining Credit Default Swap Spreads with the Equity Volatility and Jump Risks of Individual Firms,” Review of Financial Studies, 22, 5099– 5131. 43 Figures 44 Figure 1: Time evolution of CDS and bid-ask spreads The figure shows the time evolution of CDS spreads, CDS bid-ask-spreads, and the Default Sentiment Index (DSI) from 2004 to Q3:2010. The grey-shaded areas present the 20%- and 80%quantile of CDS spreads (upper panel) and CDS bid-ask-spreads (lower panel). Time evolution of cds spreads 0 400 0 −2 −1 200 CDS spread (bps) 1 600 2 20−80% quantile Mean DSI 2004 2005 2006 2007 2008 2009 2010 2 20−80% quantile Mean DSI 1 −1 0 40 −2 20 0 Bid−ask spread (bps) 60 80 Time evolution of bid−ask spreads 2004 2005 2006 2007 45 2008 2009 2010 0.20 0.10 0.00 0.30 R2 R2 2004 2006 2008 Ind. Goods and Services (32 firms) 2010 2004 2006 2008 Healthcare (17 firms) 2010 2004 2004 2008 2006 2008 Basic Resources (6 firms) 2006 Technology (6 firms) 2010 2010 2006 2008 2010 2004 2006 2008 2010 Personal and Household Goods (21 firms) 2004 Utilities (18 firms) 0.10 0.00 0.20 0.10 0.00 0.30 0.10 0.20 0.30 0.10 R2 0.00 0.20 0.30 2004 2004 2004 2008 2008 2006 2008 Automobiles and Parts (6 firms) 2006 Chemicals (11 firms) 2006 Insurance (13 firms) 2010 2010 2010 Sector specific R2 s are the average of all sample firms’ R 2 within a sector. The Telecommunications sector is omitted since it only contains one sample firm. The figure shows the time evolution of the annual commonality of the Default Sentiment Index and the Bid-Ask-Spreads of CDS contracts for the years from 2004 to 2010 per sector. The R 2 is obtained from running OLS regressions of a CDS contract Bid-Ask-Spread on the Default Sentiment Index for each year (and from January to September for 2010): BASit = α + β × Default Sentimentt + εit . Figure 2: Default Sentiment Commonality R2 R2 0.30 0.20 0.10 0.20 0.30 0.00 R2 0.00 R2 R2 R2 0.30 0.20 0.10 0.00 0.30 0.20 0.10 0.00 0.30 0.20 0.10 0.00 46 R2 R2 0.30 0.20 0.10 0.00 0.30 2004 2004 2004 2008 2008 2006 2008 Media (6 firms) 2006 Oil and Gas (20 firms) 2006 Financial Services (5 firms) 2010 2010 2004 2004 2008 2010 2004 2008 2006 2008 Travel and Leisure (9 firms) 2006 Food and Beverage (11 firms) 2006 Real Estate (17 firms) 2010 2010 2010 0.10 0.00 0.20 0.30 0.10 0.20 0.10 0.30 R2 0.20 0.00 Figure 2: Default Sentiment Commonality (continued) R2 R2 0.30 0.20 0.10 0.00 0.30 0.00 0.10 0.20 R2 0.00 R2 R2 R2 0.30 0.20 0.10 0.00 0.30 0.20 0.10 0.00 0.30 0.20 0.10 0.00 47 2008 2006 2008 Banks (4 firms) 2006 2010 2010 2004 2006 2008 2010 Construction and Material (5 firms) 2004 2004 Retail (20 firms) 2000 1500 2004 2005 2006 1st quintile 2nd quintile 3rd quintile 4th quintile 5th quintile 2007 2008 2009 2010 2004 2005 2006 1st quintile 2nd quintile 3rd quintile 4th quintile 5th quintile 2007 2008 2009 2010 Mean CDS spreads (LIQ−CSI betas) 500 0 CDS spread (bps) Mean CDS spreads (LIQ−MLIQ betas) CDS spread (bps) 500 1000 1500 2000 2004 2005 2006 1st quintile 2nd quintile 3rd quintile 4th quintile 5th quintile 2007 2008 2009 2010 Mean CDS spreads (LIQ−MCDS betas) where MLIQt is the average value of all individual bid-ask-spreads (market illiquidity), DSI t is the Default Sentiment Index, and MCDSt is the average CDS spreads in week t. For each reference entity i, we run simple OLS regressions to obtain the coefficient estimates i i i , βLIQ-DSI , and βLIQ-MCDS . The sample is split according to the quintiles of the respective beta estimate. βLIQ-MLIQ i × MCDSt + εit , BASit = α + βLIQ-MCDS i BASit = α + βLIQ-DSI × DSIt + εit , i BASit = α + βLIQ-MLIQ × MLIQt + εit , The figure shows the time evolution of average CDS spreads for firms in different quintiles of sensitivity of bid-ask-spreads. The sensitivity is measured as the OLS coefficient estimate of the following models: Figure 3: Time evolution CDS spreads in quintiles of DSI betas CDS spread (bps) 0 1000 2000 1500 1000 500 0 48 Tables 49 50 Avrg. Observations Adj. R2 Constant ΔTail beta ΔSlope ΔS&P500 ΔVIX ΔDSI ΔVolatility ΔInterest Rate ΔFirm value 348.0 0.04 0.312 (11.23)*** -45.788 (-10.02)*** (1) 350.0 0.02 0.191 (8.02)*** -31.939 (-9.76)*** (2) 346.6 0.01 0.288 (8.60)*** 8.341 (3.74)*** (3) 346.6 0.08 0.198 (7.32)*** -42.961 (-9.81)*** -26.912 (-9.58)*** 9.029 (4.20)*** (4) 350.0 0.00 0.308 (10.90)*** -1.069 (-2.76)*** (5) 346.6 0.08 0.200 (6.91)*** -43.013 (-9.85)*** -26.812 (-9.55)*** 8.981 (4.15)*** -0.813 (-1.98)** (6) 346.6 0.15 0.179 (6.32)*** -23.598 (-6.14)*** -20.895 (-8.93)*** 7.011 (3.18)*** -0.819 (-1.96)** 2.275 (9.65)*** (7) 346.6 0.18 0.240 (6.76)*** -0.312 (-8.89)*** -23.920 (-6.32)*** -16.720 (-8.62)*** 6.994 (3.23)*** -0.615 (-1.49)* (8) 346.6 0.08 0.166 (5.61)*** -19.934 (-3.29)*** -41.013 (-9.70)*** -31.981 (-8.25)*** 8.521 (3.99)*** -0.826 (-2.02)*** (9) 345.6 0.08 368.831 (1.86)** 0.193 (6.32)*** - -42.688 (-9.88)*** -26.318 (-9.52)*** 8.771 (4.04)*** -0.714 (-1.90)** (10) The main independent variable of interest is Default Sentiment Index (DSI) which is the average Google Search Volume Index of thirty search phrases from Google Trends, adjusted for seasonality, winsorized at the 5% level, and scaled by its standard deviation. Other regressors are defined in Appendix I. Values of corresponding t-statistics are calculated as outlined in Collin-Dufresne et al. (2001) and Ericsson et al. (2009) and are given below the coefficients. *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level, respectively. ΔCDSt = α + β × ΔDefault Sentimentt + γ × ΔMerton-typet + Θ × ΔControlst + εt . The table shows the results of simple OLS time-series regressions in first differences of CDS mid quotes on first differences of the Default Sentiment Index (DSI) and first differences of control variables. All regressions are of the following type: Table I: Benchmark time-series regressions in first differences. 51 Firm FE Year dummies Observations Adj. R2 Constant ΔTail beta ΔSlope ΔS&P500 ΔVIX ΔDSI ΔVolatility ΔInterest Rate ΔFirm value x x 79,334 0.011 -0.314 (-4.66)*** -73.380 (-6.86)*** (1) x x 79,792 0.004 0.602 (5.96)*** -29.995 (-9.45)*** (2) x x 79,028 0.002 -0.272 (-4.11)*** 5.944 (1.48) (3) x x 79,792 0.002 -0.325 (-4.07)*** -1.061 (-2.74)*** (4) x x 79,028 0.013 0.426 (4.90)*** -70.675 (-6.56)*** -24.727 (-9.10)*** 4.196 (0.93) -1.012 (-2.59)*** (5) x x 79,028 0.020 0.378 (4.52)*** -53.521 (-4.99)*** -19.255 (-8.47)*** 3.784 (0.87) -1.026 (-2.62)*** 2.085 (9.26)*** (6) x x 79,028 0.024 0.468 (5.24)*** -0.291 (-8.76)*** -53.355 (-5.18)*** -15.764 (-8.20)*** 3.563 (0.83) -0.806 (-2.08)** (7) x x 79,028 0.013 0.135 (1.30) -16.450 (-2.98)*** -68.759 (-6.60)*** -28.772 (-7.99)*** 3.961 (0.89) -1.025 (-2.63)*** (8) x x 78,795 0.013 23.467 (1.83)* 0.426 (4.99)*** -70.483 (-6.53)*** -24.426 (-9.03)*** 4.205 (0.93) -1.002 (-2.58)*** (9) where αi and μt are firm-fixed effects and year dummies, respectively. The main independent variable of interest is Default Sentiment Index (DSI) which is the average Google Search Volume Index of thirty search phrases from Google Trends, adjusted for seasonality, winsorized at the 5% level, and scaled by its standard deviation. Other regressors are defined in Appendix I. Values of corresponding t-statistics are corrected for clustering on the firm level and are given below the coefficients and *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level, respectively. ΔCDSit = αi + μt + β × ΔDefault Sentimentt + Θ × ΔControlsit + εit , The table shows the results of OLS panel regressions in first differences of CDS mid quotes on first differences of the Default Sentiment Index (DSI) and first differences of control variables. All regressions are of the following type: Table II: Benchmark panel regressions in first differences. 52 Firm FE Year dummies Observations Adj. R2 Constant ΔVolatility ΔInterest Rate ΔFirm value ΔCDSILLIQ ΔBASM abs ΔBASIabs ΔBASabs ΔDSI x 79,028 0.24 -73.250 (-5.75)*** -23.690 (-10.83)*** 11.073 (1.73)* 0.214 (26.88)*** -0.643 (-1.71)* 1.734 (10.30)*** (1) x 78,678 0.03 -70.602 (-6.61)*** -25.387 (-9.65)*** 6.154 (1.34) 0.215 (20.64)*** 0.941 (2.67)*** -0.814 (-2.06)*** (2) x 79,028 0.02 -72.825 (-6.61)*** -22.615 (-9.81)*** 6.244 (1.35) 0.210 (19.32)*** 2.425 (5.25)*** -0.514 (-1.36) (3) x 47,600 0.02 1.369 (7.18)*** -83.489 (-6.71)*** -35.048 (-9.04)*** 3.225 (0.64) -0.152 (-1.90)** -1.944 (-1.87)** (4) x x 79,028 0.24 -73.813 (-5.78)*** -21.696 (-10.44)*** 10.677 (1.67)* 0.362 (5.13)*** -0.648 (-1.72)* 1.732 (10.29)*** (5) x x 78,678 0.03 -71.201 (-6.65)*** -23.278 (-9.17)*** 5.726 (1.26) 0.389 (4.64)*** 0.934 (2.66)*** -0.821 (-2.08)*** (6) x x 79,028 0.02 -73.307 (-6.64)*** -20.792 (-9.22)*** 5.831 (1.27) 0.331 (4.17)*** 2.368 (5.18)*** -0.530 (-1.40) (7) x x 47,600 0.02 1.249 (6.91)*** -83.443 (-6.71)*** -32.687 (-8.76)*** 3.033 (0.60) -0.650 (-5.18)*** -1.926 (-1.85)** (8) where αi and μt are firm-fixed effects and year dummies, respectively. The main independent variable of interest is Default Sentiment Index (DSI) which is the average Google Search Volume Index of thirty search phrases from Google Trends, adjusted for seasonality, winsorized at the 5% level, and scaled by its standard deviation. The liquidity proxies include a contracts Bid-Ask-Spread BAS abs , which is the end-of-week absolute CDS bid-ask spread, calculated as daily CDS ask quote minus CDS bid price and denoted in bps, the Sector Bid-Ask-Spread BASIabs calculated as the cross-sectional average of absolute bid-ask spreads of all firms within an industry, and the Market Bid-Ask-Spread BASM abs calculated as the cross-sectional average of relative bid-ask spreads of all sample firms excluding the respective firm. CDSILLIQ is the end-of-week value of the CDSILLIQ measure proposed by Junge and Trolle (2015). Other regressors are defined in Appendix I. Columns (1)-(4) report results from regressions including firm-fixed effects while the results in (5)-(8) also include year dummies in the analyses. Values of corresponding t-statistics are corrected for clustering on the firm level and are given below the coefficients and *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level, respectively. ΔCDSit = αi + μt + β × ΔDefault Sentimentt + γ × ΔLiquidityit + Θ × ΔControlsit + εit , The table shows the results of OLS panel regressions in first differences of CDS mid quotes on first differences of the Default Sentiment Index (DSI) and first differences of liquidity proxies. All regressions are of the following type: Table III: Panel regressions of CDS spreads on liquidity. Table IV: Panel regressions of CDS spreads on liquidity residuals. The table shows the results of OLS panel regressions in first differences of CDS mid quotes on first differences of the Default Sentiment Index (DSI) and first differences of liquidity proxies. All regressions are of the following type: ΔCDSit = αi + μt + β × ΔDefault Sentimentt + γ × ΔLiquidityit + Θ × ΔControlsit + εit , where αi and μt are firm-fixed effects and year dummies, respectively. The main independent variable of interest is Default Sentiment Index (DSI) which is the average Google Search Volume Index of thirty search phrases from Google Trends, adjusted for seasonality, winsorized at the 5% level, and scaled by its standard deviation. The liquidity proxies are the residuals of regressions in I first differences of the original liquidity variables BASabs , BASM abs , and BASabs on first differences of the Default Sentiment Index. Other regressors are defined in Appendix I. Values of corresponding t-statistics are corrected for clustering on the firm level and are given below the coefficients and *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level, respectively. ΔDSI ΔBAS-ΔDSI residual (1) (2) (3) (4) (5) (6) -1.061 (-2.73)*** 1.780 (10.25)*** -1.065 (-2.73)*** -1.061 (-2.74)*** -1.031 (-2.67)*** 1.732 (10.26)*** -1.022 (-2.60)*** -1.033 (-2.64)*** ΔBASI -ΔDSI residual 0.939 (2.66)*** ΔBASM -ΔDSI residual 0.938 (2.66)*** 2.326 (5.14)*** ΔFirm value -0.278 (-3.78)*** -0.308 (-3.84)*** -0.276 (-3.47)*** -73.822 (-5.78)*** -21.693 (-10.44)*** 10.685 (1.67)* 0.376 (5.33)*** x x 79,792 0.23 x x 79,442 0.02 x x 79,792 0.01 x x 79,028 0.24 ΔInterest Rate ΔVolatility Constant Firm FE Year dummies Observations Adj. R2 53 -71.214 (-6.65)*** -23.271 (-9.17)*** 5.736 (1.26) 0.397 (4.71)*** 2.369 (5.17)*** -73.307 (-6.64)*** -20.791 (-9.22)*** 5.831 (1.27) 0.353 (4.36)*** x x 78,678 0.03 x x 79,028 0.02 Table V: Panel regressions of CDS spreads on interaction terms. The table shows the results of OLS panel regressions in first differences of CDS mid quotes on first differences of the Default Sentiment Index (DSI) and first differences of liquidity proxies and their interaction. All regressions are of the following type: ΔCDSit = αi + μt + β × ΔDSIt + γ × ΔLiquidityit + δ × ΔLiquidityit × ΔDSI + Θ × ΔControlsit + εit , where αi and μt are firm-fixed effects and year dummies, respectively. The main independent variable of interest is Default Sentiment Index (DSI) which is the average Google Search Volume Index of thirty search phrases from Google Trends, adjusted for seasonality, winsorized at the 5% level, and scaled by its standard deviation. The liquidity proxies include a contracts Bid-Ask-Spread BASabs , which is the end-of-week absolute CDS bid-ask spread, calculated as daily CDS ask quote minus CDS bid price and denoted in bps, and the Market Bid-Ask-Spread BASM abs calculated as the cross-sectional average of relative bid-ask spreads of all sample firms excluding the respective firm. ΔLiquidityit × ΔDefault Sentiment is the interaction term of the DSI and the respective liquidity proxy. Other regressors are defined in AppendixI. Values of corresponding t-statistics are corrected for clustering on the firm level and are given below the coefficients and *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level, respectively. ΔDSI ΔBASabs ΔDSI × ΔBASabs ΔBASM abs (1) -0.574 (-1.33) 1.790∗∗∗ (10.88) -0.578 ∗ (-1.84) (2) -0.365 (-0.84) 1.729∗∗∗ (11.22) -0.619 ∗∗ (-2.04) ΔDSI × ΔBASM abs ΔFirm value ΔInterest rate ΔVolatiltiy ΔVIX ΔS&P 500 ΔSlope -56.490 ∗∗∗ (-4.84) -16.980 ∗∗∗ (-8.39) 8.290 (1.60) -0.151 (-0.53) -0.260 ∗∗∗ (-6.49) -15.550 ∗∗∗ (-3.81) 0.068 (0.70) x x 79,028 0.26 -0.325 ∗∗∗ (-4.02) Firm FE x Year dummies x N 79,792 0.23 Adj. R2 t statistics in parentheses ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 Constant 54 (3) 0.093 (0.24) (4) 0.163 (0.41) 2.389∗∗∗ (5.25) -3.563 ∗∗∗ (-10.53) 1.997∗∗∗ (4.57) -3.142 ∗∗∗ (-9.43) -53.320 ∗∗∗ (-4.84) -10.930 ∗∗∗ (-5.17) 3.390 (0.82) -0.561 ∗∗ (-2.25) -0.302 ∗∗∗ (-7.28) -8.070 (-1.53) -0.046 (-0.45) x x 79,028 0.03 -0.486 ∗∗∗ (-5.62) x x 79,792 0.02 Table VI: Panel regressions of CDS spreads on DSI residuals. The table shows the results of OLS panel regressions in first differences of CDS mid quotes on first differences of the residuals of the Default Sentiment Index (DSI) and first differences of liquidity proxies. All regressions are of the following type: ΔCDSit = αi + μt + β × ΔDSI residualt + γ × ΔLiquidityit + Θ × ΔControlsit + εit , where αi and μt are firm-fixed effects and year dummies, respectively. The Default Sentiment Index (DSI) is the average Google Search Volume Index of thirty search phrases from Google Trends, adjusted for seasonality, winsorized at the 5% level, and scaled by its standard deviation. Its residuals are obtained from regressions in first differences of the original DSI variable on the I three liquidity proxies BASabs , BASM abs , and BASabs .Control variables are first differences of the Firm value, Interest rate, Volatility, VIX, S&P 500, Slope and Tail beta, which are all defined in Appendix I. Values of corresponding t-statistics are corrected for clustering on the firm level and are given below the coefficients and *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level, respectively. ΔBAS ΔDSI-ΔBAS residual (1) 1.781∗∗∗ (10.29) -0.735 ∗ (-1.96) ΔMBAS ΔDSI-ΔMBAS residual (2) (3) 2.332∗∗∗ (5.16) -0.532 (-1.43) ΔSBAS ΔDSI-ΔSBAS residual Constant -0.296 ∗∗∗ (-3.99) x x 79,792 0.23 Firm FE Year dummies Controls N Adj. R2 t statistics in parentheses ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 -0.300 ∗∗∗ (-3.76) x x 79,792 0.01 55 0.936∗∗∗ (2.67) -0.693 ∗ (-1.79) -0.319 ∗∗∗ (-3.98) x x 79,442 0.02 (4) 1.722∗∗∗ (10.22) -0.526 (-1.43) 0.085 (1.00) x x x 78,795 0.25 (5) 1.941∗∗∗ (4.47) -0.421 (-1.10) 0.067 (0.66) x x x 78,795 0.03 (6) 0.883∗∗ (2.54) -0.582 (-1.49) 0.094 (0.93) x x x 78,446 0.04 Table VII: Panel regressions of CDS spreads on retail investor attention and household sentiment. The table shows the results of OLS panel regressions in first differences of CDS mid quotes on first differences of the Default Sentiment Index (DSI) and first differences of proxies for retail investor attention (Panel A) and household sentiment (Panel B). All regressions are of the following type: ΔCDSit = αi + μt + β × ΔDSIt + γ × ΔASVIit + δ × ΔFEARSit + Θ × ΔControlsit + εit , where αi and μt are firm-fixed effects and year dummies, respectively. The main independent variable of interest is Default Sentiment Index (DSI) which is the average Google Search Volume Index (GSVI) of thirty search phrases from Google Trends, adjusted for seasonality, winsorized at the 5% level, and scaled by its standard deviation. ASVIit is the abnormal search volume on a firm’s ticker symbol in Google Trends as introduced in Da et al. (2011). FEARS is the household sentiment index proposed by Da et al. (2015) based on the GSVI of thirty economics-related search phrases. The liquidity proxies include a contracts Bid-Ask-Spread BASabs , which is the end-of-week absolute CDS bid-ask spread, calculated as daily CDS ask quote minus CDS bid price and denoted in bps, the Sector Bid-Ask-Spread BASIabs calculated as the cross-sectional average of absolute bid-ask spreads of all firms within an industry, and the Market Bid-Ask-Spread BASM abs calculated as the cross-sectional average of relative bid-ask spreads of all sample firms excluding the respective firm. CDSILLIQ is the end-of-week value of the CDSILLIQ measure proposed by Junge and Trolle (2015). Control variables are first differences of the firm value, interest rate and volatility, which are defined in AppendixI. Values of corresponding t-statistics are corrected for clustering on the firm level and are given below the coefficients and *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level, respectively. Panel A: ASVI ΔDSI ΔASVI (1) -0.956∗ (-1.92) 0.538∗∗ (2.06) ΔASVI × ΔDSI (2) -0.957∗ (-1.93) 0.543∗ (1.93) 0.036 (0.19) ΔBASabs (3) -0.570 (-1.20) 0.718∗∗∗ (2.64) 0.046 (0.24) 1.887∗∗∗ (12.06) ΔBASIabs (4) -0.751 (-1.50) 0.549∗ (1.91) 0.071 (0.35) 0.941∗∗ (2.43) 0.453∗∗∗ (3.31) 59,476 0.012 0.453∗∗∗ (3.31) 59,476 0.012 0.361∗∗∗ (3.36) 59,476 0.243 0.411∗∗∗ (3.10) 59,134 0.028 2.606∗∗∗ (4.45) 0.347∗∗∗ (2.75) 59,476 0.023 (1) -1.118∗∗∗ (-2.90) 1.095∗∗ (2.16) (2) -0.769∗∗ (-2.09) 1.260∗∗∗ (2.73) 1.733∗∗∗ (10.29) (3) -0.937∗∗ (-2.39) 1.207∗∗ (2.45) (4) -0.659∗ (-1.77) 1.349∗∗∗ (2.66) (5) -2.724∗∗ (-2.55) 3.826∗∗∗ (4.63) ΔBASM abs Constant N Adj. R2 Panel B: FEARS ΔDSI ΔFEARS ΔBAS ΔSBAS 0.935∗∗∗ (2.66) ΔMBAS 2.372∗∗∗ (5.18) ΔCDSILLIQ Constant N Adj. R2 Firm FE Year dummies Controls (5) -0.408 (-0.85) 0.673∗∗ (2.34) 0.029 (0.15) 0.396∗∗∗ (4.60) 79,028 0.01 x x x 0.327∗∗∗ (4.67) 79,028 0.24 x x x t statistics in parentheses ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01 56 0.356∗∗∗ (4.35) 78,678 0.03 x x x 0.294∗∗∗ (3.77) 79,028 0.02 x x x 1.342∗∗∗ (7.30) -0.632∗∗∗ (-5.07) 47,600 0.02 x x x
© Copyright 2025 Paperzz