The Information Intermediary Role of Short Sellers

Transcription

1 01/27/03 The Information Intermediary Role of Short Sellers by Grace Pownall Goizueta Business School Emory University and Paul Simko Darden Graduate School of Business University of Virginia January 2003 We are grateful to the New York Stock Exchange for providing the short interest data used in this paper, and to Sudipta Basu, Jennifer Francis, Carla Hayn, Elizabeth Keating, Katherine Schipper, Nicholas Valerio, Greg Waymire, an anonymous hedge fund manager, workshop participants at the Middle States Accounting Research Symposium, the Southeast Summer Accounting Research Colloquium, London Business School, Emory and Northwestern Universities, and the Universities of Houston and Wisconsin for helpful comments. The Goizueta Business School of Emory University provided generous financial support for this research.

2 The Information Intermediary Role of Short Sellers I. Introduction In this paper we provide evidence on the extent to which the activities of short sellers are associated with information reflected in stock price. Specifically, we investigate: (1) the observable fundamental characteristics of firms targeted by short sellers; (2) the degree to which the price response to the revelation of large increases in short selling is conditional on these fundamentals; and (3) whether the positions taken by short sellers are profitable. Anecdotal accounts by short sellers suggest that these traders' positions are based largely on analyses of publicly-available information. For instance, Staley (1997) and Asensio (2001) assert that short sellers target firms characterized by "price bubbles," deteriorating fundamentals, and poor performance masked by manipulative accounting practices. Presumably, short sellers' use of publicly available data is dictated, in part, by the hostility of corporate managers to short selling, making it difficult to obtain inside information (e.g., Frank 1996). The extent to which observable firm characteristics explain increases in short sales, and the market response to these increases, is important for understanding both the short selling phenomenon and the role of accounting fundamentals for financial analysis. The possibility of profitable short selling based on publicly available data is suggested by various studies showing a relation between fundamental analysis and future stock returns (Ou and Penman 1989; Lev and Thiagarajan 1993; Abarbanell and Bushee 1997). These studies can be characterized as part of a broader literature on the value of investment advice, mostly that provided by sell-side analysts. This literature has investigated the nature of firms followed by analysts, the extent to which analysts' recommendations have information content for stock prices, and whether trading positions based on such recommendations are profitable. For instance, Jegadeesh and Kim (2001) find that analysts are more likely to provide positive 2

3 recommendations for high growth and momentum stocks, and McNichols and O'Brien (1997) document that analysts are more likely to initiate coverage on firms with favorable future prospects. Others document that analyst recommendations are informative for prices (e.g., Stickel 1985; Lys and Sohn 1990), but the extent to which trading rules based on investment recommendations are profitable is ambiguous (see Francis and Soffer 1997; Mikhail et al. 1997; Barber et al. 1998; Jegadeesh and Kim 2001). Several studies have posited that favorable news is more readily reflected in stock price because of the activities of analysts acting on positive expectations about firms' future earnings (e.g., McNichols and O'Brien 1997). In our paper, we likewise view short sellers as financial intermediaries issuing (explicit or implicit) "forecasts" of earnings news, but through their observable trades. These forecasts would fall in the lower tail of the earnings expectation distribution often not covered by analysts. We contribute to this literature by comprehensively examining a similar set of issues (i.e., the nature of firms targeted by short sellers, the information content of large increases in short interest, and the profitability of short positions) using statistically large short sales in NYSE common stocks. Two factors motivate our analysis. First, prior research suggests that prices are less accurate on the downside than on the upside with respect to firms' future prospects (see Finn et al. 1999; Lee and Swaminathan 2000). This suggests that returns to fundamental analysis may be greater for short positions, for which the objective is to profit from downward price revisions, than for long positions. Second, and more importantly, the availability of short selling data facilitates the analysis of actual trading decisions as opposed to hypothetical trading rules or the experimental formation of hedge portfolios that may not be implementable in real time. With these data we can assess the extent to which (presumably negative) information is exploited by short sellers and whether investors learn from these trading decisions and update prices accordingly. 3

4 This paper also contributes to the literature investigating the stock price patterns associated with short trading. Only a handful of studies examine the extent to which the revelation of actual short sales results in significant contemporaneous downward price revisions and subsequent underperformance (Senchack and Starks 1993; Desai et al. 2002; Asquith and Meulbroek 1995; and Dechow et al. 2001). Our objective is to build on this work (all examining short interest data from prior to 1994) by examining more recent short sales, controlling for market-wide movements in short interest, and providing more comprehensive evidence on the economics of intermediation by short sellers. In our first set of tests, we examine the observable characteristics of firms that experience abnormally high short interest increases in their common stock (i.e., short spikes ) as announced in the Wall Street Journal (WSJ) during the years 1988 through We measure a set of variables, both before and after the reported short spike, reflecting firms' "earnings quality" (Lev and Thiagarajan 1993) as well as other conventional accounting metrics used to gauge performance and risk (e.g., earnings growth, size, and leverage). In these simple tests we find that firms targeted by short sellers tend to exhibit little systematic difference from the population of NYSE firms during the period immediately prior to the short spike. However, in the year subsequent to the short spike our sample of firms experience significant declines in key earningsbased fundamentals such as earnings-to-price and earnings growth. In our second set of tests, we examine the nature of stock returns at the time short selling activity is revealed to the market, and further whether abnormal announcement period returns are associated with key accounting fundamentals. If short selling is based on correct perceptions of overvaluation, we predict that stock returns around short spikes will be negative, and that abnormal returns will be more negative if accounting fundamentals relative to price are deteriorating. Consistent with Senchack and Starks (1993), we find that the sample-wide 4

5 abnormal return at the time of short spike announcements is negative but modest. We estimate various regression models of cumulative abnormal returns on accounting fundamentals and proxies for price bubbles and the expected costs of short selling. These regression analyses indicate that in the five day interval starting on the WSJ announcement date, abnormal returns are more negative the larger is earnings relative to price in the prior quarter and the smaller is the growth in earnings-to-price in the current quarter. In addition, abnormal returns are more negative the larger the stock price run-up in the several months prior to the short spike. These relationships are robust to several sensitivity analyses and are stronger for data prior to 1994 (the time period prior to large increases in the incidence of hedge fund trading). The results are also stronger for short positions greater than 2.5 percent of outstanding shares, for non-optionable shares, and for firms experiencing multiple short spikes in our study period. The proxies for other accounting fundamentals and for the costs and risks associated with short selling are not significantly associated with abnormal announcement period returns. A third and final set of tests examines the profitability of short selling. The basic hypothesis we test is that stock returns subsequent to increases in short interest for the firm will, on average, be negative. We test this hypothesis by examining abnormal returns from the announcement period through the month short interest returns to normal levels (or zero). The results of these analyses support the following conclusions. First, almost 20 percent of the short spikes in our sample revert to pre-spike levels within a month of the announcement, and over half of the short spikes have reverted by the fourth month. Second, the short positions in our sample are on average profitable for short sellers, with mean cumulative returns from the announcement date to the reversion date of almost 5 percent. These cumulative abnormal returns are significantly larger for pre-1994 spikes, non-optionable shares, short positions of greater than 2.5 percent of outstanding shares, and single spikes for a given sample firm, with total returns to 5

6 these spikes ranging from 6 percent to over 9 percent. These returns are heavily influenced by short positions that remain open for at least nine months, however, with the largest returns earned on short positions that are open for more than 12 months. Therefore, it is not clear that the sample short positions are ultimately profitable considering the trading risk of extended open short positions. Finally, we also document that short sellers' profits are associated with our proxies for accounting fundamentals and price bubbles, including a weak association with the Lev and Thiagarajan (1993) fundamental signal scores. In contrast to the announcement period results, the percentage of outstanding shares shorted is strongly positively related and firm size is significantly negatively related to short sellers' profits, consistent with short sellers incurring increased risk of a short squeeze when their positions are information-based. The rest of the paper is organized as follows. In section II, we briefly discuss the institutional background of short selling and the primary hypotheses we examine. Sample selection and data are described in section III. Results are reported in section IV, and section V provides a summary of the major results and their implications. II. Background and Hypotheses Rule 3b-3 of the Securities Exchange Act of 1934 defines a short sale as "any sale of a security which the seller does not own or any sale which is consummated by the delivery of a security borrowed by, or for the account of, the seller". To deliver a security sold short to the purchaser, the short seller borrows the security, typically from a broker-dealer or an institutional investor. 1 The short seller closes the position by returning the security to the lender, typically by repurchasing it in the open market. Short sellers in general expect to either profit from a future price decline or hedge against the risk of a position in the same or a related security. 1 Fabozzi (1997) and Reed (2001) describe the equity lending market in further detail. 6

7 According to the SEC, short selling provides two benefits to the capital market (see SEC 1999). First, short selling by market makers and large institutional traders provides market liquidity by offsetting temporary imbalances in the supply and demand for equity securities. Second, short sellers contribute to market efficiency because their information search and processing activities inform the capital market of future price decreases. However, the SEC has regulated short selling of securities traded on organized stock exchanges in the US for over 60 years, as have the exchanges and the Federal Reserve Board. The justification for these regulations is that short selling may be used to manipulate stock prices. 2 Both the SEC and the NYSE prohibit short selling except on a plus tick or a zero plus tick: that is, short selling is prohibited in a falling market (SEC Rule 10a-1 and NYSE Rule 440B). In addition, Federal Reserve Board Regulation T requires short sellers to place 150 percent of the current market value of most securities sold short in a margin account until the securities are replaced in the account from which they were borrowed. The NYSE also imposes a 30 percent margin maintenance requirement on short sellers, but typically the brokers' margin requirements (which vary among brokerage firms) are more stringent than the exchange requirements. The short seller does not have access to the proceeds of the short sale, and earns only a small fraction of the return on the margin collateral, called the "rebate rate", which varies depending upon the market interest rate and supply and demand conditions for the borrowed securities. 3 2 These fears date back at least to the 1929 stock market crash and the market's sustained inability to recover from the crash amid allegations of "bear raids" in which securities were allegedly sold short to drive down the price by creating an imbalance of sell orders in the market (see SEC 1999). 3 The return on the margin collateral after payment of the rebate is split between the beneficial owner of the securities (the client) and the lender (the broker). Reed (2001) quotes Bargerhuff & Associates in their "Securities Lending Analytics" for the 2nd quarter of 2000 as asserting that a 75/25 percent client/lender fee split is typical according to industry statistics, making equity lending a very profitable business in which a securities brokerage can engage. 7

8 The expected benefit of short selling depends on the reason for selling the securities short, only one of which is a short sale driven by firm-specific information suggesting overvaluation. Staley (1997, 4-9) considers information-based short positions to include those based on predictions of future firm-specific stock price declines due to accounting manipulation (e.g., overstated earnings), overvaluation associated with speculative price "bubbles" (e.g., high prices relative to fundamentals as alleged for Internet stocks in early 2000), and/or increased likelihood of adverse consequences due to changes in the firm's external environment (e.g., firms selling "fad" products expected to experience a decline in demand). Further, Staley (1997, ) asserts that much of the information used by short sellers is already in the public domain, especially in various SEC filings There are other reasons for short selling that are unrelated to firm-specific information (Dechow et al. 2001; SEC 1999; Senchack and Starks 1993), and the prevalence of hedge strategies has increased dramatically during the 1990 s. 4 Pairs trading involves a short position to hedge the risk of a long position taken in a different (but related) firm. It is also common for some traders to go long in the shares of a merger target while short selling a bidder's shares. Alternatively, some traders may take short positions in a takeover target if they believe it is unlikely that a firm will be acquired. Finally, short positions may be hedges against options exposures, commodity price risks, interest rate sensitivity, and/or expected regulatory changes. 5 The institutional context for short selling suggests that the expected cost of short selling depends primarily on the opportunity cost of the collateral that must be held in the margin 4 Tremont Partners and TASS (1999, 5) indicate that the annual growth in the number of hedge funds during a ten year period from the late 1980 s to the late 1990 s was 25.7 percent, or a total increase of 648 percent over the ten year period. Many hedge funds invest in a style that involves some type of equity short selling, but less than one percent of these funds are engaged in short selling as their primary focus (Tremont Partners and TASS (1999, 8-14). 5 The tendency for non-information-based short selling to increase over time is not universal. For instance, the IRS has tried to restrict certain tax-based incentives for short selling. Thus, certain strategies such as dividend capture strategies or selling against the box were eliminated in many circumstances by the Taxpayer Relief Act of 1997 (see CCH 1997a, 1997b). 8

9 account against the short position as well as the supply of the securities to be sold short. Securities that are scarcer in the equity lending market (e.g., those with lower market capitalization or those with lower share turnover) are more expensive to borrow because the rebate is smaller (Reed 2001). In addition, they expose the short seller to the risk of being caught in a "short squeeze", which can occur if a lender demands the securities back while they are trading at a higher price, and a sufficient number of shares are not available to cover the short position. We hypothesize that abnormal returns around short spike announcements are more negative when the level of firm-specific fundamental variables indicates a higher probability that the short spike is due to information-based short selling. Thus, we examine variables indicative of overstated earnings, stock price bubbles, and weak accounting fundamentals. Because theoretical results in Diamond and Verrecchia (1987) and evidence in Senchack and Starks (1993) suggest that the profitability of information-based short selling is higher in stocks without traded options, we expect these relations will be stronger in non-optionable stocks. In terms of short selling costs, we hypothesize that short selling for risk hedging or other purposes is more likely to occur in large, highly liquid stocks. Thus, we expect price responses to short spikes to be less negative as the market capitalization and float of the shorted stock increase, and when the percentage of outstanding shares shorted is lower. We conduct regression tests using both stock price changes at the time of short spike announcements as well as buy and hold returns during the period that the short seller maintains his/her position. The former provides an indication of the extent to which short selling is interpreted by investors as an indicator of bad news. The latter regression based on buy and hold returns provides evidence on the relation between the gross gains to short selling (before risk and transaction costs) and various firm-specific characteristics. 9

10 III. Sample Selection and Data Our analyses are based on a sample of abnormally large positive monthly changes in short interest for common stocks identified from a comprehensive dataset provided by the New York Stock Exchange. This dataset includes monthly short interest levels data on NYSE common stocks for the 144 months between January 1988 and December We first identified all US firms listed on NYSE and also included in the CRSP and Compustat databases during the study period. We computed book to market ratios for each fiscal year end between January 1989 and December 1998 for the US firms listed on NYSE and also on CRSP and Compustat, and excluded the top and bottom 1 percent of the distribution of book to market ratio observations, leaving 18,674 firm-year observations. The mean book to market ratio (B/M) for the population of firms (0.629) is shown on the first line of panel A of table 1, along with the Fundamental Signal Score (FSS) for the 18,674 firm-year observations, computed as in Lev and Thiagarajan (1993) 6. We present these population-level statistics for purposes of comparison as we refine the sample as described in panel A of table 1. The second line of the panel gives the FSS and mean B/M for the population of firm-years which are not included in the sample because there were no short interest observations on the NYSE database for them. The FSS and B/M for the firm-year observations which were not targeted by short sellers is similar to the same statistics for the population. [Insert Table 1 About Here] The next step in selecting a sample was to estimate time series models that identify abnormally large monthly firm-specific changes in short interest. For each firm, we computed the time series of short interest (SI) defined as the number of shares shorted in month t divided 6 The calculation of FSS is described in detail in Section IV and in the notes to table 3. 10

11 by the number of common shares outstanding. We next differenced this series to calculate the change in short interest for the firm in month t ( SI). We then regressed SI (using OLS) for firm i against the contemporaneous average monthly short interest change for NYSE common stocks in month t. The regression model is: SI it = γ 0 + γ 1 SI mt + ε it (1) Based on the firm-specific intercept and slope coefficients from this regression, we estimated a series of abnormal short interest changes ( ASI) as the difference between SI and its predicted value. After computing the standardized abnormal short interest change for the firm in month t ( SASI, equal to ASI divided by the variance of residuals from the short interest market model described above), we identified all firm-month observations meeting two criteria: (1) SASI exceeded 2.0 in month t but was less than 1.2 in month t 1; and (2) for the prior 12 months the observation was not preceded by another short spike that had not yet reverted to prespike short interest levels. 7 Together, these criteria are intended to ensure that the sample contains only large discrete jumps in short interest that are not mere continuations of prior increases in short interest. From the observations identified in this manner, we eliminated all observations occurring either in calendar 1988 or calendar We eliminated 1988 short spikes because the lack of 1987 data made it impossible to identify whether the 1988 spike was a continuation of 1987 trading decisions. Observations from calendar 1999 were deleted since lack of data for the year 2000 made it impossible to identify the final outcome for many 1999 short spikes. As a result, the initial sample from the NYSE dataset includes 3,572 spikes associated with 2,060 firms. 7 As further discussed in Section IV, we define a spike in month t as having reverted if the short interest in the subsequent month is less than a level equal to month 1 short interest plus 25 percent of the jump in short interest during the spike month. 11

12 Table 1 describes additional sample selection criteria. We excluded 97 firms with less than 20 monthly short interest observations (out of the 144 possible), and 165 short spikes occurring in the month of a stock split, which reduced the sample to 3,298 positive short spikes for 1,905 firms. We excluded 379 financial services firms (SIC codes 60-67) and 193 regulated firms (SIC codes 48-49). As a result, our primary sample includes 2,304 short interest spikes associated with 1,333 firms. Panel A describes the FSS and B/M for the observations excluded by each sample selection filter and for the primary sample, measured at the fiscal year ends prior and subsequent to the short spikes. On average, the financial firms and regulated firms excluded from the sample have higher B/M ratios, and the sample of short spike firms have lower B/M ratios, than the population from which the sample was drawn. On the other hand, the primary sample of short spikes is associated with FSS statistics that are less than 2 percent different than the mean FSS for the population. Panel B shows summary statistics on the short interest market model parameters for the 1,333 firms in the final sample. Three points are noteworthy. First, the average explanatory power of the model is limited; the mean (median) adjusted R 2 is.02 (.008). Second, the residuals exhibit modest negative serial correlation. The mean and median correlation between successive residuals is about Finally, the mean slope coefficient exceeds the median by a factor of two suggesting that the model likely identifies a significant relation for a subset of the sample. 8 Panel A of table 2 provides data on the incidence of single and multiple spike firms in the sample. Over half of the 1,333 firms have only one short spike and, in total, 251 firms have three or more short spike events. Panel B shows the incidence of short spikes by calendar month 8 An inspection of the firm-specific adjusted R 2 values confirms that the explanatory power of the model is considerably higher for a subset of the sample. For example, the third quartile of the cross-sectional distribution of firm-specific adjusted R 2 values equals.03. This suggests that a failure to account for market-wide short interest 12

13 and year. These data indicate no obvious clustering by month although the frequency is somewhat lower in January, perhaps because of tax-related short selling (Brent et al. 1990). In terms of years, there is a general upward trend in the frequency of short spikes beginning in 1992: 138 enter the sample in 1992, and this frequency increases each subsequent year to a high of 424 short spikes in [Insert Table 2 About Here] Panel C shows summary statistics on the cross-sectional distribution of the percentage of outstanding shares shorted with observations aligned in event time (month 0 equals short spike month). Summary statistics are shown from months 2 through +3 as well as at six and 12 month lags after the short spike. 9 These results indicate that even though all observations met the criteria for classification as short spikes, many of the resultant positions represent only a small percentage of total shares outstanding. The mean and median percentage of outstanding shares shorted is 2.95 and 1.60 percent, respectively, for the short spike month (i.e., month 0). Finally, there is considerable cross-sectional variation in the level of short interest. For month 0, the minimum percentage of outstanding shares shorted is.0006 percent (the minimum is zero in all other months) and the maximum is percent. Panel C also reports similar summary statistics on changes in short interest for months 1 through +3. While (not surprisingly) these data are consistent with those on short interest levels, they highlight two additional aspects of our sample. First, changes in month 1 are more likely to have a positive sign than a negative sign despite truncating large positive t-1 short interest changes in selecting the sample. Second, the changes in months +1 through +3 are negative at changes could result in a higher incidence with which a firm s short interest changes were incorrectly attributed to firm-specific factors. 9 The sample size declines from 2,304 in month 0 to 2,136 in month +12 because of mergers or takeovers, liquidations, and delistings that occur within 12 months following the spike date. 13

14 both the mean and median, suggesting that the average firm is likely to experience at least a modest decline in short interest in the month following a short spike. IV. Empirical Results Fundamental Characteristics of Short Spike Firms Staley (1997), among others, asserts that short sellers target firms with deteriorating fundamentals, and further that these fundamentals are manifest in publicly available financial reports. This section explores this issue for our sample of short spikes. We choose as our proxies for firm fundamentals the twelve financial components identified by Lev and Thiagarajan (1993), and aggregate the measures into a single, firm-specific FSS consistent with their approach. Below are the twelve characteristics that comprise FSS: Inventory - the percentage change from t-1 to t in inventory (Compustat item 78 or 3) minus the percentage change in sales (Compustat item 12). Accounts Receivable - the percentage change in accounts receivable (Compustat item 2) minus the percentage change in sales (Compustat item 12). Capital Expenditure - the annual (t-1 to t) percentage change in capital expenditures for the firm's 2-digit SIC industry (Compustat item 30) minus the annual percentage change in capital expenditures for the firm. R&D - the annual percentage change in R&D for the firm's 2-digit SIC industry (Compustat item 46) minus the annual percentage change in R&D for the firm. Gross Margin - the annual percentage change in sales (Compustat item 12) minus the annual percentage change in gross margin (Compustat item 12 minus Compustat item 41). Sales & Administrative Expenses - the annual percentage change in SGA (Compustat item 189) minus the annual percentage change in sales. Provision for Doubtful Accounts - the percentage change in gross receivables (Compustat items 2 plus 67) minus percentage change in doubtful receivables. Effective Tax - pretax earnings (Compustat item 170) divided by beginning of the fiscal year price (from CRSP) times T (the effective tax rate) at t-1 minus T at t; where T=63/( ) in Compustat item numbers. But, if both the numerator and the 14

15 denominator of T are negative, the observation is excluded from the calculation; and if the numerator is positive and the denominator is negative, then T is replaced with.34. Order Backlog - the annual percentage change in sales minus the annual percentage change in order backlog (Compustat item 98). Labor Force - the annual percentage change in sales per employee [sales (Compustat item 12) divided by number of employees (Compustat item 29)]. LIFO - set to 0 if the firm uses LIFO or replacement cost and one otherwise, from Compustat item 59. Audit Qualification - equals one for an unqualified audit report and zero otherwise, from Compustat item 149. Following Lev and Thiagarajan (1993), each nonmissing fundamental signal component is replaced with a one if the component has a positive value, with zero if it has a negative value, or with a missing value code. FSS is then the sum of the components divided by the number of nonmissing values. Thus, FSS can range from zero to one, with lower values of FSS implying stronger firm fundamentals. [Insert Table 3 About Here] Panel A of Table 3 summarizes the composite FSS and the twelve fundamental characteristics: (i) for the sample in the fiscal year prior to the short increase (top row of each variable), and (ii) for the sample observations in the fiscal year subsequent (bottom row). For comparative purposes each variable is also provided for the population of 1989 through 1998 NYSE firm-years. It is evident from these distributions that there is little meaningful temporal shift in fundamentals from before the short increase to after, or between the cross-section of short increases and the full NYSE population. Specifically, for only three (five) of 12 components is the mean (median) fundamental significantly weaker than the full population, and for only two (two) components does the fundamental significantly deteriorate across time. 10 In the last line of 10 For some of the mean and median comparisons the differences are too negligible to be observed in the rounded statistics reported on table 3. 15

16 Panel A, the aggregate FSS score is trivially lower but statistically indistinguishable at both the mean and median from the aggregate FSS score for the population of Compustat firm-years. Our conclusion from these comparisons is that neither the individual FSS components nor the aggregate FSS score provide evidence that the sample firms' accounting fundamentals are very different from the population accounting fundamentals, nor that the sample firms' accounting fundamentals are deteriorating through time. Panel B provides similar summary statistics but for other relevant descriptive characteristics of the sample observations: earnings-to-price, book-to-market, return-on-equity, earnings growth, total assets, and financial leverage. At each quartile there exists a general deterioration in the three earnings based fundamentals (e.g., median EPS growth from 16 percent to 11 percent), but only the deterioration in earnings-to-price (at the mean and median) and on EPS growth (at the median) is statistically significant. The comparisons between the sample observations and the population of firm years are more striking, and uniformly statistically significant, however. The comparisons reveal that the sample of short interest observations are for firms that are smaller (i.e., less assets) and with less leverage. The firms which experienced significant increases in short interest had higher earnings-to-price ratios prior to the short interest event but lower earnings-to-price subsequent to the event, and book-to-price ratios that were lower than the population of firm-years both before and after the event. Return-on-equity for the shorted firms was significantly lower prior to the short interest event, but statistically indistinguishable subsequent to the event. Likewise, EPS growth was significantly higher (at the median) prior but significantly lower after the short interest spike than the population. Taken as a whole, the results in Table 3 indicate that the Lev and Thiagarajan (1993) accounting fundamentals do not reflect deteriorating conditions that short sellers target ex ante, although the shorted firms are reliably smaller, with less leverage, higher market values relative 16

17 to their book values, and higher earnings relative to market prices. In our regressions reported in the next section, we confirm that the Lev and Thiagarajan (1993) FSS measures do not explain abnormal returns around the announcement of short spikes, but explore the ability of other variables suggested by Staley (1997) and others to explain abnormal announcement period returns. Stock Price Behavior Associated with Short Spike Announcements Primary Model Our primary tests are based on an OLS regression model in which abnormal stock return from the day of the short spike announcement (day 0) to five days after the announcement (day +5) is the dependent variable. 11 The independent variables are proxies for overstated earnings, weak accounting fundamentals, overvaluation due to price bubbles, and costs and risks associated with short selling. The primary model is: AR 0,+5 = α + δ 1 E Q-1 + δ 2 E Q-1,0 + δ 3 E Q0,+1 + δ 4 B Q-1 + δ 5 B Q-1,0 + δ 6 B Q0,+1 + δ 7 AR -60,-21 + δ 8 TURN PRE + δ 9 MV PRE + δ 10 SI + δ 11 FSS PRE + δ 12 FSS + ε i (2) where 12 : RELDATE = the short interest announcement date; AR 0,+5 = buy and hold market-adjusted return from RELDATE to trading day +5; E Qt = earnings per share before extraordinary items for the quarter t=-1 (ending between 106 and 195 calendar days prior to RELDATE), t=0 (ending between 16 and 105 calendar days prior to RELDATE), or t=+1 (ending between 15 days prior to 75 days subsequent to RELDATE) divided by price at trading day -21 relative to RELDATE; E Q-1,0 = change in earnings per share between quarter t=-1 and quarter t=0 divided by price at day -21; 11 As reported later, we replicated these regressions with several other specifications of the dependent variable, including the cumulative holding period return through the month of reversion for each short spike. 12 All variables are firm-specific, but firm subscripts are suppressed for simplicity. 17

18 E Q0,+1 = change in earnings per share between quarter t=0 and quarter t= +1 divided by price at day -21; B Qt = book value per share at the end of quarters t=-1, t=0, or t=+1 divided by price at day -21; B Q-1,0 = change in book value per share from quarter t=-1 to quarter t=0 divided by price at day -21; B Q0,+1 = change in book value per share from quarter t=0 to quarter t=+1 divided by price at day -21; AR -60,-21 = buy and hold market-adjusted return from 21 to 60 days before RELDATE; TURN PRE = natural log of the total volume of shares traded over the earliest 21 trading day period in the interval beginning 141 trading days prior to RELDATE and ending on the day before RELDATE; MV PRE = natural log of the market value of the sample firm (in millions of US dollars) on the earliest trading day in the interval beginning 120 trading days prior to RELDATE and ending the day before RELDATE; SI = number of shares sold short divided by total number of common shares outstanding on RELDATE; FSS PRE = FSS computed as in Lev and Thiagarajan (1993) as of the last fiscal year end at least one month but no more than 13 months prior to RELDATE; FSS = change in FSS between the last fiscal year end at least one month but no more than 13 months prior to RELDATE and the first fiscal year end no more than one month before but no more than 11 months after RELDATE. Earnings levels and changes are included to capture overstated earnings which, when revealed to the market, would lead to stock price declines. We expected overstated earnings to be associated with more negative announcement period abnormal returns, and earnings decreases to be associated with more negative abnormal returns. Thus, our expectation for δ 1 is negative and those for δ 2 and δ 3 positive. Similarly, we expected that stock price run-ups prior to short selling would indicate overvaluation; hence, we expected AR i(-60, -21) to enter the regression with a negative coefficient. 18

19 We expected abnormal returns to be inversely related to the costs of short selling, our proxies for which are share turnover, firm size, and short interest level. Turnover relates to share liquidity, so selling short less liquid shares would imply the position would be viewed as more risky and therefore more likely to be an information-based short sale. Conversely, selling short shares in a large firm is less risky and therefore more likely to be a hedge against underlying industry or location-specific risks and less likely to be an information-based short sale. SI -- the percentage of outstanding shares shorted -- is a proxy for the risk of a short squeeze, and therefore we expected higher levels of SI to be associated with information-based short positions and more negative abnormal returns. Finally, book- to-price and FSS are proxies for accounting fundamentals and should have a negative relation with abnormal returns at the time of a short spike. Weaker accounting fundamentals, reflected in higher book-to-market and FSS, should be associated with more negative stock returns when the short spike is announced. Increases in the book-to-price and in FSS, reflecting deteriorating fundamentals, should also be negatively related to abnormal returns. Panel A of table 4 summarizes the distribution of each variable used in the estimation of equation (2). The first three lines describe the abnormal returns data used in the regression analyses, and include the primary dependent variable (AR i(0, +5) ) as well as cumulative abnormal returns over days -5 to 0 (AR i(-5, 0) ) and abnormal returns on day 0 alone (AR i(0) ). The mean of the abnormal returns measures on day 0 is zero, and the medians of all three abnormal returns measures are negative but small, consistent with earlier research (Senchack and Starks 1993). All three abnormal return measures have a substantial amount of cross-sectional variability, however, with an interquartile range of approximately 5.5 percent (2.1 percent) for the six (one) day return intervals. [Insert Table 4 About Here] 19

20 Panel A of table 4 shows that the short spikes are associated with relatively low earnings, quarterly earnings growth, and book values (all relative to price) around the time of the reported short spikes. The abnormal return from day -60 to day -21, our proxy for run-ups in stock price reflecting price bubbles, varies considerably across the sample, ranging from -10 percent at the first quartile to +7 percent at the third quartile. The average (median) percentage of outstanding shares sold short across the sample is about 3 (2) percent, as reported on table 2 panel C. The mean and median change in FSS from just before to just after the short spike are both zero, but the interquartile range is from -11 to +13 percent. The correlation structure of the regression variables (not tabulated) indicates that AR 0,+5, AR 0, and AR -5,0 are positively and significantly correlated. However, with correlations ranging from.12 to.42, it is apparent that each return measure captures a different set of information. In addition, the correlation between raw returns and abnormal returns is in excess of.90. Finally, the correlations among the independent variables are particularly high between market value and share turnover, between FSS PRE and FSS, and between consecutive earnings levels, book value levels, changes in earnings, and changes in book values. Since these pairs of independent variables are so highly correlated, diagnostic regressions with reduced sets of variables may be especially relevant. Results from the estimation of our primary regression model, equation (2), are reported on table 5. Consistent with expectations, coefficient estimates that capture overstated earnings and price bubbles are associated with negative announcement period returns. In particular, coefficients on prior earnings and the run-up in stock price two and three months prior to the short spike are both negative and significant. No other coefficients are statistically significant We replicated the results reported in column 3 using cumulative raw returns instead of abnormal returns as the dependent variable, with similar inferences except that the coefficient on book-to-price was positive and marginally significant. We also replicated the results reported in column 3 without the variables that would not have been 20

21 To assess the sensitivity of the results to the correlation structure in the independent variables, columns 4 and 5 of table 5 report results from estimating reduced versions of equation (2), where in each regression one of each pair of highly correlated variables are dropped from the regression. In each reduced regression, the negative and significant coefficients on prior earnings and prior stock price run-up recur, as do the insignificant coefficients on the proxies for the costs and risks of short selling and on the proxies for accounting fundamentals. However, when earnings and earnings changes are included in the regressions with book-to-price but no other earnings or book value variables, the coefficient on earnings changes is positive and significant. Across the full and two reduced regressions, the R 2 is constant at 11 percent. Reported extensions and diagnostics on equation (2) will be based on the reduced model in column 4, although in all cases inferences similar to those reported are supported by using the full primary regression or either of the reduced versions in columns 4 and 5. Taken together, the results reported on table 5 suggest that firms for which the announcement of short spikes is associated with significant stock price decreases are more likely to be those with overstated earnings in the prior period, decreases in earnings from the quarter before to the quarter of the short spike, and recent run-ups in stock price. Diagnostics and Extensions to the Primary Regression Tests Table 6 summarizes extensions and diagnostics of our primary regression tests. Column 2 of table 6 repeats the coefficient estimates and statistics of the reduced version of the primary regression for purposes of comparison. Columns 3, 4, and 5 report estimations of three partitions of the sample: (i) short spikes during the years ; (ii) short spikes representing greater than 2.5 percent of outstanding shares; (iii) short spikes in securities that were not optionable; observable at the time of the short spike announcement (change in earnings and change in book value, both relative 21

22 and (iv) short spikes in the securities of all sample firms which experienced more than one short spike during the study period. We focus on the first half of the sample because our data indicates increases in short interest over our time period. We speculate that a substantial part of this increase comes from the risk management activities of hedge funds adopting short positions to hedge industry- and time-period-specific risk associated with long positions or derivatives positions. 14 Therefore, we expected short spikes from to be associated with a higher probability of being information-based short selling, and thus that our regression model would better explain abnormal announcement period returns in the earlier half of the period. 15 We focus on short spikes to greater than 2.5 percent of outstanding shares because those are the observations for which the risk of a short squeeze is highest. Therefore, we expected that these positions were more likely to be information-based, and thus that our regression model would better explain abnormal returns for the largest short spikes relative to shares outstanding. 16 Senchack and Starks (1993) found that optionability of sample firms' common stock was significantly related to abnormal returns around the time of short interest announcements, so we expected that short spike observations associated with optionable stocks would be more likely to be hedges against derivatives positions. Therefore, the short spikes in non-optionable stocks are more likely to be information-based, and we expected them to be associated with larger negative abnormal returns. Finally, we speculated that multiple spikes are more likely to be contexts in which conflicts of opinion between short sellers and firm management were extended and acrimonious, to price, from the quarter of the short spike to the quarter after the short spike), with nearly identical results. 14 This speculation has been confirmed by a hedge fund manager who wishes to remain anonymous. 15 In addition, the earlier half of the period corresponds to the time period studied in most prior literature (e.g., Senchack and Starks 1993; Asquith and Meulbroek 1995; Dechow et al. 2001). 16 These observations are also interesting because they are similar to the substantial short positions studied by Dechow et al. (2001). 22

23 and therefore that these spikes were more likely to be information-based than the single spikes associated with firms whose securities were not of repeated interest to short sellers. The third column of table 6 reports the coefficients for the pre-1994 subsample of 475 observations with complete data. This specification yields the highest R 2 among the regressions at.33. As in the primary regression, coefficients on prior earnings, contemporaneous earnings changes, and prior stock returns are significant and of the predicted sign. Also consistent with the primary regression, the coefficients on the proxies for the risks and costs of short selling are not significantly different than zero, nor is the coefficient on FSS. However, the coefficient on book-to-price is significantly negative, providing some support for traders choosing to short firms with weak accounting fundamentals in the early part of the sample period. Results for estimating the regression for the 648 observations for which SI is greater than 2.5 percent of shares outstanding are reported in column 4. Although the adjusted R 2 is higher at.20, the coefficient estimates and significance levels are very similar to those of the primary regression. Results in column 5 for estimating the regression on the 1,049 non-optionable observations are similar to the results for the full sample: the adjusted R 2 is about the same (.14), and the coefficient estimates and significance levels are very similar. Likewise, results for estimating the regression on the 1,146 multiple spike observations with complete data, reported in column 6, are also very similar to those for the full sample, with a higher R 2 but nearly the same coefficient estimates and significance levels. Taken together, the results on table 6 suggest that the primary regression results are robust to time period, magnitude of the short interest relative to shares outstanding, optionability of the securities shorted, and frequency of short spikes for a given firm during the sample period. In each case, we find that higher earnings and higher returns prior to the short spike, and earnings decreases from the quarter before to the quarter of the short spike, were associated with 23

24 more negative abnormal returns. We are unable to document that FSS or proxies for the risks and costs associated with short sales are associated with abnormal returns in the announcement period, nor to confirm that the book-to-price ratio was associated with abnormal returns at the time of short spikes after Duration and Profitability of Short Positions Associated with Spike Announcements Table 7 provides evidence on the duration and profitability of short positions associated with short spikes. Panel A shows the frequency of short spike reversion in each of the 12 months following the spike month. For each month, a spike is classified as reverting if the short interest as of that month has declined to its pre-spike levels, plus 25 percent of the initial short spike (column (1)). In addition, we also classify a stock as having reverted if the firm is acquired by another company, is delisted or liquidated, or drops from CRSP for other reasons (shown in columns (2) through (4)). The fifth column shows the number and percentage of firms that have not yet reverted at months +1 through greater than +12. [Insert Table 7 About Here] Table 7 reveals that most of the short spikes in our sample are of short duration, with about 20 percent of the short spikes reverting in month +1. After four months, more than half of the spikes have reverted, suggesting that assuming a holding period of one or two years (as in Dechow et al. 2001) would be inappropriate for the majority of our sample. Panel B reports the cross-sectional mean holding period return, calculated as the marketadjusted buy-and-hold return from the announcement date through the midpoint of the reversion 17 In unreported diagnostics, we also replicated the regression on AR 0 and AR -5,0. These regressions had very low adjusted R 2, but the pattern of significant coefficients in the two alternate event interval specifications was slightly different than in the primary regressions. Prior earnings and both earnings change variables had positive and significant coefficients, unlike in the primary regressions. In addition, change in book-to-price ratio from the quarter of to the quarter after the short spike had a negative and significant coefficient. Run-up in stock price two and three 24