Online Appendix to Do Short-Sellers Trade on Private Information or False Information? by Amiyatosh Purnanandam and Nejat Seyhun December 12, 2017 Purnanandam, amiyatos@umich.edu, University of Michigan, Ann Arbor; Seyhun, nseyhun@umich.edu, University of Michigan, Ann Arbor. We thank an anonymous referee, Hank Bessembinder (the editor), Brad Barber, Sunil Wahal, Kent Womack, and seminar participants at the University of Michigan, Koc University, and RMA UNC Academic Forum for several useful suggestions. All remaining errors are our own. 1
We provide some additional results in this Appendix that have been omitted from the main text of the paper for brevity. A.1. Daily Short-Sales In our tests in the paper, we use monthly short-interest data to separate the information discovery hypothesis from manipulation hypothesis. If manipulation occurs within a shorter frequency than one month, then our monthly data analysis will not have the power to detect it. To overcome this challenge, in this section we present an analysis based on a sample of daily short-sales and insider trading data. The short-sale transactions data come from NYSE TAQ dataset pursuant to Regulation SHO reporting. As per this regulation, a pilot program was instituted to collect data on a set of securities from January 2005 to May 2007. The analysis in this section, therefore, can also be taken as a refinement of our results in the rest of the paper. We obtain data on daily short-sales from this dataset and match it with insider trading data on a daily basis. We drop invalid and cancelled transactions from the SHO dataset. For each firm in the sample, we create a measure of abnormal short-selling by subtracting the average short-sales of the last 60 days from the day s short-sales and dividing it by the standard deviation of short-sales measured over the same window. We assign stocks, everyday, into nine portfolios based on the activities of insiders and short-sellers. Along the insider trading dimensions, we create three buckets based on whether insiders sell, do not trade, or buy their company s stock on a net basis on a given day. On the shortselling dimension, we assign them into three buckets based on whether a given stock s abnormal short-selling falls in top, middle, or bottom 1/3rd of the distribution. In total we have data from 545 trading days (we miss 60 trading days in the beginning of the sample since we need data on historical mean and standard deviation). There are 2
1520 stocks in our portfolio on an average day giving us a total sample size of about 828,000 stock-days over this period. We have 33,697 stock-days (4.1% of the sample) in which insiders are net sellers and 4,850 stock-days (0.59% of the sample) in which insiders are net buyers. The average short-sale is about 0.16% of the shares outstanding of the firm. We compute the average stock return for several holding periods, 1-day, 3-days, 15-days, and 30-days after the portfolio formation. Using the Fama-MacBeth approach we regress average stock returns on eight dummy variables, representing the eight portfolio assignments. We leave out the portfolio with the neutral signal from both insiders and short-sellers as the baseline variable in the regression model. The time series mean and t-statistics, after Newey-West correction to account for overlapping return days, are presented in Table A.2. Our results show that when both signals are negative (insiders selling and shorts increasing their position), the next day s return is -0.06% which is significant. On the other extreme when both agents act in the positive direction, the next day s return is a large 0.15% which is also significant at the 1% level. More important, we do not find any evidence of return reversal for 3, 15 or 30 days out in future, after the daily portfolio formation. In other words, the price impact of these trades seems to reflect permanent information and not any short-lived manipulative pressure on the stock price. Overall these daily results corroborate our monthly analysis. A.2. Merger Months Recent research has shown that short-sellers can also earn profits by trading on public information (see Drake at al. (2011), and Engelberg et al. (2012)). Our hypothesis regarding the private information channel is not mutually exclusive to this alternative. Indeed, short-sellers can earn profits by engaging in both of these activities: by discovering private information and by interpreting publicly available information. In our 3
research design, our identification assumption is that insider trades provide a relatively cleaner signal of private information as compared to this alternative channel. If the insider trading signals highly correlate with months where short-sellers make profits on publicly-available signals, then our explanation that short-sellers bring additional value relevant private information to the market can be questionable. Our results on the lagged short-seller data alleviates this concern to some extent. To establish the robustness of our results to this alternative, we focus on one of the most visible and value relevant public signal the announcement of corporate mergers. We obtain data on all corporate mergers during our sample period form Thomson Reuters s SDC dataset. From our sample, we exclude firm-month observations where the firm has been involved in a merger either as an acquirer or as a target. The exclusion of these months ensures that our results are not driven by at least one key public information event. Results are provided in Table A.3. Our results remain almost identical to the base case analysis. While ruling out the effect of all other publicly available information is a daunting task, the robustness of our results to merger-excluded sample provides confidence in our main argument that short-sellers bring some value relevant information to the market. 4
Table A.1: Number of Observations in Portfolios This Table provides the number of observations in each cell formed on the basis of the intersection of insider trading and short-selling signal. The percentage of observations that fall in each row and column are also provided in the Table. Frequency of Observation Short\Insider Sell Neutral Buy % of Total Increasing 25994 156202 6004 33.33% Neutral 22013 160874 5389 33.33% Covering 14688 167406 6024 33.33% % of Total 11.10% 85.81% 3.08% 5
Table A.2: Fama-MacBeth regression with Daily Data This table provides results from Fama-MacBeth regression results. The dependent variable is the average one-day stock return (in percentage) for different holding periods after the portfolio formation. High Insider Low Shorts is a variable that equals one if insiders are net buyers in the preceding day and the stock s standardized short-interest at the end of the preceding day falls in bottom 1/3 rd, zero otherwise. Low Insider High Shorts is a variable that equals one if insiders are net sellers and standardized short-interest falls in top 1/3 rd, zero otherwise. Other interaction variables are defined similarly. We estimate this cross-sectional regression on a daily basis for 545 days and then report the time-series means and Newey-West (NW) corrected t-statistics based on these estimates. NW correction lag is chosen to equal the number of holding days in returns minus one. Holding Period: 1-day 3-day 15-day 30-day Dependent Var: % daily return beta beta beta beta (t-stat) (t-stat) (t-stat) (t-stat) I Low Insider x High Shorts -0.0589-0.0460-0.0279-0.0230 (-3.17) (-3.82) (-3.87) (-3.58) II Low Insider x Medium Shorts 0.0014-0.0194-0.0219-0.0193 (0.07) (-1.63) (-2.62) (-2.42) III Low Insider x Low Shorts 0.0125-0.0059-0.0082-0.0090 (0.60) (-0.47) (-1.03) (-1.44) IV Medium Insider x High Shorts -0.0056-0.0051-0.0058-0.0056 (-0.88) (-1.21) (-2.25) (-3.29) V Medium Insider x Low Shorts 0.0042 0.0042-0.0014-0.0001 (0.69) (1.14) (-0.61) (-0.03) VI High Insider x High Shorts 0.3980 0.3593 0.1315 0.0812 (6.50) (11.69) (8.17) (5.43) VII High Insider x Medium Shorts 0.2994 0.2437 0.0665 0.0395 (5.60) (7.82) (4.27) (3.22) VIII High Insider x Low Shorts 0.1526 0.1878 0.0622 0.0328 (3.25) (5.45) (3.31) (2.08) 6
Table A.3: Fama-French regressions: Excluding Merger Months This table provides results from Fama-French calendar time portfolio return regressions for trading strategies based on different combinations of short-interest and insider-trading signals. We provide regression results for the portfolio that buys stocks with positive signals from both insiders and short-sellers (i.e., insiders are buying and shorts covering) and sells stocks with negative signals from both (i.e., insiders selling and shorts increasing). In the Base case, we reproduce the Fama-French regression result of the base case portfolio formation strategy documented in the earlier Table. The Excluding row refers to a portfolio formation strategy in which we exclude all firm-month observations in which a firm was either a merger target or an acquirer. Equal-Weighted Returns Value-Weighted Returns Alpha Market SMB HML UMD R 2 Alpha Market SMB HML UMD R 2 Excluding 1.84-0.09 0.29-0.02-0.44 22.05% 1.46 0.09 0.00 0.03-0.39 17.87% (5.89) (-1.13) (3.35) (-0.16) (-4.91) (4.46) (0.86) (0.02) (0.23) (-6.53) Base 1.85-0.11 0.30-0.02-0.43 21.52% 1.44 0.11 0.06-0.06-0.38 17.46% (5.92) (-1.20) (3.54) (-0.12) (-4.75) (4.37) (0.96) (0.54) (0.33) (-6.42) 7