Short selling and firm investment efficiency

The question of interest in this paper is whether short-sellers have an information advantage about a firm’s investment inefficiency before the public. Generally speaking, if short-sellers are informed about the investment inefficiency before it becomes publicly available, it should have a significant relationship between short-selling activities preceding the release of the financial report and the information about inefficiency contained in the report. The level of abnormal short interest will increase as the investment inefficiency worsens. The primary hypothesis is as follows:

Investment inefficiency manifests itself mainly in two ways: either overinvestment or underinvestment. There are three main differences documented in the literature regarding overinvestment and underinvestment.

First, Bebchuk and Stole (1993) show that underinvestment occurs when the market has incomplete information about the level of investment undertaken, while overinvestment occurs when the market can observe the level of investment but not its productivity. Following this signaling theory, if investment inefficiency can be detected by short-sellers, it should be more related to overinvestment rather than underinvestment.

Second, Gilchrist et al. (2005) and Grullon et al. (2015) document that overvaluation leads to overinvestment by artificially reducing the cost of capital. Therefore, overinvestment reflects some information about overvaluation. As a result, short-sellers search for overvalued stocks to make a profit, so they pay more attention to overinvesting stocks.

Third, according to Franzoni (2009), the incentives behind underinvestment and overinvestment are different. Overinvestment occurs when management has empire-building incentives and when corporate governance for the firm is poor, both of which provide negative signals. On the other hand, the motivation behind underinvestment is not clear. It might happen when firms are financially constrained, leading to insufficient funds for investment. Under this case, the signal is negative, and market reactions tend to be significant. However, underinvestment could also be due to other reasons, with market reactions being indistinguishable from zero. These two effects make the response to underinvestment unclear.

Following these arguments, the second hypothesis is formulated as follows:

Recently, researchers find that board independence can significantly impact short-selling activities. For instance, Rahman et al. (2021) argue that a lack of board independence facilitates short-selling due to increased information asymmetry. Additionally, studies suggest that insufficient board independence leads to a deterioration in disclosure quality (Byrd and Hickman 1992; Fama and Jensen 1983; Beasley 1996; Yekini et al. 2015). Short-sellers are known as traders who exploit their informational advantage arising from the information asymmetry. Consequently, a lack of board independence may enable short-sellers to capitalize on their informational edge concerning investment inefficiencies.

Furthermore, board independence can mitigate a firm’s investment inefficiency (Liu et al. 2015), indicating that firms lacking board independence are more susceptible to such inefficiencies. Souther (2021) finds that firms with less board independence are associated with lower firm value, making them more attractive to short-sellers. In other words, the lack of board independence enhances the informativeness of short-sellers regarding a firm’s investment inefficiencies.

Short-selling is closely related to managers’ equity compensations (De Angelis et al. 2017; Shi et al. 2021) because more equity-based compensation can mitigate short-selling threats by aligning managers’ personal interests with corporate stock prices. Firms with less equity-based compensation are more vulnerable to short-selling, as their managers are more likely to engage in value-destroying investments (Mehran 1995; Morck et al. 1988; McConnell and Servaes 1990; Jensen and Murphy 1990; Minnick et al. 2011). As short-sellers profit from downward stock price movements, they tend to avoid firms whose managers are cautious and diligent in stock price movements. Therefore, short-sellers are more likely to pay attention to the investment inefficiency of firms whose managers are not bound by stock performances.

I employ two measures proposed by Biddle et al. (2009) and McNichols and Stubben (2008) to capture a firm’s investment inefficiency. These measures align with neoclassical theory, which posits that a firm’s primary objective is to maximize its net present value while adhering to the constraints of the production function. This pursuit of profit maximization shapes a firm’s optimal investment policy, and any deviation from the optimal level of investment can be considered inefficient.

The two measures diverge in their approach to determining the optimal investment level. Drawing from the acceleration theory, Biddle et al. (2009) assert that the level of capital is directly proportional to the output level. When an economy operates at full capital utilization, firms adapt their investment in response to changes in demand. Biddle et al. (2009) model current investment as a function of past output growth, proxied by sales growth.In this equation, \({\textit{TI}}_{i,t}\) represents firm i’s total investment,⁶ and \({\textit{SG}}\) denotes firm i’s percentage sales growth from year \(t-1\) to year t. \(\epsilon _{i,t}\) represents the residual in the regression, capturing deviations from the expected level of investment.

I utilize its absolute value, standardized by the standard deviation of its industry and year distribution, as the first measure of investment inefficiency.A higher level of \({\textit{Inefficiency}}^B\) indicates lower efficiency in the firm’s investment. I estimate Eq. (1) for each industry-year using the Fama and French 48-industry classification. Consistent with Biddle et al. (2009), I exclude industries with fewer than 20 observations in a given year.

Similarly, in line with Biddle et al. (2009), I calculate the total investment in a given year as the sum of capital expenditures, R &D expenditures, and acquisitions minus the sales of property, plant, and equipment (PPE). This total is then scaled by the lagged total assets. This approach employs an accounting-based framework to estimate total investment as the difference between total investment and asset sales (Richardson 2006). It also takes into account various types of investments, including capital expenditures and acquisitions.

The second measure of investment inefficiency, based on McNichols and Stubben (2008), assumes that investment can be predicted by investment opportunities. The use of Tobin’s Q as a measure of investment opportunities is prevalent in the literature. For example, Goodman et al. (2014) use it as the proxy for investment opportunities and also control for cash flows, growth in assets, and past investment to allow for variations in the relationship between investment and Tobin’s Q. McNichols and Stubben (2008) augment the model by further adding three interaction terms, which individually interact the lagged Tobin’s Q with an indicator variable that equals 1 if the lagged Tobin’s Q is in the second, third, or fourth quartile of its industry and year distribution. The model adopted from McNichols and Stubben (2008) is as follows:where \(Q_{i,t-1}\) is the lagged Tobin’s Q.⁷\(Q*QRT2_{i,t-1}\) (or \(Q*QRT3_{i,t-1}\), \(Q*QRT4_{i,t-1}\)) equals \(Q_{i,t-1}\) times an indicator variable that equals 1 if \(Q_{i,t-1}\) is in the second (third, fourth) quartile of its industry and year distribution. \(CF_{i,t}\) stands for net cash flows from operations (Compustat data item 308) scaled by the total book value of assets (Compustat data item 6). \({\textit{AG}}_{i,t-1}\) stands for the lag of asset growth, which equals the natural log of the total book value of assets at the end of year \(t-1\) divided by the total book value of assets at the end of year \(t-2\). \({\textit{TI}}_{i,t-1}\) is the lag of total investment.

I estimate Eq. (3) for each industry-year based on the Fama–French 48-industry classification and use the absolute value of the residuals from Eq. (3), standardized by the standard deviations of their corresponding industry and year distributions, as our second measure of investment inefficiency, \({\textit{Inefficiency}}^M\):Table 2 reports the regression results from the investment inefficiency models of Biddle et al. (2009) and McNichols and Stubben (2008). I calculate the cross-sectional means of the coefficient estimates for all industries in a given year and report the time-series average of the cross-sectional means of the coefficient estimates. We can see that all the factors in the models of Biddle et al. (2009) and McNichols and Stubben (2008) are significantly correlated with the firm’s investment level, which is consistent with the assumptions behind these models. Besides the total investment, I also use the investment excluding R &D expenditure, \({\textit{TIW}}\), to estimate Eqs. (1) and (3) and calculate these measures of investment inefficiency. They will be denoted by \({\textit{Inefficiency}}^{BW}\) and \({\textit{Inefficiency}}^{MW}\), respectively, and used for a robustness check.

The primary independent variable is abnormal short interest. I employ the method of Karpoff and Lou (2010) to calculate it. First, I independently sort stocks into three groups by size, market-to-book ratio, and momentum, so that each stock is assigned to one of 27 constructed portfolios (\(3\times 3\times 3\)). I further categorize each stock into industry groups based on Fama and French’s 48-industry classifications. In particular, I define the model as follows:where \({\textit{SI}}_{i,t}\) represents the short interest level of firm i in month t. The first three independent variables are dummy variables that collectively determine the 27 size, market-to-book, and prior month return portfolios. For example, if firm i is sorted into the lowest market-to-book portfolio in month t, then \({\textit{MB}}_{i,{\textit{low}},t}=1\) and \({\textit{MB}}_{i,{\textit{medium}},t}=0\). The industry dummy \({\textit{Ind}}_{ikt}=1\) if firm i is assigned to industry k in month t, and K is the total number of industries in the sample.

I conduct monthly cross-sectional regressions of Eq. (5) and then calculate the time-series averages of the coefficient estimates. I use these time-series averages in conjunction with the dummy variables to derive the expected short interest \(E({\textit{SI}}_{i,t})\) and calculate the abnormal short interest, \({\textit{ASI}}_{i,t}\):In an unpublished table,⁸ I find that the smallest firms have the lowest short interest, growth stocks (with the highest market-to-book ratio) are more heavily shorted, and momentum has a U-shaped relationship with short interest as the coefficients on \({\textit{Momentum}}_{{\textit{low}}}\) and \({\textit{Momentum}}_{{\textit{medium}}}\) have different signs. These findings are consistent with Dechow et al. (2001) and Karpoff and Lou (2010).

The sample consists of stocks listed on NYSE, AMEX, and NASDAQ for the period from 1996 to 2018. Following Fang et al. (2016), I exclude firms in the financial services industry (SIC 6000–6999) and utilities industry (SIC 4900–4949) as the accounting rules, investment purposes, and processes significantly differ for these industries. Several steps are taken to clean the data. I require firms to have available information about investment and various variables needed for the calculation of investment inefficiency to remain in the sample. In particular, I exclude firms without capital expenditure information since it is a major component of total investment. The data used in the calculation of investment inefficiency are retrieved from the Compustat-Fundamentals Annual database. Additionally, following Biddle et al. (2009), I exclude industries with less than 20 observations in a given year. After these screenings, there are 61,947 firm-year observations, which are then used to calculate investment inefficiency.

Next, I match the measure of investment inefficiency with the short interest data and require firms to have available information on short interest before the release date of the annual report to remain in the sample. I retrieve short interest data, defined as the number of all open short positions scaled by total common shares outstanding, from the Compustat-Supplement Short Interest File. Short interest is measured on the 15th of each month or the preceding business day if the 15th is not a business day. The lag between the date of the short interest and the date of the annual release of the financial report should not exceed one month. I impose this lag between the release date of the annual report and the short sale date so that I can test whether short-sellers can detect investment inefficiency ex-ante, in other words, whether they are informed traders who possess superior information.

Finally, I require firms to have available information about various control variables. Data for the control variables are retrieved from the Compustat-Fundamentals Quarterly database because some control variables are measured from the closest quarterly report released before the release date of the annual report. After data cleaning, there are 40,594 firm-year observations with available information about investment inefficiency, short interest, and various control variables, making up the main sample. Appendix shows the variables used in my empirical analysis and how to construct them. Table 1 reports the summary statistics of variables in the whole sample. I report the results of the main dependent and independent variables and other controlling variables. For each variable, I report the number of observations (N), mean, standard deviation (STD), and the 25th percentile, median, and 75th percentile values. I winsorize all variables at the 1 and 99% levels to mitigate the impact of outliers. These variables are used in the following analysis.

In the economic significance analysis, I collect stock return information from the Center for Research on Security Prices (CRSP) database. I require firms to have available stock return information for the past two years since we need to calculate the volatility of the stock return, which equals the standard deviation of monthly stock returns over the past 24 months. Therefore, we have 37,690 firm-year observations in the economic significance analysis.

For the subsample analyses, I use information about board independence and CEO incentive pay. Information about board independence is retrieved from the RiskMetrics database, which covers Standard & Poor’s (S &P) 500, S &P MidCaps, and S &P SmallCap firms. Information about CEO incentive pay is obtained from the ExecuComp database.

Table 2 presents the regression results of the investment inefficiency model. Both models are run for each industry-year based on the Fama-French 48-industry classification. Columns (1) and (2) report the results of the models by Biddle et al. (2009) and McNichols and Stubben (2008), respectively.

As shown in column (1) of Table 2, the coefficient of sales growth is highly significant. Therefore, investment is significantly correlated with output growth, which is consistent with the assumption of acceleration theory. The results in column (2) show that the coefficients of various variables related to investment opportunities are highly significant. These results indicate that investment is strongly correlated with investment opportunities.⁹

In this subsection, I present the results of a univariate test to examine the relation between short-selling and investment inefficiency. If short-sellers are informed about a firm’s investment inefficiency, then short-selling should be more severe for firms with higher investment inefficiency. Following Hirshleifer et al. (2011), I sort these sample firms into deciles based on the abnormal short interest closest to, but before, the annual release of the financial report for each year. The abnormal short interest is measured on the 15th of each month or the preceding business day if the 15th is not a business day. The lag between the date of abnormal short interest and the date of the annual release of the financial report should not exceed one month. I then test whether the mean of investment inefficiency in the top decile of abnormal short interest is significantly higher than the mean of investment inefficiency in the bottom decile. To do this, I first calculate the cross-sectional means of investment inefficiency for the top and the bottom abnormal short interest deciles in each year. I then calculate their difference. After that, I calculate the time-series average of the difference and check whether the mean difference is significantly different from zero.

Table 3 reports the results of the univariate test. I consider two measures of investment inefficiency, \({\textit{Inefficiency}}^B\) and \({\textit{Inefficiency}}^M\), calculated by Eqs. (1) and (3), respectively. Table 3 shows that the means of \({\textit{Inefficiency}}^B\) in the top and bottom deciles of abnormal short interest are 0.689 and 0.611, respectively. The difference in mean is 0.078 and is significant at the 1% level (t-value = 4.43). It accounts for 12.15% of the difference in the standard deviations of \({\textit{Inefficiency}}^B\). The difference in mean of \({\textit{Inefficiency}}^M\) between the top and bottom deciles is 0.075, which is also significant at the 1% level (t-value = 4.45). It accounts for 10.65% of the difference in the standard deviations of \({\textit{Inefficiency}}^M\). These results support the hypothesis that short-sellers are informed about firms’ investment inefficiency. There is a positive relation between abnormal short-selling activity and investment inefficiency.

In this subsection, I present a baseline panel regression using the whole sample data to investigate the relation between short-selling and future investment inefficiency. In particular, I follow Jiao et al. (2016) to run the following panel regressions using all data:where \({\textit{ASI}}_{;i,t-1}\) represents the abnormal short interest before the release of the annual financial statement. \({\textit{Inefficiency}}_{i,t}\) serves as the measure of investment inefficiency. \(X_{i,t-1}\) refers to a list of control variables, including firm size, market-to-book ratio, leverage, and return on assets, all measured with a lag of one quarter. Furthermore, prior research finds a significant correlation between short interest and operating accruals (Hirshleifer et al. 2011; Fang et al. 2016; Massa et al. 2015; Park 2017). I further add operating accruals as control variables.¹⁰\(Y_{i,t-1}\) refers to the list of independent variables used in the investment inefficiency models of Biddle et al. (2009) and McNichols and Stubben (2008). According to Chen et al. (2018), it leads to biased coefficients and standard errors if the researcher decomposes a dependent variable into its predicted and residual components in the first regression and uses the residuals as the dependent variable in a second regression. As discussed above, the investment inefficiency measures take advantage of the residuals in the models of Biddle et al. (2009) and McNichols and Stubben (2008). Hence, there is the concern of biased coefficients and standard errors here because the equation above uses investment inefficiency as the dependent variable. One of the approaches to address this problem is to regress the residuals on the combination of all the independent variables in the first regression and the independent variables in the second regression (Chen et al. 2018). Following this approach, I further add regressors in the models of Biddle et al. (2009) and McNichols and Stubben (2008) as independent variables in the above model.

Table 4 reports the results of the baseline panel regressions. I consider both firm and year fixed effects and cluster the standard errors at the firm level. Columns (1) and (2) report the results of \({\textit{Inefficiency}}^B\) and \({\textit{Inefficiency}}^M\), respectively. The coefficient of abnormal short interest on \({\textit{Inefficiency}}^B\) is 0.003, as shown in column (1), which is significant at the 1% level (t-value = 3.58). The coefficient of abnormal short interest on \({\textit{Inefficiency}}^M\) is 0.003, as shown in column (2), which is also significant at the 1% level (t-value = 3.37). Therefore, the result of the baseline regressions demonstrates that short-selling reflects information about a firm’s investment inefficiency. If a firm is becoming less efficient in investment, then short-sellers react by holding a larger short position in the firm’s stock before this information is made public. In summary, the result of the baseline regression is consistent with that of the univariate test, and both of them support Hypothesis 1.

To spell out why I choose to measure the short-selling activities just before the release of the annual report, I draw the dynamic changes of the coefficient on \({\textit{ASI}}\). I run the baseline model of Eq. (7) but replace \({\textit{ASI}}_{t-1}\) with the abnormal short interest levels 2 months, 4 months, 6 months, and 8 months before the release of the annual financial report. I include the same controlling variables and consider the year and firm fixed effects. I also cluster the standard errors at the firm level. Figure 1 shows that the relationship between abnormal short interest and investment inefficiency is stronger when it is closer to the release of the annual financial report. As short-sellers can observe more information and have a more comprehensive understanding of a firm’s investment inefficiency when it is closer to the fiscal year end, I investigate whether short-sellers are informed about a firm’s investment inefficiency at the time point when the information is considered sufficient and comprehensive.

Figure 1 also shows that the lower bound of the 95% confidence interval for the coefficient on \({\textit{ASI}}\) crosses over zero at the 4-month mark before the annual report release. This suggests that the relationship between short-selling activities and investment inefficiency becomes significant approximately 4 months prior to the annual report release, coinciding with the release of the 3rd quarter report. As discussed in previous sections, the investment inefficiency measures are derived from information disclosed in the annual financial report. This implies that short-sellers gather and analyze information about a firm’s investment behaviors along the way, enabling them to make informed conjectures about a firm’s investment inefficiency one quarter before it becomes evident in the annual financial report.

To address the concern of omitted variables, I further include additional control variables in the model, as shown in Appendix. These variables encompass dividend cuts, competition, momentum, and misstatements in the financial reports. However, the inclusion of these additional variables does not alter the significant relationship between short-selling activities and investment inefficiency.

Texas Industries Inc. (Ticker Symbol: TXI) is a publicly traded company listed on the New York Stock Exchange (NYSE). We analyze key financial metrics, including the mean values of the natural logarithm of market value (\({\textit{Ln}}_{\textit{MV}}\)), the natural logarithm of total book value of assets (\({\textit{Ln}}_{\textit{Assets}}\)), the market-to-book ratio (\({\textit{MB}}_{\textit{Ratio}}\)), leverage (\({\textit{Leverage}}\)), and return on assets (\({\textit{ROA}}\)), over the period from 2005 to 2009. These metrics exhibit the following means during this period: \({\textit{Ln}}_{\textit{MV}}\) (7.081), \({\textit{Ln}}_{\textit{Assets}}\) (7.146), \({\textit{MB}}_{\textit{Ratio}}\) (2.364), \({\textit{Leverage}}\) (0.398), and \({\textit{ROA}}\) (2.436%).

These values collectively suggest that Texas Industries Inc. does not display extreme financial characteristics during this timeframe, indicating a relatively stable performance. However, in the year 2006, the investment inefficiency of the company worsens. Specifically, the measures of investment inefficiency (\({\textit{Inefficiency}}^{B}\) and \({\textit{Inefficiency}}^{M}\)) increase from 0.560 and 0.345 in 2005 to 0.573 and 0.748, respectively. Consequently, I observe an increase in short-selling activities before the annual report release. For instance, the levels of short interest and abnormal short interest rise from 3.627% and − 2.646% in 2005 to 10.251% and 3.932%, respectively, in 2006.

After 2006, the investment inefficiency of Texas Industries Inc. continued to deteriorate until 2009. Consequently, the short-selling activities before the annual report release also increased. Figure 2 depicts the short interest before the annual report release,¹¹ clearly illustrating an ascending pattern from 2005 to 2009. Consistent with expectations, as short-selling activities before the annual report release intensified for Texas Industries Inc. after 2005, the abnormal stock return around the annual report release also turned negative, with a clear decreasing trend evident in Fig. 2

I find a statistically significant positive correlation between short-selling before the annual release of financial information and investment inefficiency. In this subsection, I study whether short-sellers can also use this information to generate economic profits.

To measure the abnormal returns of a stock, I deduct the individual stock returns by the returns of an equal-weighted benchmark portfolio, constructed following Daniel et al. (1997). Specifically, I sort all firms traded on NYSE, AMEX, and NASDAQ, excluding American Depository Receipts (ADRs),¹² into quintiles based on market capitalization. In each quintile of market capitalization, I sort firms into book-to-market ratio quintiles to generate \(5\times 5\) groups. Then, in each of these 25 groups, I sort firms into momentum quintiles using the stock’s past one-year returns while skipping the most recent month. This results in 125 (\(5\times 5\times 5\)) groups, and then I construct the equal-weighted portfolio for each group. The benchmark portfolio for a stock is that of the group to which the stock belongs.

Next, I use the model in Jiao et al. (2016) to run a Fama–MacBeth regression at a yearly frequency:where \({\textit{CAR}}_{i,[t+1,t+6]}\) represents the accumulation over 6 months of abnormal returns. \({\textit{Inefficiency}}_{i,t}\) serves as the measure of investment inefficiency. \(X_{i,t}\) refers to a list of control variables used in Jiao et al. (2016), including the natural logarithm of the stock price (\({\textit{Ln}}\_{\textit{Price}}_{i,t}\)), the natural logarithm of the standard deviation of stock returns over the past 24 months (\({\textit{Ln}}\_{\textit{Volatility}}_{i,t}\)), the natural logarithm of firm age (\({\textit{Ln}}\_{\textit{Age}}_{i,t}\)), the natural logarithm of stock turnover (\({\textit{Ln}}\_{\textit{Turnover}}_{i,t}\)), the dividend (\({\textit{Dividend}}_{i,t}\)), and a dummy variable equal to one if the firm is in the S &P 500 index (\(SP500_{i,t}\)). I also add operating accruals (\({\textit{Operating}};{\textit{Accruals}}_{i,t}\)) as a control variable since it is well-documented in the literature to have return predictability (Hirshleifer et al. 2011).

Table 5 reports the time-series average of the coefficient estimates from the yearly cross-sectional regressions with their t-values (in brackets) adjusted using the Newey–West method with three lags. The result in Table 5 shows that the coefficients of the four measures of investment inefficiency are significantly negative, above the 10% level. This means that investment inefficiency forecasts negative abnormal returns. It is also worth noting that the coefficients of operating accruals are significantly negative, which is consistent with the previous literature, which shows that accruals forecast negative returns. Therefore, the results suggest that short-selling based on information inefficiency information can generate significant profits. Short-sellers increase their short position in the stocks with a higher level of investment inefficiency and receive significant returns from these activities.

I document a significantly positive correlation between short-selling and upcoming firm investment inefficiency information. However, this finding is subject to an endogeneity concern, as unknown factors influence both abnormal short interest and firm investment inefficiency. In other words, abnormal short interest may be correlated with the error term in the regression. To address this potential endogeneity issue, I follow Brav et al. (2018) and Park (2017) in adopting a propensity score matching procedure. This procedure reduces the potential bias that can result either from the omission of observable variables or the specification of an improper functional form for the relation between the observable variables and the outcome variable of interest (Armstrong et al. 2010; Park 2017). In the matched design, each treatment observation is matched with a control observation that does not receive that treatment but is indistinguishable from it in terms of other relevant dimensions. Therefore, any difference in the outcome between the treatment and control groups is attributed to the treatment effect (Armstrong et al. 2010; Park 2017).

To run the propensity score matching, I first rank my sample firms into quintiles based on abnormal short interest before the annual release of the financial statement for each year. I consider the top quintile of abnormal short interest as the treatment group because firms in the top quintile of abnormal short interest are the ones most heavily shorted. The firms in the other groups are treated as control groups. Second, I establish a dummy variable, \({\textit{DASI}}_{i,t}\), that equals one if it is in the top quintile of abnormal short interest and zero otherwise. Third, I run a logistic regression that sets \({\textit{DASI}}_{i,t}\) as the dependent variable:Here, \(X_{i,t}\) constitutes the vector of variables considered as determinants of short-selling activities in previous literature, including firm size, return on assets, market-to-book ratio, momentum, leverage, and operating accruals (Hirshleifer et al. 2011; Karpoff and Lou 2010; Grullon et al. 2015; Diether et al. 2009). Industry and year fixed effects are also controlled for in the regression. The predicted value in the logit regression serves as the propensity score for matching purposes, and one-to-one matching with replacement is employed. In other words, each observation in the top quintile of abnormal short interest is matched with one observation in the control group with the closest score. Observations in the control group may be used multiple times in the matched sample. Additionally, an additional “maximum 1%” constraint is imposed, meaning that matched pairs should have a propensity score difference of less than 1

To assess the effectiveness of the propensity score matching, I compare the means of various firm fundamentals and characteristics between the treatment and control groups before and after matching. Table 6 reports the results. Before matching, significant differences exist between the treatment and control groups in terms of \({\textit{Ln}}\_{\textit{MV}}\), \({\textit{Ln}}\_{\textit{Asset}}\), \({\textit{MB}}_{\textit{ratio}}\), and \({\textit{Leverage}}\). When all sample firms are considered, larger firms, higher market-to-book ratios, and higher leverages are associated with more short-selling. For example, the difference in \({\textit{Ln}}\_{\textit{MV}}\) between the treatment and control groups before matching is 0.343 and significant at the 1% level (t-value = 17.74). The differences in \({\textit{Ln}}\_{\textit{Asset}}\), \({\textit{MB}}_{\textit{ratio}}\), and \({\textit{Leverage}}\) are 0.170, 0.424, and 0.022, respectively, all of which are significant at the 1% level. However, after matching, differences in firm characteristics between the treatment and control groups become indistinguishable. None of the differences are significant after matching. These results suggest that this propensity score matching effectively mitigates endogeneity concerns.

Next, I present the results of the following panel regression using the propensity score matching sample to study the relationship between the level of abnormal short interest and firm investment inefficiency.Here, \(X_{i,t-1}\) constitutes the vector of variables considered as determinants of short-selling activities in previous literature, including firm size, market-to-book ratio, leverage, and return on assets, all measured with a lag of one quarter. \(Y_{i,t-1}\) refers to the list of independent variables used in the investment inefficiency models of Biddle et al. (2009) and McNichols and Stubben (2008). A positive coefficient of \(\beta _1\) suggests that a higher level of abnormal short interest before the release date is associated with a higher level of investment inefficiency. Similarly, I control for industry and year fixed effects and cluster standard errors at the firm level.

Table 7 reports the regression results. Columns (1) and (2) report the results of \({\textit{Inefficiency}}^B\) and \({\textit{Inefficiency}}^M\), respectively. The results show that for the matched samples, the coefficients of \({\textit{DASI}}_{i,t}\) are both significantly positive. The coefficients of \({\textit{DASI}}\) for \({\textit{Inefficiency}}^B\) and \({\textit{Inefficiency}}^M\) are 0.050 and 0.046, respectively. Both of them are significant at the 1% level. These results suggest the positive relationship between abnormal short interest and investment inefficiency is robust after controlling for endogeneity concerns.

Some other alternative explanations for the relationship between short-selling activities and investment inefficiency bring up endogeneity concerns. There are two major alternative explanations. First, weak corporate governance is associated with investment inefficiency due to increased agency problems (Shleifer and Vishny 1997). If short-sellers target firms with weak corporate governance, then corporate governance tends to become endogenous. Second, certain events prior to the annual financial report release may be simultaneously related to short-selling activities and investment inefficiency. However, I suggest that these concerns are considerably minimal.

First, although short-sellers can identify firms with significant internal control weaknesses, the evidence primarily focuses on material internal governance failures (Singer et al. 2022). Considering the large sample size in this paper, the relationship should not be driven solely by these rare instances of material internal governance failure.

More importantly, if short-sellers target firms with material internal governance failure, the significant relationship should appear around the discovery of such failures rather than the financial report release. Furthermore, since material internal governance failures tend to be enduring, the significant relationship should not be limited to the period around the financial report release.

The concern regarding events prior to the financial report release is also considered minimal. First, events preceding the financial report release are reporting-specific, and short-selling activities related to these events occur within a very short period surrounding them. For example, Dai et al. (2021) find that the sharp increase in short-selling activities for firms delaying financial report release focuses on the month prior to the scheduled release.

However, Fig. 1 suggests that short-sellers have begun to detect investment inefficiency four months prior to the financial report release. The short-selling activities related to these reporting-specific events demonstrate a different pattern from that documented in the present paper.

Second, the number of such reporting-specific events is relatively scarce compared to the sample size in the present paper. For instance, the average number of late reporting events each year in Dai et al. (2021) is 127, while the average number of firms each year in the present paper is 1765. The late-reporting events account for only around 7% of the firms in the present study. For the subsequent analysis, I have identified 502 severe misstatement events, which constitute only about 1% of the firms each year in the present study. It is highly unlikely that the significant relationship between short-selling activities and investment inefficiency is biased by these rare cases.

To support this point, I have conducted an additional analysis as shown in Appendix. Short-selling activities are related to serious financial misstatements (Karpoff and Lou 2010), and there is concern that these misstatements are related to investment inefficiency. I have implemented the following model to investigate whether the filing of serious financial misstatements will affect the relationship between short-selling activities and investment inefficiency:where \({\textit{ASI}}_{i,t-1}\) is the abnormal short interest before the release of the annual financial statement. \({\textit{Restatement}}_{i,t}\) is a dummy variable that indicates whether a material inadvertent (unintentional) or fraudulent (intentional) error exists in a firm’s financial statements. \({\textit{ASI}}_{i,t-1} \times {\textit{Restatement}}_{i,t}\) is the interaction term between \({\textit{ASI}}_{i,t-1}\) and \({\textit{Restatement}}_{i,t}\). \({\textit{Inefficiency}}_{i,t}\) is the measure of investment inefficiency. \(X_{i,t-1}\) refers to a list of control variables included in Table 4. \(Y_{i,t-1}\) refers to the list of independent variables that are used in the investment inefficiency models of Biddle et al. (2009) and McNichols and Stubben (2008). I consider both firm and year fixed effects and cluster the standard errors at the firm level.

The result in Appendix shows that the coefficient of \({\textit{ASI}}_{;i,t-1} \times {\textit{Restatement}}_{;i,t}\) is insignificant, indicating that the filing of serious financial misstatement does not affect the relationship between short-selling activities and investment inefficiency.

I use a firm’s total investment (TI) to calculate the measures \({\textit{Inefficiency}}^B\) and \({\textit{Inefficiency}}^M\), which equal the sum of capital expenditures, R &D expenditures, and acquisitions minus the sale of property, plant, and equipment. The literature documents several reasons for separating capital expenditures and R &D expenditures. First, according to the accounting principles generally accepted in the USA (US GAAP), all R &D expenditures must be expensed when incurred, while capital expenditures must be capitalized and depreciated over their economic useful lives (Amir et al. 2007). Therefore, R &D expenses could be used to manage earnings (Baber et al. 1991; Dechow and Sloan 1991; Graham et al. 2005; McNichols and Stubben 2008). For example, Graham et al. (2005) indicate that 80% of executives are willing to decrease R &D to meet an earnings target. As previous literature has highlighted the relation between short sales and earnings management (Desai et al. 2006; Karpoff and Lou 2010; Hirshleifer et al. 2011; Boehmer and Wu 2013; Massa et al. 2015; Fang et al. 2016; Park 2017), to test whether the relation between short sales and investment inefficiency is robust, it is worthwhile to exclude R &D expenditures and focus on physical investments that are unrelated to earnings management. Second, according to Amir et al. (2007), R &D expenditures contribute more to earnings volatility than capital expenditures, especially in R &D-intensive industries. There are fundamental differences between R &D expenditures and capital expenditures, which make it more challenging to investigate the subsequent economic benefits of R &D expenditures than those of capital expenditures.

To test the robustness of the relationship between abnormal short interest and firm investment inefficiency, I use total investment excluding R &D expenditures (TIW) to construct a measure of investment inefficiency. I run the same models as in Eqs. (1) and (3) but use \({\textit{TIW}}\) as the dependent variable. I define the measures of investment inefficiency as \({\textit{Inefficiency}}^{BW}\) and \({\textit{Inefficiency}}^{MW}\) using Biddle et al. (2009) and McNichols and Stubben (2008), respectively.

Table 8 reports the results. Panel A of Table 8 reports the results of the univariate test. The results show that the means of the measures of investment inefficiency for the top deciles of abnormal short interest are significantly higher than those of the low decile groups. The differences between High and Low (H–L) and between the decile 9 and decile 2 portfolios (Decile 9–Decile 2) are all significantly positive. Panel B of Table 8 reports the results of the panel regressions. The left columns report the results using all firms, while the right columns report the results using the matched sample. When I use all firms, the coefficients of \({\textit{ASI}}\) for \({\textit{Inefficiency}}^{BW}\) and \({\textit{Inefficiency}}^{MW}\) are 0.004 and 0.003, respectively. Both of them are significant at the 1% level. When I use the matched sample, the coefficients of \({\textit{DASI}}\) for \({\textit{Inefficiency}}^{BW}\) and \({\textit{Inefficiency}}^{MW}\) are 0.050 and 0.038, respectively. They are significant above the 5% level. These results suggest that the positive relationship between the abnormal short interest level and firm investment inefficiency is robust when excluding R &D expenditures.

I also conduct the economic significance analysis using \({\textit{Inefficiency}}^{BW}\) and \({\textit{Inefficiency}}^{MW}\) and report the results in columns (3) and (4) of Table 5. The results continue to demonstrate significant economic profits from short-selling that utilizes information about investment inefficiency. The coefficients of \({\textit{Inefficiency}}^{BW}\) and \({\textit{Inefficiency}}^{MW}\) on CAR are significantly negative at the 10% level or above.

First, I test the relationship between historical abnormal short interest and investment inefficiency for the subsamples of overinvestment firms and underinvestment firms. Initially, I define overinvestment and underinvestment firms following Chen et al. (2011a). I sort all firms into two groups each year based on the standardized residuals (\(\frac{\epsilon _{i,t}}{\sigma _{j,t}}\)) from Eqs. (1) and (3). Firms in the top group are classified as overinvestment firms, and firms in the bottom group are categorized as underinvestment firms. Then, I run Eq. (7) for both subsamples and present the results in Table 9. The coefficients of \({\textit{ASI}}\) on \({\textit{Inefficiency}}^B\), \({\textit{Inefficiency}}^M\), \({\textit{Inefficiency}}^{BW}\), and \({\textit{Inefficiency}}^{MW}\) are 0.006, 0.005, 0.006, and 0.004, respectively, for the overinvestment firms. All of them are statistically significant at 5% or above. Meanwhile, none of the four coefficients for the underinvestment firms is significant, except for \({\textit{Inefficiency}}^M\). Although the coefficient of \({\textit{ASI}}\) on \({\textit{Inefficiency}}^M\) is significant in the subsample of underinvestment firms, the magnitude and the significance level of the coefficient are lower than those in the subsample of overinvestment firms. The results indicate that the relationship between historical abnormal short interest and investment inefficiency is significant for the overinvestment subsample but insignificant for the subsample of underinvestment firms.

The coefficients of historical abnormal short interest for the subsample of overinvestment firms are also larger than those of the whole sample. For example, the coefficient of \({\textit{ASI}}\) on \({\textit{Inefficiency}}^B\) is 0.003 (\(t=3.58\)) for the whole sample, while it is 0.006 (\(t=3.47\)) for the overinvestment firms. The results using other measures are similar. Overall, the significantly positive relationship between historical abnormal short interest and investment inefficiency only exists for the overinvestment firms, supporting Hypothesis 2.

I test the relationship between historical abnormal short interest and investment inefficiency for the subsamples of firms with little board independence and firms with a high degree of board independence. As an important component of internal governance, board independence can mitigate investment inefficiency (Liu et al. 2015), which makes firms lacking board independence more susceptible to investment inefficiency and more difficult to correct such investment inefficiency.

More importantly, compared with other general governance measures, board independence is linked with short-selling activities (Rahman et al. 2021). Since a lack of board independence will damage disclosure quality and increase information asymmetry, it will facilitate short-selling activities by enhancing the information advantage of short-sellers (Byrd and Hickman 1992; Fama and Jensen 1983; Beasley 1996; Yekini et al. 2015; Rahman et al. 2021).

First, I sort all firms into two groups each year based on board independence. A firm with a board that has a fraction of independent directors above the median is regarded as a firm with high board independence, and a firm with a board that has a fraction of independent directors below the median is regarded as a firm with little board independence. Then, I run Eq. (7) for both subsamples and report the results in Table 10. The results indicate that the relationship between historical abnormal short interest and investment inefficiency is significant for firms with little board independence, as shown in columns (1) to (4), but insignificant for firms with a high degree of independence, as shown in columns (5) to (8). The coefficients of \({\textit{ASI}}\) using \({\textit{Inefficiency}}^B\), \({\textit{Inefficiency}}^M\), \({\textit{Inefficiency}}^{BW}\), and \({\textit{Inefficiency}}^{MW}\) are 0.006, 0.005, 0.006, and 0.004, respectively, for firms with little board independence. All of them are significant at 5% or above. Meanwhile, none of the four coefficients for firms with a high degree of board independence is significant.

The coefficients of historical abnormal short interest for the subsample of firms with little board independence are also larger than those of the whole sample. For example, the coefficient of \({\textit{ASI}}\) using \({\textit{Inefficiency}}^B\) is 0.003 (\(t=3.58\)) for the whole sample, while it is 0.006 (\(t=3.31\)) for firms with little board independence. The results using other measures are similar. Overall, there is only a significantly positive relationship between historical abnormal short interest and investment inefficiency for firms with little board independence, which supports Hypothesis 3.

I will next examine whether the relationship between historical abnormal short interest and investment inefficiency varies with CEO incentive pay. Linking managers’ compensation to firm performance motivates managers to make more value-maximizing decisions (Mehran 1995; Morck et al. 1988; McConnell and Servaes 1990; Jensen and Murphy 1990; Minnick et al. 2011). Short-sellers’ threats are significantly lower if a manager’s compensation is closely related to stock performance, as managers will be diligent and pay attention to market movements.

To investigate whether the relationship between historical abnormal short interest and investment inefficiency depends on CEO incentive pay, I will sort all firms into two groups each year based on CEO incentive pay. CEO incentive pay is calculated bywhere \({\textit{Total}}\_{\textit{CEO}}\_{\textit{compensation}}\) is item TDC1 in the ExecuComp database, which equals the sum of salary, bonus, other annual, the total value of restricted stock granted, the total value of stock options granted calculated using the Black–Scholes option-pricing formula, long-term incentive payouts, and all other totals. A firm with CEO incentive pay above the median is regarded as having high CEO incentive pay, and a firm with CEO incentive pay below the median is regarded as having low CEO incentive pay. Then, Eq. (7) is run for both subsamples. I also introduce firm and year fixed effects and cluster the standard errors at the firm level.

The results in Table 11 indicate that there is only a significant relationship between historical abnormal short interest and investment inefficiency for firms with low CEO incentive pay. For firms with high CEO incentive pay, the relationship is insignificant. Moreover, the coefficients of \({\textit{ASI}}\) for firms with low CEO incentive pay are larger than those of the whole sample. For example, the coefficient of \({\textit{ASI}}\) using \({\textit{Inefficiency}}^B\) is 0.003 (\(t=3.58\)) for the whole sample, while it is 0.005 (\(t=3.09\)) for firms with low CEO incentive pay. Short-sellers focus more on the investment inefficiency of firms with low CEO incentive pay when making short selling decisions. Overall, the results in Table 11 support Hypothesis 4.