Short Selling and Firm Investment Efficiency: Evidence from a Natural Experiment

Short Selling and Firm Investment Efficiency: Evidence from a Natural Experiment Zhihong Chen Hong Kong University of Science and Technology E-mail: aczh@ust.hk Tel.: +852 2358-7574 Ke Wang University of Alberta E-mail: k.wang@ualberta.ca Tel.: +1 780-492-1630 This version: August 2017 Abstract We examine the causal link between short selling and firm investment efficiency using the experiment of the SEC Regulation SHO (Reg SHO) pilot program that relaxes short-selling constraints on a group of randomly selected firms during 2005 2007. Based on a difference-indifferences approach, we find significant improvements in investment efficiency for pilot firms compared with non-pilot firms during the pilot period, which manifests itself in reductions in both underinvestment and overinvestment. We also find that the pilot program increases the risk of overinvesting pilot firms being targeted by short sellers. In turn, pilot firms increase their propensity to make downward adjustments to their capital expenditure plans when following the plans would result in a high level of overinvestment. External financing of pilot firms with a high ex-ante likelihood of underinvestment also increases relative to similar non-pilot firms. Crosssectional analysis shows a more pronounced effect of the pilot program among firms with a poorer information environment and greater divergence in investor opinion. Collectively, our study suggests that reducing short-selling constraints improves investment efficiency by reducing market frictions. Keywords: Short selling, Reg SHO, pilot program, investment efficiency, market frictions. JEL classification: G14, G31, M41, M48 We thank Russell Investments for providing the list of Russell 3000 index firms for 2004 and 2005. All remaining errors are our own.

1. Introduction We examine whether there is a causal link between short selling and investment efficiency. Answering this question is important given the controversial role of short selling in capital markets and the real economy, especially during the recent financial crisis (e.g., Boehmer, Jones and Zhang, 2013; Scannell and Strasburg, 2008). Moreover, real investment efficiency is a key determinant of economic productivity and critically determines a nation s success in global competition (Thurow, 1992; Biddle and Hilary, 2006; Chen, Hope, Li and Wang, 2011). Short selling can potentially reduce overinvestment due to the agency problem of overvaluation by reducing the likelihood of overvaluation in the first place (Jensen, 2005). Short sellers also target firms that have engaged in misallocation of capital or other misconduct. 1 Therefore, prices become more efficient as the negative consequences of inefficient resource allocation are revealed earlier (Boehmer and Wu, 2013; Fang, Huang and Karpoff, 2016). As a result, shareholders have stronger incentives and abilities to intervene and managers have weaker incentives to misallocate capital (Massa, Zhang and Zhang, 2013). The improved price efficiency can also mitigate underinvestment by facilitating external financing (Rajan and Zingales, 1998). Nevertheless, due to arbitrage costs such as noise trading, short selling may not reduce overvaluation when prices deviate significantly from fundamental values and overinvestment is most likely (Shleifer and Vishny, 1997; Jensen, 2005). In addition, manipulative short selling may lead managers to cancel value-enhancing projects, exacerbating underinvestment (Goldstein and Guembel, 2008). The impact of short selling on investment efficiency is thus ultimately an empirical question. 1 For instance, a Bloomberg article on 28 June 2016 reports that short sellers have recently targeted Turkish airlines for its massive expansion in spite of the shrinking prospects of the country s tourist industry (Levitov and Ozsoy, 2016). 1

We attempt to answer this question by empirically examining the effect of relaxing short selling constraints on corporate investment efficiency. Our identification strategy is based on a pilot program (hereafter the pilot program) under Rule 202T of Regulation SHO of the U.S. Securities and Exchange Commission (SEC). The pilot program randomly selected a group of firms (hereafter pilot firms) from the Russell 3000 index. During the pilot period (May 2, 2005 to August 6, 2007), short-sale price tests were suspended for pilot firms but not for the remaining firms in the index (hereafter non-pilot firms). Thus, the pilot program represents an exogenous decrease in short-selling constraints for pilot firms relative to non-pilot firms (Diether, Lee, and Werner, 2009; Grullon, Michenaud, and Weston, 2015). This exogenous shock in short-selling constraints enables us to conduct a difference-in-differences (DiD) test to draw causal inference on the link between short-selling constraints and investment efficiency. We compare pilot firms with non-pilot firms in terms of changes in investment efficiency from the pre-pilot period (2001 2003) to the pilot period (2005 2007). Following prior literature (e.g., Biddle, Hilary and Verdi, 2009; Cheng, Dhaliwal and Zhang, 2013; García Lara, García Osma and Penalva, 2016), we measure a firm s investment inefficiency based on the sensitivity of investment to a proxy for its ex-ante tendency to underinvest or overinvest. Using a DiD approach, we document evidence that the pilot program improves investment efficiency. More specifically, compared with non-pilot firms, pilot firms show a more pronounced decrease in both overinvestment and underinvestment in the pilot period. The results remain statistically significant after controlling for other economic determinants of investment efficiency documented in the prior literature. Compared with similar non-pilot firms, pilot firms with the highest ex-ante tendency to underinvest increase the investment by 3.9% of total assets in the pilot period, while pilot firms with the highest ex-ante tendency to overinvest reduce the 2

investment by 3.3% of total assets. These numbers represent about 24.7% and 21.5% of the mean investment levels of the sample (15.7% of total assets). Thus, the effects are also economically significant. We conduct a series of robustness tests on the baseline findings. First, we find similar results using a sample of pilot and non-pilot firms matched based on a propensity score (Cheng et al., 2013). We also perform placebo tests by assigning a pseudo-pilot program start month before the actual start month. We do not find significant differences in the changes in investment efficiency between pilot and non-pilot firms around the pseudo-events, ruling out non-parallel trends between the two groups of firms as an explanation for the baseline results. Finally, we find similar results when we measure overinvestment and underinvestment by the deviation of actual investment from the expected levels of investment (Goodman, Neamtiu, Shroff and White, 2014). We then proceed to examine the mechanisms through which the pilot program increases investment efficiency. First, we find that pilot firms observe a more pronounced increase in the sensitivity of short-selling activities to overinvestment in the pilot period than non-pilot firms. Pilot firms with higher levels of overinvestment in the pre-pilot period also have more negative market reactions to the announcement of the pilot program. This evidence suggests that the pilot program does facilitate short sellers in targeting overinvesting firms, and rational investors anticipate more active short selling in the pilot period. Second, using a subsample with manually collected capital expenditure plan data, we examine how the pilot program affects the way firms adjust to their capital expenditure plans. We find that, relative to non-pilot firms, pilot firms increase the propensity to make downward adjustments to planned capital expenditure when following the plan is likely to result in a high level of overinvestment. These results, taken together, suggest that pilot firms decrease overinvestment to reduce the likelihood of being 3

targeted by short sellers. Lastly, we find a more pronounced increase in external financing among pilot firms with a high ex-ante tendency to underinvest compared with similar non-pilot firms. This evidence suggests that the pilot program mitigates underinvestment by facilitating external financing. Finally, we examine the cross-sectional variation in the effect of the pilot program on investment efficiency. First, we find that the effect of the pilot program on investment efficiency is more pronounced for pilot firms with greater increase in short interest in the pilot period and more negative market reaction to the announcement of the pilot program. This evidence suggests that pilot firms change their investment behaviors anticipating changes in short-selling activities. Second, we find a more pronounced improvement in the investment efficiency for pilot firms with a poor information environment as reflected in higher absolute discretionary accruals and high bid-ask spread. This evidence supports the view that the pilot program increases investment efficiency by reducing market frictions arising from information asymmetry. Finally, theory suggests that short selling constraints matter more when divergence in investor opinion is greater (Miller, 1977). If the pilot program increases investment efficiency by relaxing short-selling constraints, the effect should be more pronounced for firms with higher level of divergence in investor opinion. We find evidence consistent with this prediction. The increase in investment efficiency for pilot firms is more pronounced for firms with higher trading volume and analyst earnings forecast dispersion before the pilot-period. Overall, the evidence from these crosssectional tests further increases our confidence to attribute the increase in investment efficiency for pilot firms to decreases in short-selling constraints. Our study contributes to the growing literature on the economic consequences of short selling. Several papers examine the capital market impact of short selling activity (e.g., Diether 4

et al., 2009). Other studies examine how short selling affects firms decisions on financial reporting, voluntary disclosure, governance choice, and compensation contracting (Fang et al., 2016; Chen, Cheng, Luo and Yue, 2014; Li and Zhang, 2015; Massa et al., 2013, 2016; De Angelis, Grullon and Michenaud, 2017). Our paper contributes by examining the effect of short selling on real investment efficiency. Two related papers also examine the effects of short selling on corporate investment (Grullon et al., 2015; Chang, Lin and Ma, 2016). Grullon et al. find that the pilot program reduces firms investment, especially for small firms. Chang et al. find a negative association between shares available for lending and corporate investment. Both studies suggest relaxing short-selling constraints results in a lower corporate investment, which could be driven by a decrease in overinvestment or an increase in underinvestment, or both. Our study, however, finds that relaxing short-selling constraints reduces investment only when firms are most likely to overinvest ex ante. In contrast, for firms that are most likely to underinvest ex ante, relaxing short-selling constraints actually increase investment. Our paper also contributes to the literature on the economic determinants of real investment efficiency. Several prior studies find that firms with higher quality of financial reporting or disclosures invest more efficiently (Biddle and Hilary, 2006; Hope and Thomas, 2008; Biddle et al., 2009; Francis, Huang, Khurana and Pereira, 2009; Chen et al., 2011; Cheng et al., 2013). Other studies find that accelerating the incorporation of bad news into accounting earnings improves investment efficiency (Francis and Martin, 2010; Bushman, Piotroski and Smith, 2011; García Lara et al., 2016). Chen, Huang, Kusnadi and Wei (2016) and Edmans, Jayaraman and Schneemeier (2016) find that restricting insider trading improves investment efficiency by improving price efficiency and transparency. Our study suggests that short selling that accelerates the incorporation of bad news into prices improves real investment efficiency. 5

The rest of the paper proceeds as follows. Section 2 introduces the institutional background of Reg SHO and develops hypotheses. Section 3 discusses the research design of this study. Section 4 reports the empirical results. Section 5 concludes the paper. 2. Hypothesis Development 2.1. Short-sale price tests and the SEC Regulation SHO Short-sale price tests were introduced by the SEC in 1938 to deter short sellers from manipulating the market or, more specifically, intentionally driving down stock prices and subsequently making profits by unwinding their short positions. These tests take two forms: (i) the tick test for the NYSE under Rule 10a-1; and (ii) the bid test for Nasdaq under NASD Rule 3350. 2 On October 28, 2003, the SEC proposed Regulation SHO, which includes a pilot program that would temporarily remove the price tests (SEC, 2003). On July 28, 2004, the SEC adopted the new regulation for short sales and announced a list of randomly selected pilot firms, for which the price tests would be suspended during the one-year pilot period starting on January 3, 2005 (SEC, 2004a, 2004b). SEC later extended the pilot period, initially from May 2, 2005 to April 28, 2006 (SEC, 2004c), to a new closing date of August 6, 2007 (SEC, 2006). 2.2. Hypothesis development Prior studies have shown that firms invest sub-optimally due to market frictions arising from moral hazard and adverse selection (Jensen and Meckling, 1976; Myers and Majluf, 1984; Stein, 2003). On one hand, self-interested managers overinvest as they prefer managing larger as opposed to more profitable firms to pursue private benefits such as social status, power, 2 See SEC (2003, 2007) for details about the short-sale price tests. 6

compensation, and prestige (Jensen, 1986, 1993). Managers are also reluctant to discontinue loss-making projects to avoid negative reputation effects or to enjoy a quiet life (Kanodia, Bushman and Dickhaut, 1989; Bertrand and Mullainathan, 2003). In addition, overinvestment may stem from the agency problem of overvaluation when managers need to satisfy overoptimistic growth expectations of investors that underlie over-valuation (Jensen, 2005). On the other hand, agency conflicts between managers and shareholders can also lead to underinvestment. Bertrand and Mullainathan (2003) show that a preference for a quiet life could result in underinvestment when managerial decisions concern whether to create a new line of business. Alternatively, underinvestment may arise because managers are more risk-averse than shareholders due to career concerns or non-diversifiable human capital risks (Amihud and Lev, 1981). Overinvestment squanders corporate resources, whereas underinvestment costs firms in terms of competitive advantage, both of which can lead to poor future performance. We hypothesize that Reg SHO improves the investment efficiency of pilot firms by mitigating moral hazard and adverse selection. Prior studies suggest that restricting short selling exacerbates overvaluation when investors hold diverse opinions (e.g., Miller, 1977; Chen, Hong and Stein, 2001; Hong and Stein, 2003; Boehmer, Danielsen and Sorescu, 2006). Prior studies have also shown that short sellers are informed investors who target over-valued equities (e.g., Dechow, Hutton, Meulbroek and Sloan, 2001). Jensen (2005) argues that the most effective remedy for the agency problem of overvaluation is to prevent overvaluation from arising. Thus, reducing short-selling constraints can mitigate overinvestment arising from overvaluation by reducing the likelihood of equities being overvalued. Moreover, Kedia and Philippon (2009) posit that manipulating reported profits is necessary for managers of low-productivity firms to mimic high-productivity firms by overinvestment because investment and reported profits must 7

be consistent in any signaling equilibrium. Relaxing short-selling constraints can also mitigate overinvestment by deterring earnings manipulation (Fang et al., 2016; Massa et al., 2015). In addition, reducing short-selling constraints can motivate investors to analyze and acquire information (Edmans, 2009; Engelberg, Reed and Ringgenberg, 2012) and accelerate the incorporation of information into prices (Diamond and Verrecchia, 1987; Alexander and Peterson, 2008), which in turn leads to more efficient equity prices (Boehmer and Wu, 2013; Drake, Myers, Myers and Stuart, 2015; Fang et al., 2016). Existing studies also show that short selling provides information to the debt market and improves its efficiency (Kecskés, Mansi and Zhang, 2013; Griffin, Hong and Kim, 2016). More efficient prices reveal the negative outcomes of underinvestment and overinvestment earlier and more accurately. This increases the ability of directors, shareholders, creditors, financial analysts and other market participants to take timelier actions (Edmans and Manso, 2011). Furthermore, when negative outcomes are incorporated into prices more quickly, shareholders are less likely to exit at inflated prices and thus have greater incentive to intervene in sub-optimal managerial decisions (Massa et al., 2013). All of these effects increase the risks to managers of reduced compensation, job loss, and reputational damage, which in turn reduce the ex-ante incentives for both overinvestment and underinvestment. Reducing short-selling constraints also reduces underinvestment by mitigating capital rationing due to adverse selection and moral hazard. Prior studies have shown that adverse selection arising from information asymmetry between managers and external capital suppliers can result in capital rationing. In addition, investors and creditors ration capital when they are concerned that managers overinvest funds in negative-npv projects (Keeton, 1979; Stiglitz and Weiss, 1981). Capital rationing can lead to underinvestment because firms with promising 8

investment opportunities may not have sufficient internal funds to finance their investments. Thus, to the extent that reducing short-selling constraints improves price efficiency and mitigates overinvestment, it also relieves external financing constraints and reduces underinvestment. The above discussion leads to our hypothesis: H1: Overinvestment and underinvestment by pilot firms decrease during the pilot program period. However, Reg SHO may not be effective at improving investment efficiency for several reasons. First, due to other arbitrage costs such as noise trading, short selling may be ineffective in extreme cases when prices diverge significantly from their fundamental values and drive overinvestment (Shleifer and Vishny, 1997; Lee, 2001; Jensen, 2005). Kaplan, Moskowitz and Sensoy (2013) show that short selling restrictions only have a small effect on overvaluation. Second, manipulative short selling could exert downward price pressure and destabilize the market, which undermines the capital allocation role of the financial market. For example, Goldstein and Guembel (2008) show that short selling facilitates price manipulation by uninformed speculators, which may lead managers to change their assessment of investment opportunities and cancel value-enhancing investment projects. Several studies show that equity prices are more sensitive to negative news in the pilot program period (e.g., De Angelis et al. 2017; Li and Zhang 2015). Thus, a decrease in short-selling restrictions could even exacerbate underinvestment. These possibilities, if valid, add tension to our research question and make it more difficult to find improvements in investment efficiency for pilot firms during the pilot period. 3. Research design 9

3.1. Model specification and variable definition Our empirical model builds on prior literature that measures a firm s under- and overinvestment based on the sensitivity of investment to a proxy for the ex-ante likelihood of under- and overinvestment (Biddle et al., 2009; Cheng et al., 2013; García Lara et al., 2016). More specifically, INVEST i,t = b 1 + b 2 OverTEND i,t-1 + CONTROLS i,t-1 + i,t (1) where OverTEND is the ex-ante tendency to overinvest. We follow Biddle et al. (2009) and define OverTEND as the average of standardized decile ranks of cash holdings and the standardized decile ranks of negative leverage. A low or high value of OverTEND means a high ex-ante tendency to underinvest or overinvest respectively. Thus the coefficient b 1 measures a firm s investment when the ex-ante likelihood of underinvestment is the highest (i.e., OverTEND = 0). A more negative value of b 1 implies greater underinvestment. (b 1 +b 2 ) then measures a firm s investment when the ex-ante likelihood of overinvestment is the highest (i.e., OverTEND = 1). A more positive value of (b 1 +b 2 ) implies greater overinvestment. We then model coefficients b 1 and b 2 as functions of the pilot program using a DiD approach, controlling for other determinants of under- and overinvestment. That is, b 1 = α 1 + α 2 PILOT i + α 3 POST t + α 4 PILOT i POST t + α 5 AQ i,t-1 + α 6 InstOwn i,t-1 + α 7 Analyst i,t-1 + α 8 InvG i,t-1 + α 9 MissingG i,t-1 (2) and b 2 = 1 + 2 PILOT i + 3 POST t + 4 PILOT i POST t + 5 AQ i,t-1 + 6 InstOwn i,t-1 + 7 Analyst i,t-1 + 8 InvG i,t-1 + 9 MissingG i,t-1 (3) where PILOT is a dummy variable that equals one for pilot firms and zero otherwise. POST is a dummy variable that equals one for observations in the pilot program period and zero otherwise. The coefficients α 2 and (α 2 + 2 ) measure the difference in underinvestment and 10

overinvestment respectively between the pilot and non-pilot firms before the pilot program. The coefficients α 3 and (α 3 + 3 ) gauge the change in underinvestment and overinvestment during the pilot program period for non-pilot firms. The coefficients of interest are α 4 and 4. The coefficient α 4 captures the impact of the pilot program on firms underinvestment, and (α 4 + 4 ) captures the impact of the pilot program on firms overinvestment. H1 predicts a negative (α 4 + 4 ) and a positive α 4. Following prior literature (Biddle et al., 2009; Cheng et al., 2013; García Lara et al., 2016), we control the impact on under- and overinvestment of accrual quality (AQ), institutional ownership (InstOwn), financial analyst following (Analyst), and shareholder rights as measured by the negative G-index (InvG) developed by Gompers, Ishii, and Metrick (2003). We set InvG at zero for observations with a missing G-index and include an indicator variable (MissingG) that equals one for these observations and zero otherwise. Substituting (2) and (3) into (1), we have our baseline model specification. INVEST i,t = α 1 + α 2 PILOT i + α 3 POST t + α 4 PILOT i POST t + α 5 AQ i,t-1 + α 6 InstOwn i,t-1 + α 7 Analyst i,t-1 + α 8 InvG i,t-1 + α 9 MissingG i,t-1 + 1 OverTEND i,t-1 + 2 PILOT i OverTEND i,t-1 + 3 POST t OverTEND i,t-1 + 4 PILOT i POST t OverTEND i,t-1 + 5 AQ i,t-1 OverTEND i,t-1 + 6 InstOwn i,t-1 OverTEND i,t-1 + 7 Analyst i,t-1 OverTEND i,t-1 + 8 InvG i,t-1 OverTEND i,t-1 + 9 MissingG i,t-1 OverTEND i,t-1 + CONTROLS i,t-1 + i,t (4) INVEST is total investment, measured as net capital expenditure plus R&D expenditure and acquisitions, and scaled by lagged total assets. We multiply INVEST by 100 for expressional convenience. Finally, we follow prior studies (e.g., Biddle et al. 2009; Cheng et al., 2013; García Lara et al., 2016) and control for industry fixed effects and a battery of other control variables (CONTROLS), including firm size, firm age, market-to-book ratio, operating cash flows, dividend payments, operating cycle, asset tangibility, industry leverage, volatility of sales, 11

volatility of past investment, volatility of operating cash flows, and an indicator of loss. The Appendix contains detailed definitions of all variables. Following prior literature (Biddle et al. 2009; Fang et al., 2016), we adjust the standard errors for clustering at both firm and year levels (Gow, Ormazabal and Taylor, 2010). 3.2. Sample and data We obtain the 2004 and 2005 versions of the Russell 3000 index constituents from Russell Investments. Following prior studies (e.g., Diether et al., 2009; Grullon et al., 2015), we require our sample firms to be included in both versions. We then merge the constituent list with the list of 986 pilot firms obtained from the SEC. 3 This process generates 884 pilot firms and 1,764 non-pilot firms. We follow prior literature (e.g., Biddle et al., 2009; Grullon et al., 2015) and drop financial institutions and utilities firms as they are inherently different in nature of investment. We then merge the firm list with the data required to estimate Equation (4). We obtain financial accounting data and short interest data from Compustat, stock returns and trading volume data from CRSP, institutional ownership data from Thomson Reuters Institutional Holdings (Form 13F), analyst coverage data from I/B/E/S, and shareholder rights data from ISS (formerly RiskMetrics). Our pre-pilot program period sample includes observations between 2001 and 2003, and the pilot program period sample includes observations between 2005 and 2007. We omit 2004 because the SEC approved the pilot program and announced pilot firms in mid-2004 and thus it is unclear to what extent investment in 2004 is affected by the pilot program (e.g., Fang et al., 2016; Hope, Hu and Zhao, 2016). After deleting 3 The list is obtained from SEC Release No. 50104 (http://www.sec.gov/rules/other/34-50104.htm). 12

observations with missing variables required to estimate Equation (1), our final sample contains 5,154 firm-years observations for 1,151 firms, including 401 pilot firms and 750 non-pilot firms. 3.3. Descriptive statistics Table 1 presents the summary statistics for the variables used in the baseline regression. On average, total investment represents 15.7% of total assets for our sample firms, comparable to the statistics in prior studies (e.g., Biddle et al., 2009). The mean and median values of firm size are 6.626 and 6.486 respectively. The mean/median values of institutional ownership and number of analysts are 0.662/0.714 and 11.010/9.000, respectively. These numbers are larger than those in Biddle et al. (2009) but are comparable with the figures reported by extant studies on the pilot program (e.g., Li and Zhang, 2015). This is unsurprising as we only include firms in the Russell 3000 index while Biddle et al. (2009) use all Compustat firms. [Insert Table 1 here] 4. Empirical Results 4.1. The effects of the pilot program on investment efficiency Table 2 presents the results of the test for H1. Column (1) reports the results without control variables. The coefficient of PILOT POST is positive and statistically significant (3.784, t = 6.27). The results suggest that pilot firms with the highest ex-ante likelihood of underinvestment increase their investment relative to non-pilot firms under similar conditions. The evidence is thus consistent with our hypothesis that the pilot program mitigates underinvestment. The coefficient of PILOT POST OverTEND is negative and statistically significant (-7.998, t = -4.67), and the sum of the coefficients of PILOT POST and 13

PILOT POST OverTEND is also negative (-4.214, p < 0.001). Thus the results suggest that pilot firms with the highest ex-ante likelihood of overinvestment reduce their investment relative to non-pilot firms with similar characteristics after the start of the pilot program. This evidence is consistent with our hypothesis that the pilot program mitigates overinvestment. Column (2) shows the results of our baseline regression. The conclusion holds after controlling for other determinants of investment efficiency. The coefficient of PILOT POST is positive (3.886, t = 6.00), and the sum of the coefficients of PILOT POST OverTEND and PILOT POST OverTEND is significantly negative (-3.328, p = 0.005). In terms of economic significance, our baseline regression results suggest that the pilot program increases the investment of pilot firms with the highest ex-ante likelihood of underinvestment by 3.9% of total assets. The pilot program also reduces the investment of pilot firms with the highest ex-ante likelihood of overinvestment by 3.3% of total assets. These numbers represent 24.7% and 21.5% of the sample mean of investment respectively (15.7%, reported in Table 1) and are economically significant. With respect to the control variables, we generally find positive coefficients for accrual quality, institutional ownership, analyst following and shareholder rights. The coefficients of the interactions between OverTEND and the above variables are negative. The sign of the coefficients are generally consistent with prior literature (e.g., Biddle et al., 2009; Cheng et al., 2013). However, these coefficients are not statistically significant except those of institutional ownership and the interaction between shareholder rights and overinvestment tendency. This could be because our sample is limited to firms covered by the Russell 3000 index and thus only 14

includes relatively large firms. Since large firms are subject to greater market scrutiny, the effects of financial reporting quality and other governance mechanism are likely to be weaker. 4 [Insert Table 2 here] 4.2. Robustness tests 4.2.1. Propensity score matching Our baseline analysis includes all non-pilot firms as the control group as pilot firms are randomly selected from the Russell 3000 index. As shown in Panel A of Table 3, the differences in most firm characteristics related to investment between pilot and non-pilot firms are insignificant except market to book ratio, volatility of past investment, Altman (1968) Z-Score, firm age, and the loss indicator. To further check whether our baseline results are driven by differences in firm characteristics between pilot and non-pilot firms, we estimate the baseline regression using a sample matched based on propensity scores. 5 Following Cheng et al. (2013), we include all variables listed in Panel A of Table 3 as covariates to estimate the propensity score. 6 The right four columns of Panel A show that none of the firm characteristics show significant differences between pilot and non-pilot firms after matching, suggesting that our matching procedure yields balanced covariates between the treatment and control samples (Tucker 2010; Shipman, Swanquist and Whited, 2017). The regression results based on the propensity score matching sample are reported in Panel B of Table 3. The results are 4 When we include all Compustat firms in our sample period, we obtain significantly positive coefficients of AQ, and the sum of coefficients of AQ and AQ OverTEND is negative and significant (results untabulated). 5 For each pilot firm, we find, without replacement, the non-pilot firm with the closest propensity score within the same industry as the matched control firm. 6 Untabulated results show that only firm age shows a significant coefficient in the logistic regression. The pseudo R-squared of the logistic regression is 1.7%. 15

qualitatively similar to our baseline results. The coefficient of PILOT POST is positive and significant (5.337, t = 4.81). The sum of the coefficients of PILOT POST and PILOT POST OverTEND is significantly negative (-2.993, p = 0.021). [Insert Table 3 here] 4.2.2. Placebo tests To further address the concern that pilot and non-pilot firms may have already been experiencing different changes in investment efficiency before the start of the pilot program (Roberts and Whited, 2011), we repeat the baseline test in Table 2 using a series of pseudo-pilot programs. Specifically, we assign July of year T (T = 1999 to 2002) as the pseudo-pilot program start date. We then repeat the test in Table 2 using the observations of pilot and non-pilot firms in years [T 3, T 1] and years [T+1, T+3]. Years [T 3, T 1] are defined as the pre-pseudo pilot program period and years [T+1, T+3] are defined as the post-pseudo pilot program period. 7 The results are reported in Table 4. We do not find significant differences between pilot and non-pilot firms in terms of changes in investment efficiency around any of the pseudo-pilot program start dates. This further confirms that the baseline findings are not driven by confounding events during the pilot program period. It also rules out non-parallel trends as an explanation for the baseline findings. [Insert Table 4 here] 4.2.3. Alternative measures of overinvestment and underinvestment 7 When T = 2002, we only include 2003 as the post-pseudo-pilot program period. When T = 2001, we only include 2002 and 2003 as the post-pseudo-pilot program period to ensure that the post-pseudo-pilot program period does not overlap with the actual post-pilot program period. 16

Our baseline results are based on investment conditional on a proxy for ex-ante likelihood of underinvestment and overinvestment. We also test whether our results are robust to an unconditional measure of investment efficiency, namely the deviation of actual investment from expected investment given a firm s investment opportunities (e.g., Biddle et al., 2009; Richardson, 2006; McNichols and Stubben, 2008; Goodman et al., 2014). Specifically, we measure investment inefficiency by the residual terms from the following regression (Goodman et al., 2014). INVEST i,t = α 1 + α 2 MTB i,t-1 + α 3 CFO i,t + α 4 AGRW i,t-1 + α 5 INVEST i,t-1 + t (5) where MTB is market-to-book ratio, CFO is operating cash flows scaled by lagged total assets, and AGRW is asset growth. We estimate regression (5) within each industry and year. Following Biddle et al. (2009) and Goodman et al. (2014), we first sort firms into quartiles based on the residuals obtained from the above regressions. We then define firm-year observations in the bottom quartile (i.e., the most negative residuals) as those having underinvested and firmyear observations in the top quartile (i.e., the most positive residuals) as those having overinvested. Firm-year observations in the middle two quartiles are used as the benchmark group. We then estimate a multinomial logit model that links the probability that a firm lies in the underinvestment or overinvestment quartile as opposed to the two benchmark quartiles as a function of the pilot program as follows: INVEST_STATE i,t = α 1 + α 2 PILOT i + α 3 POST t + α 4 PILOT i POST t + CONTROLS i,t-1 + i,t (6) where i and t are firm and year indicators respectively. INVEST_STATE i,t equals 2 if firm i is classified as neither overinvesting nor underinvesting firm in year t, 1 if firm i is classified as an underinvesting firm (UnderINVEST = 1) in year t, and 3 if firm i is classified as an overinvesting firm (OverINVEST = 1) in year t. We follow Biddle et al. (2009) and further 17

control firm leverage and slack in the regression in addition to the control variables included in regression (4). Firms with INVEST_STATE = 2 (OverINVEST = 0 and UnderINVEST = 0) serve as the benchmark. Hypothesis H1 predicts a negative coefficient of PILOT POST for both the underinvestment and overinvestment regressions. The results are reported in Table 5. Consistent with our hypothesis, we find a more pronounced decrease in the likelihood of both underinvestment and overinvestment for pilot firms than for non-pilot firms. The coefficient of POST PILOT is significantly negative for both underinvestment (-0.194, z = -1.76) and overinvestment (-0.230, z = -2.32). These results are consistent with the findings based on the conditional tests and further support our hypothesis. [Insert Table 5 here] 4.3. Further analysis We hypothesize that the pilot program improves investment efficiency by enhancing the monitoring role of short sellers. In addition, it may improve market efficiency and facilitate external financing, which further reduces underinvestment. If this is the case, pilot firms that make sub-optimal investments are more likely to be targeted by short sellers in the pilot period. Anticipating the increased threat from short sellers, firms are expected to not only plan their investment more efficiently but also adjust their planned investment more actively when these plans are inefficient. Finally, firms with a high ex-ante likelihood of underinvestment are expected to raise more external capital to finance investment. We then conduct several further analyses to answer the following questions. First, does the pilot program increase the propensity of short sellers to target firms that make sub-optimal investments? Towards this end, we examine the change in the sensitivity of subsequent shortselling activities to sub-optimal investment. Second, does the pilot program increase firms 18

propensity to correct their investment plans when such plans are sub-optimal? We attempt to answer this question by examining how the pilot program affects firms adjustments to their capital expenditure plans using a subsample with manually collected capital expenditure plan data. Finally, we examine whether the pilot program increases the external financing of pilot firms with a high ex-ante likelihood of underinvestment. These tests provide further insight into the mechanisms through which the pilot program enhances investment efficiency. We explain detailed research design and present empirical evidence below. 4.3.1. The pilot program and the sensitivity of short selling to sub-optimal investment We adopt a similar method to that of Desai, Krishnamurthy and Venkataraman (2006) and estimate the following regression to examine the effect of the pilot program on the propensity of short sellers to target firms making sub-optimal investments. SHORT i,t = α 1 + α 2 PILOT i + α 3 POST t + α 4 PILOT i POST t + 1 OverINVEST i,t-1 + 2 PILOT i OverINVEST i,t-1 + 3 POST t OverINVEST i,t-1 + 4 PILOT i POST t OverINVEST i,t-1 + 1 UnderINVEST i,t-1 + 2 PILOT i UnderINVEST i,t-1 + 3 POST t UnderINVEST i,t-1 + 4 PILOT i POST t UnderINVEST i,t-1 + CONTROLS i,t-1 + i,t (7) Following Desai et al. (2006), we measure short-selling activities as the change in short interests. OverINVEST t and UnderINVEST t are dummy variables that equal one if the firm is classified as an overinvesting or underinvesting firm respectively in year t and zero otherwise. The classification algorithm is the same as that used in the unconditional test reported in section 4.2.3. We separate the overinvesting and underinvesting firms to allow the possibility that short sellers treat the two types of firms differently. 4 and 4 measure the effect of the pilot program on the propensity of short sellers to target overinvesting and underinvesting firms respectively 19

and are the coefficients of interest. Following Desai et al. (2006), we control for firm size, market-to-book ratio, price momentum, return volatility, accruals, and industry fixed effects. The results are reported in Table 6. The coefficient of PILOT POST OverINVEST is positive and significant (0.023, t = 2.64), which supports the notion that the pilot program facilitates short sellers in targeting overinvesting pilot firms. In contrast, the coefficient of PILOT POST UnderINVEST is insignificant (-0.002, t = -0.44), suggesting that the pilot program does not affect the propensity of short sellers to target underinvesting firms. A possible explanation for the asymmetric effect is that investment is at least partially irreversible (Dixit and Pindyck, 1994). Thus, once a negative NPV project has been initiated, firms may still incur significant costs even if it is discontinued. On the other hand, if a firm has the option of waiting, underinvesting in the current period does not necessarily imply losing the investment opportunities. Overall, the evidence reported in Table 6 is consistent with the view that the pilot program increases the disciplinary role of short sellers. [Insert Table 6 here] In addition, if rational investors anticipate more active short-selling activities targeting overinvesting firms, we also expect to observe a more negative market reaction to the announcement of the pilot program for pilot firms that massively overinvest prior to the announcement. To test this prediction, we estimate the following model. CAR i = α 1 + α 2 OverINVEST i + α 3 UnderINVEST i + 1 PILOT i + 2 PILOT i OverINVEST i + 3 PILOT i UnderINVEST i + CONTROLS i + i (8) Following Grullon et al. (2015), we measure market reaction by cumulative abnormal returns (CAR) from ten trading days before the announcement date to one trading day after (i.e., [-10, +1]) to incorporate potential information leakage. We also treat overinvesting and 20

underinvesting firms separately. OverINVEST and UnderINVEST are indicator variables for firms that overinvest or underinvest respectively in the fiscal year immediately before the announcement date. 2 and 3 measures the difference in market reaction between the underinvesting and overinvesting pilot firms and the benchmark pilot firms respectively. The control variables include firm size as measured by the logarithm of market value of equity (logmv) and market-to-book ratio (MTB). The results are reported in Table 7. We find that the coefficient of PILOT OverINVEST is significantly negative (-0.023, t = -2.17), suggesting that investors do expect the pilot program to increase the risks of overinvesting pilot firms being targeted by short sellers. Nevertheless, the coefficient of PILOT UnderINVEST is insignificant (-0.006, t = -0.49). This is consistent with the findings in Table 6 that the pilot program does not increase the propensity of short sellers to target underinvesting pilot firms. [Insert Table 7 here] 4.3.2. The pilot program and adjustments to investment plans We estimate the following regression to test the effect of the pilot program on firms propensity to adjust their planned capital expenditure when following such plans is likely to lead to significant underinvestment or overinvestment. CAPX_ADJUST i,t = α 1 + α 2 PILOT i + α 3 POST t + α 4 PILOT i POST t + 1 OverPLAN i,t + 2 PILOT i OverPLAN i,t + 3 POST t OverPLAN i,t + 4 PILOT i POST t OverPLAN i,t + CONTROLS i,t-1 + i,t (9) CAPX_ADJUST i,t is the adjustment to capital expenditure plans, defined as the actual capital expenditure in year t minus the planned capital expenditure for year t, scaled by total assets at the end of year t-1. A positive sign of CAPX_ADJUST means an upward adjustment to 21

the planned capital expenditure while a negative sign implies a downward adjustment. OverPLAN is a measure of degree of planned overinvestment. We adopt a method similar to regression (5) to define planned overinvestment or underinvestment. Specifically, we replace the dependent variable in (5) with the planned capital expenditure scaled by lagged total assets and measure the inefficiency of the capital expenditure plan using the residuals. We define OverPLAN as the standardized decile ranks of the residuals. Firms in the top or bottom decile are most likely to overinvest or underinvest respectively when they follow their capital expenditure plans. The coefficients in interest are α 4 and (α 4 + 4 ). A positive coefficient of α 4 means that the pilot program increases firms propensity to make upward adjustments to their planned capital expenditure when following the plan would result in a high level of underinvestment. A negative (α 4 + 4 ) suggests that the pilot program increases firms propensity to reduce their capital expenditure from the planned level when following the plan would result in a high level of overinvestment. We include the same control variables as those in regression (4). We manually collect the capital expenditure plan data from Form 10-Ks. After merging the capital expenditure plan data with the sample of our baseline regression, the sample size is reduced to 1,975, about 38% of the baseline regression sample. 8 The results are reported in Table 8. The sum of the coefficients of PILOT POST and PILOT POST OverPLAN is negative and significant (-7.062, p = 0.045). Thus, pilot firms increase their propensity to make downward adjustments if following their capital expenditure plans is likely to result in high levels of 8 The capital expenditure plan for fiscal year t is disclosed in Form 10-K for year t-1. We manually searched 19,460 10-K filings for fiscal years 2000 2006 and found capital expenditure plan disclosures in point or close range forecasts for 38.5% (7,478) of the observations. We use the point estimate when it is provided and the mid-point of the close range otherwise. 22

overinvestment compared with similar non-pilot firms. This evidence suggests that pilot firms cut overinvestment due to the increased concerns of being targeted by short sellers. 9 The effect of the pilot program on the adjustments is weaker for pilot firms with capital expenditure plans that are likely to result in the greatest degree of underinvestment. The coefficient of PILOT POST is positive, albeit insignificant (3.596, t = 1.56). One possible reason for the low significance is that firms may require more time to increase investments than to cut a planned investment. 10 [Insert Table 8 here] 4.3.3. The pilot program and external financing regression. We examine the effect of the pilot program on external financing with the following DebtIssue i,t (EquityIssue i,t ) = α 1 + α 2 PILOT i + α 3 POST t + α 4 PILOT i POST t + 1 OverTEND i,t-1 + 2 PILOT i OverTEND i,t-1 + 3 POST t OverTEND i,t-1 + 4 PILOT i POST t OverTEND i,t-1 + CONTROLS i,t-1 + i,t (10) Equation (10) is identical to equation (4) except that the dependent variable is external financing. We consider debt financing (DebtIssue) and equity financing (EquityIssue) separately. The coefficient of interest is α 4, which measures the effect of the pilot program on the external financing of firms with the highest ex-ante likelihood of underinvestment. 9 Untabulated results suggest that pilot firms also make more efficient investment plans. Specifically, we re-estimate regression (4), replacing INVEST with planned capital expenditure. We continue to find that the coefficient of PILOT POST is significantly positive and that the sum of the coefficients of PILOT POST and PILOT POST OverTEND is significantly negative. In addition, we do not find that the pilot program changes the likelihood of pilot firms disclosing capital expenditure plans. Thus selection bias does not appear to be a serious concern. 10 For example, firms may need extra time to plan the details of the new investment, coordinate with existing plans, and raise extra finance. 23

The results are reported in Table 9. The coefficient of PILOT POST is significantly positive for both debt financing (3.242, t = 2.50) and equity financing (1.852, t = 2.19) regressions. Thus, the pilot program increases the external debt and equity financing of pilot firms with a high ex-ante tendency to underinvest relative to non-pilot firms under similar conditions. This evidence suggests that the pilot program mitigates underinvestment by facilitating external financing. In contrast, the pilot program does not appear to affect external financing for firms with the highest ex-ante likelihood of overinvestment. The sum of the coefficients of PILOT POST and PILOT POST OverTEND (i.e., α 4 + 4 ) is insignificantly negative in the debt financing regression (p = 0.135) and insignificantly positive in the equity financing regression (p = 0.434). [Insert Table 9 here] 4.4. Cross-sectional variations in the effect of the pilot program on investment efficiency In this section, we conduct several cross-sectional analyses to further substantiate our hypothesis. First, if the difference in the change in investment efficiency between pilot and nonpilot firms around 2004 is driven by shocks of short-selling constraints, the difference in investment efficiency is expected to be correlated with the change in short-selling activities after the start of the program. Second, if the pilot program increases investment efficiency because it improves the information environment and reduces market frictions, then the effect is expected to be stronger for firms with a poorer information environment before the pilot program. Finally, since short selling matters more when divergence in investor opinion is high, we expect to observe a more pronounced impact of the pilot program for firms with higher divergence. 24

To test these predictions, we first separate the sample into two sub-samples based on a partitioning variable that proxies for the change in short-selling activities, information environment, or divergence in investor opinion. We then estimate the baseline regression for each sub-sample and then compare the key coefficients between the sub-samples. 4.4.1. Changes in short-selling activities after the pilot program We use two proxies to measure the change in short-selling activities driven by the pilot program. Our first proxy is the change in average monthly short interest scaled by shares outstanding from 2003 to 2005. Since the change in short interest can be affected by changes in firms investment decisions, we also use the narrow-window market reactions to the announcement of the pilot program (i.e., CAR[-10,+1]) to measure the expected change in shortselling activities. The results are reported in Panel A of Table 10. We only report the statistics for the key variables for brevity. We find that the improvement in investment efficiency is more pronounced for firms with stronger capital market effects of the pilot program. In column (1), we find a significantly positive coefficient of PILOT POST for firms with large increases in short interest (5.028, t = 4.48), whereas the coefficient for firms with small increases in short interest is insignificant (1.054, t = 1.33). The difference between the two coefficients is also statistically significant (p < 0.001). The sum of the coefficients of PILOT POST and PILOT POST OverTEND is significantly negative for firms with large increases in short interest (p = 0.003). However, the sum of coefficients even becomes positive for firms with small increases (p = 0.077). The difference in the sum of the coefficients is also significant (p < 0.001). The evidence thus suggests that the effects of the pilot program on reducing both underinvestment and overinvestment are more pronounced for firms with larger increases in 25