Short Selling And Corporate Bond Returns - 國立臺灣大學

effects that could also determine future bond returns. Particularly, we show that our findings are not a manifestation of bond return autocorrelation or stock-bond spillover effect. Additional robustness tests confirm that the predictability of bond returns using short selling is unlikely driven by illiquidity of our sample bonds. As we affirm the robustness of our results and rule out plausible explanations, the next logical and essential question is related to the nature of information in short selling. The negative relation between abnormal short selling and bond returns is consistent with two hypotheses. The first is that short sellers target firms with negative future prospect, and their trades deliver credible signals to both equity and bond markets. This proposition suggests that an increase in shorting activity is followed by weak firm performance. We utilize various market and accounting metrics to test this hypothesis. The second is that short sellers identify overvalued firms, and they realize their profit once share prices move more closely to the firms’ intrinsic values. It is important to note that both information signaling and overvaluation accord with the negative association between short selling and stock returns, and the former can also explain the relation between short selling and bond returns. However, to apply the second hypothesis to explain bond price reactions, we require overvaluation on both stock and bond. Although existing studies provide ample evidence of overvaluation in the stock market, there are few empirical studies suggesting the same in the corporate bond market.5 The task for this research is even more challenging as we need to identify common channels through which a firm’s stock and bond(s) are overvalued simultaneously. To test the overvaluation explanation, we recognize one such possible channel: earnings accruals. Ample studies have suggested that high (low) accrual stocks are relatively overvalued (undervalued)6, and Hirshleifer, Teoh, and Yu (2011) find that 5 Intuitively, a stock’s market value equals discounted future cash flows, and investors could be overoptimistic about the firm’s future cash flows. However, unlike stock, a bond’s value equals discounted pre-determined interests and principal. Thus, under normal conditions, its value is capped and unless the market yield is significantly underestimated, it is unusual for bonds to be overvalued. 6 See, e.g., Sloan (1996), Hirshleifer, Hou, Teoh, and Zhang (2004), Richardson, Sloan, Soliman, and Tuna (2005), and Chan, Chan, Jegadeesh, and Lakonishok (2006). 5

short sellers tend to target high accrual stocks. Bhojraj and Swaminathan (2009) further document that mispricing of accruals could occur in corporate bond market. Because of the evidence from existing papers, we test the overvaluation hypothesis using the accrual measures. In examining the sources of information in short selling, we find that firms with high abnormal short selling and low bond returns are associated with significantly higher credit risk and lower earnings per share and earnings surprises than are other firms in the subsequent quarters. Further, these firms are more likely to suffer from reductions in dividend payouts and shrinking total assets. In contrast, our results provide little support for the overvaluation explanation in the corporate bond market. Together, our findings are more consistent with the proposition that short sale activities provide credible information to equity and bond markets about the firms’ future prospects. As noted, we construct monthly abnormal short selling measures by employing daily short sale transactions rather than monthly net open short positions, which are short interest data reported monthly by major exchanges.7 Our dataset possesses a number of advantages.8 First, our measures are more precise in capturing persistent short selling during a month than short interest which could net out useful information given the evidence that short positions are usually short-lived (Geczy et al. (2002) and Boehmer et al. (2008)). Second, we are able to disaggregate short trades initiated by speculators (customers) versus dealers (market makers). As suggested in existing studies, speculative trades are typically motivated by information while dealers trade mostly for bona fide market making. Notwithstanding the benefits of our dataset, a legitimate concern is whether our results can be replicated by using short interest data, and whether our short selling measures provide valuable information beyond that of short interest. We conduct additional tests to investigate these issues. As expected, our short selling measures are highly 7 8 During our sample period, Nasdaq released short interest data once a month. As of September 2007, the frequency was changed to bimonthly. We also note that our dataset has a limitation: the short time period which severely weakens our likelihood of uncovering significant results. 6

correlated with short interest. More importantly, the multivariate analysis including both short interest and our abnormal short selling measures demonstrate that short interest has little predictive power for bond returns while the short selling measure remains statistically and economically significant. This evidence suggests that our short-selling measures, instead of the monthly short interest, successfully capture the information content of short-sale transactions. This research expands our understanding of short selling in several dimensions. First, it is the first study that uses daily short sale data to analyze the spillover effect of equity short selling in the corporate bond market. Previous research concentrates on the impact of short selling in the equity market. We provide evidence that short sale activities in the equity market are informative in the corporate bond market. Second, we further show that our short selling measures provide valuable information beyond that of monthly short interest data. For example, a recent paper by Kecskes, Mansi, and Zhang (2012) uses monthly short interest data in June of each year as a snapshot of bond pricing information and investigate the impact of large and sudden changes in short interest in June on the firm’s cost of debt. They document that a sudden positive (negative) spike in monthly short interest could signal a firm’s heightened (decreased) expected credit risk. Our findings corroborate and extend theirs by using a finer dataset and importantly, by examining more general correlations between short selling and bond returns. We show a significant and timely association between short selling and bond returns. Third, we provide important and novel results that abnormal short selling exhibits strong predictive power for future bond returns. Lastly, we disentangle the two commonly cited hypotheses explaining the negative relation between abnormal short selling and future equity returns and examine whether the theories offer support for the negative relation between short selling and future bond returns. The remainder of the paper is organized as follows. Section 2 describes the sample construction and summary statistics of the combined sample used in the study. Section 3 analyzes the relation between short selling and contemporaneous bond returns. Section 4 7

investigates the impact of short selling on future bond returns. Section 5 discusses extended results and robustness checks. Section 6 examines the information content in short selling and further distinguishes between the two hypotheses explaining the impact of short selling on security prices. Section 7 concludes. 2. Sample Construction and Descriptive Statistics 2.1. Sample of Short Selling The source of the short selling data is Nasdaq’s Automated Confirmation Transaction (ACT) Service, which processes the vast majority of transactions in Nasdaq-listed stocks during our sample period that spans from September 13, 2000 to July 10, 2001. The dataset includes, on a daily basis, aggregate long and short shares, closing price, and total trading volume for each stock. The aggregate shares contain all, except for small odd-lot trades, processed trades during the regular trading session (9:30am-4:00pm); thus, after hour trades are excluded. An important advantage of the ACT transaction data is that trades can be grouped based on trader identity: market makers vs. customers. Specifically, we use Nasdaq’s daily file of quotations to separate short transactions in our ACT database into dealer short vs. customer short. That file enables us to identify who, during each day of our sample period, actively posted bid and ask quotes in each stock (i.e., conducting market making activity).9 Accordingly, the short trades made by those traders are classified as dealer shorts. The rest of short trades that were not made by dealers acting as market makers on the day of the trade were classified as customer shorts. Another important distinction of our dataset is that it categorizes the trades based on whether they are designated as “exempt” vs. “non-exempt” of the Nasdaq Short Sale Rule.10 9 During the sample period, NASD rules required market makers to flag all of their trades as such (Revised NASD Rule 6130(d)(6)). Other trades, according to trade reporting rules and trade records, are classified as initiated by customers. 10 Nasdaq’s Short Sale Rule (Rule 3350) was analogous to the “uptick” rule for NYSE-listed securities in our sample period. The major difference was that Rule 3350 used a bid-test instead of the tick-test 8

Specifically, dealers can use the exemption designation to execute a trade at the prevailing inside bid following a down tick as long as the trade is claimed for the purpose of bona fide market making.11 As for customers, Nasdaq allows them to use the exempt designation for short sales related to certain activities including arbitrage positions on options, convertible bonds, or foreign markets and hedging of deliveries due in a few days. Because the exemption-marked short trades are for non-speculative activities and could well be the target of an eventual audit for potential abuse of the exemption, exempt-marked shorted shares by dealers and customers are generally considered non-informative shorting.12 By using the above two distinctions, daily short sales can be accurately partitioned into a two-by-two matrix: customers vs. dealers and exempt vs. non-exempt. Clearly, dealer exempt and non-exempt short trades are mostly driven by market making moves or by proprietary investing for either the dealer’s desk or other units within the firm, and customer exempt short trades are for aforementioned purposes, they are less likely to be informative. For the analysis conducted herein, we focus on non-exempt short trades by customers. 2.2. Sample of Firms and Stock Returns The initial firm sample consists of Nasdaq firms with share codes 10 or 11 in the CRSP/Compustat Merged Database that traded during the period 1999-2001. We collect accounting and stock-price related variables one year before short sales occurred to avoid lookahead bias. Firms are excluded with missing or non-positive observations for any of the following variables: book value of equity, book value of total assets, and total shares outstanding. In addition, we delete any stock that had an end-of-year 1999 price less than or equal to 1. applied by the NYSE. Generally, the rule prohibited short-selling at the bid if it was lower than the preceding bid. 11 Dealers were required to mark this distinction between regular short and short exempt, as described in Chapter 9 of the Nasdaq Trader Manual (revised 2000). 12 During our sample period, the average daily short exempt volume to total trading volume is only 0.09% for customers and 1.68% for dealers. 9

Finally, since we examine corporate bond returns, we delete firm-year observations with zero or missing long-term debt (Compustat items dd1 dltt). The firm sample after initial screenings contains 1,997 firms. We then delete 261 firms since they do not have short selling data. In addition, to avoid potential issues posed by thinly traded shares and to focus on stocks that are liquid enough to attract short sellers, we follow previous literature13 and restrict our analysis to those stocks satisfying two conditions: (1) the stock trades (long or short) every day during the sample period, and (2) the total number of short trades from customers is positive during the sample period.14 We remove 645 and 41 firms due to the first and second rules, respectively. We then use this filtered firm dataset (1050 firms) to collect corporate bond data. 2.3. Sample of Corporate Bonds Using the merged sample of Nasdaq firms with valid short sale data described above, we collect a sample of US corporate fixed-coupon bonds and exclude those that are convertible, putable, or asset-backed. Two major sources for bond data are Moody’s/Mergent Bond Record and Standard and Poor’s Bond Guide. For each Nasdaq firm in our sample, we hand collect monthly bond prices for each firm’s bonds that have the following information: issue and maturity dates, coupon rate, and issue size. Since Moody’s and S&P report valid end-of-month transaction prices or bid quotes, there are no matrix prices in our final dataset. The collected bond prices are reliable and robust for our analysis. In addition, we collect each bond’s Moody’s rating if available, and otherwise collect the Standard & Poor’s rating. Our final sample contains 156 unique bond issues from 86 firms. The overall distribution is as follows: (number of bonds: 13 14 For example, see Boehmer et al. (2008), Diether et al. (2009), and Christophe et al. (2010). Another commonly applied filter is to screen out stocks with a stock price less than 5 because short sellers are less likely to short those stocks (D’Avolio (2002)). We originally required firms with a stock price of at least 5 in the month of September 2000. It turns out that this constraint is not binding in our final sample of firms with available corporate bonds since firms with bonds outstanding are larger and have higher stock prices than other firms without bonds. 10

number of firms): (1: 58), (2: 12), (3: 6), (4: 6), (5: 1), (8: 2), and (11: 1). Consistent with other studies, the majority of our sample firms have only one bond outstanding. 2.4. Construction of Variables of Interest Two major variables of interest for this research are bond returns and equity short selling. Bond returns (BDRET) are calculated based on the full price which equals flat price plus accrued interest: BDRETi,t (BPi,t – BPi,t-1 AIi,t)/BPi,t-1, (1) where BP is bond price and AI is accrued interest in a month. To minimize potential impacts of extreme outliers on our results, we winsorize bond returns at 1% and 99% percentiles. We then follow the approach utilized in prior studies (e.g., Warga and Welch (1993), Billett et al. (2004), Wei and Yermack (2010), and Ambrose et al. (2011)) and compute the excess bond returns (BDABR): BDABRi,t BDRETi,t – INDEXi,t, (2) where INDEX is the index return on a matching Lehman Brothers US Corporate Bond Index and High Yield Index over the same month. Bond index returns are obtained from Datastream. There are eleven Standard and Poor’s bond rating categories (AAA, AA, A, BBB, BB, B, CCC, CC, C, D, and NR) in each of the two maturity categories (long-term and intermediate-term), resulting in 22 different bond indices. We match each sample bond with a bond index by rating and maturity to control for credit and term risks, respectively. As noted by Bessembinder et al. (2009), Lehman Brothers indexes are based on bond returns from both going-concern and bankrupt firms. This fact (the inclusion in the indices of bond returns from bankrupt firms) leads to an upward bias in our BDABR measure. Because we hypothesize that the impact of short selling on bond returns is either negative or insignificant, this bias makes our results more conservative. Two challenges emerge in our estimation of each company’s bond returns. First, although the majority of our sample firms have only one bond outstanding, we still have 28 firms 11

with multiple bonds. Second, the sample size is relatively small due to the fact that Nasdaq firms are generally small in size and most do not have public debt outstanding. Prior studies address the first complication by using the weighted average of the firm’s multiple bond returns, selecting a random representative bond and using that bond return only, or treating each bond as a separate observation. Bessembinder et al. (2009) argue that using the weighted average of the return on the firm’s outstanding bonds is more parsimonious, but using a small sample of bonds is likely to result in an empirical framework that has no power to detect abnormal returns – even when they are in fact present. Considering the complexity and tradeoff between parsimony and power, we follow the methodology of Gebhardt, Hvidkjaer, and Swaminathan (2005), Asquith, Au, Covert, Pathak (2012) and others, and treat each bond as a separate observation. This empirical approach, however, biases the standard error of the mean return downward and the associated t-statistics upward (Eberhart and Siddique (2002)). To account for the intra-firm clustered correlation and to eliminate biases in standard errors, we apply the methodology suggested by Petersen (2009) and Thompson (2011). The second main variable of interest is short selling. Since our collected bond prices are monthly, we construct a monthly measure of short selling using daily short trades while preserving useful information in these observations. Short selling, SS, is the daily ratio of shares shorted by non-exempt customers to thousands of shares outstanding of a given firm. Specifically, to test the relation between abnormal short selling and bond return

bond returns. The impact of abnormal short selling on bond returns is robust to various risk, liquidity, and characteristics controls. In examining the source of information in short selling, we find that firms associated with heavy short selling and lower bond returns are likely to have lower earnings surprises and higher credit risk in the .