A. Individual Measures of Information-Driven Insider Trading

The literature offers a plethora of different methodologies for the identification of potentially information-driven insider trades, covering, for example, insider trading patterns, the timing of insider trades, the general profitability of past insider trades, the nature of the insider position in the company, and general firm characteristics which are associated with greater information asymmetries.

Our goal is not to implement all proposed information-driven measures but to provide a sufficiently wide range of individual measures for the construction of a meaningful composite. In this context, some signals suggested in the U.S.-based literature cannot be implemented appropriately in our global setting due to lack of sufficient data availability.Footnote ⁵ In addition, the individual signals must be consistent in terms of their temporal availability to enter the resulting composite measure with a holding period of 1 month. The baseline composite measure always assumes that in calendar month $ t $ following the information-driven insider trade (signal) in month $ t-1 $ , a long or short position is established according to the expected predictive power of the transaction (signal) direction. The monthly rebalanced trading strategy is based on the actual trade date, disregarding the reporting lag of the transaction. This implies that the strategy mimics information-driven insider trades with some delay for trades that occur at the beginning of month $ t-1 $ . Note that an outside investor would potentially not be able to mimic insider trades that occur at the end of month $ t-1 $ or trades with greater reporting lags, as those trades would not be observable in real time at the beginning of month $ t $ (see Table 1 for the median reporting lags of each country). Consequently, we focus more on investigating the existence of insiders’ short-term informational advantages rather than the actual real-time implementability (see Section V.D).

Given the aforementioned constraints, we implement the following individual measures to identify information-driven insider trades and provide a brief description of the particular economic logic and/or intuition behind the identification:

• • “Opportunistic trades” (Ali and Hirshleifer (Reference Ali and Hirshleifer2017)): We assume that particularly profitable trades shortly before quarterly earnings announcements (QEAs) are more likely to be information-driven trades (Q3).

• • “Strong trades,” “Short horizon trades,” and “Unexpected trades” (Akbas et al. (Reference Akbas, Jiang and Koch2020)): We assume that insiders with a high trading frequency, and therefore a short investment horizon, are more likely to execute information-driven trades (SHOR). We apply a similar rationale to insider trades with greater market impact, that is, high daily volume (STRO), and insider trades that are unexpected given the regular trading behavior of an insider (UNEX).

• • “Insider silence” (Hong and Li (Reference Hong and Li2019)): We assume that sudden insider silence, that is, discontinuing to purchase (PPN) or sell (SSN), following several consecutive purchases/sales, is informative.

• • “Realized loss sales” (Kelly (Reference Kelly2018)): We assume that a sale at a loss is more “painful” than a sale at a gain, and thus it is assumed to have a higher (negative) information content (LOSS).

• • “CFO purchases” (Wang, Shin, and Francis (Reference Wang, Shin and Francis2012)): We assume insider purchases by CFOs to be more likely information-driven, as CFOs are in the best position among corporate insiders to have access to informational advantages and also to use them effectively (CFO).

• • “Accrual trades” (e.g., Beneish and Vargus (Reference Beneish and Vargus2002), Bergstresser and Philippon (Reference Bergstresser and Philippon2006)): We assume that insider transactions by top-level executives are more likely to be information-driven if all company insiders as a group trade in the same direction, while the level of accruals in the last fiscal year suggests a possible manipulation of accruals in favor of the transaction direction (ACC).

• • “Research & Development trades” (Aboody and Lev (Reference Aboody and Lev2000)): We assume that higher R&D expenditures lead to higher information asymmetries regarding future developments, which increase the likelihood of information-driven insider trades by corporate officers (R&D).

• • “Idiosyncratic volatility trades” (e.g., Jagolinzer et al. (Reference Jagolinzer, Larcker and Taylor2011), Ben-David et al. (Reference Ben-David, Kim, Moussawi and Roulstone2023)): We assume that higher stock-level idiosyncratic volatility leads to higher information asymmetries regarding future developments or to more pronounced mispricing, both of which increase the likelihood of information-driven insider trades (IVOL).

• • “Clustered trades” (Alldredge and Blank (Reference Alldredge and Blank2019)): We assume that insider trades in the same direction made by multiple insiders of the same firm within a certain temporal proximity are more likely to be information-driven, as this synchronous trading suggests that the same (shared) information might be exploited by multiple insiders (CLUS).

• • “Analyst coverage trades” (e.g., Frankel and Li (Reference Frankel and Li2004), Ellul and Panayides (Reference Ellul and Panayides2018)): We assume that an unusually high degree of analyst coverage (controlling for firm size) may indicate greater public interest in and scrutiny of the firm, which makes it harder to exploit nonpublic information for insiders. On the contrary, less attention from analysts might result in more information-driven trades (ANA).

• • “Residual media coverage trades” and “Selected residual media coverage trades” (e.g., Dai, Parwada, and Zhang (Reference Dai, Parwada and Zhang2015), Sun, Dutta, Huang, and Zhu (Reference Sun, Dutta, Huang and Zhu2021), and Kyung and Nam (Reference Kyung and Nam2023)): We assume that an unusually high degree of media coverage (controlling for firm size) may indicate greater public interest in and scrutiny of the firm, which makes it harder to exploit nonpublic information for insiders. On the contrary, less press attention could result in more information-driven trades (RMC). We apply the same rationale to narrower press coverage dealing with insider trades and related topics (SRMC) such as corporate responsibility or investor relations.

• • “Press release trades” (e.g., Cheng and Lo (Reference Cheng and Lo2006)): We assume that voluntary firm-initiated news disclosure, that is, press releases, in advance of a planned insider transaction could be motivated to influence the stock price in a favorable direction (VOLD).

• • “Conditional conservatism trades” (Khalilov and Osma (Reference Khalilov and Osma2020)): We assume it to be more difficult for insiders to speculate on negative news when these are incorporated in the accounting numbers in a timely manner, that is, in times of high conditional accounting conservatism. Therefore, we regard sales of insiders as more likely to be information-driven if the company has low conditional conservatism. On the other hand, we assume that high conditional conservatism leads to a potential undervaluation, as profits are not recognized until the associated cash flows are realized, which offers the opportunity to speculate on good news (CONS).

• • “Multiple firm insider sales” (Karamanou, Pownall, and Prakash (Reference Karamanou, Pownall and Prakash2021)): We assume that when an insider who is active in multiple firms executes a sale transaction in one firm that is accompanied by at least one purchase transaction in one of his or her other firms in the same month, it is less likely that the sale transaction is liquidity-driven. Therefore, we assume the sale to be information-driven (MFS).

• • “Persistently profitable trades” (Cline, Gokkaya, and Liu (Reference Cline, Gokkaya and Liu2017)): We assume that insiders with persistently high abnormal returns might exploit information (PROF).

Following economic logic, we interpret sell (buy) transactions as a negative (positive) signal for future performance. We provide a detailed description of the implementation of all information-driven trade measures adapted to our global sample in Table OA.4 of the Supplementary Material. Tables OA.5 and OA.6 of the Supplementary Material provide descriptive statistics for the appearance of each information-driven buy and sell signal across countries, respectively.

B. Composite Measures of Information-Driven Insider Trading

The likelihood that information about a stock’s future performance can reliably be drawn from insider trades presumably increases with the number of trade signals. In addition, individual information-driven signals are not necessarily unambiguous. Hence, different measures can indicate opposing trade directions. Composition rules out such ambiguities by requiring a certain number of unique signals to be considered relevant. We classify each firm-month as information-driven (Buy = 1/Sell = −1) or non-information-driven (0) using the following composition methodology based on a total of 16 buy measures and 17 sell measures:

(1) $$ {CID}_{i,t}^N=\left\{\begin{array}{ll}1& \mathrm{if}{\sum}_{j=1}^{16} Buy{Signal}_{j,i,t}-{\sum}_{k=1}^{17} Sell{Signal}_{k,i,t}\ge \hskip1em N\hskip1em (Buy)\\ {}-1& \mathrm{if}{\sum}_{j=1}^{16} Buy{Signal}_{j,i,t}-{\sum}_{k=1}^{17} Sell{Signal}_{k,i,t}\le -N\hskip1em (Sell)\\ {}0& \mathrm{else},\end{array}\right. $$

where $ Buy{Signal}_{j,i,t} $ is a dummy variable equal to 1 if any information-driven insider buy occurred in firm $ i $ during month $ t $ according to the individual information-driven buy measure $ j $ , $ Sell{Signal}_{k,i,t} $ is a dummy variable equal to 1 if any information-driven insider sell occurred in firm $ i $ during month $ t $ according to the individual information-driven sell measure $ k $ , and $ N $ is the required surplus needed to qualify as an information-driven composite buy ( $ CID=1 $ ) or sell ( $ CID=-1 $ ) signal, respectively. In the baseline approach, we use $ N=2 $ , which is denoted $ {CID}^2 $ , for brevity henceforth $ CID $ . For example, we would consider a firm-month to have an information-driven buy signal if three information-driven buy signals and no more than one information-driven sell signal occurred in that month.

To better understand the operation of $ CID $ in equation (1), Figure 1 illustrates how many trade signals are generated in the pooled country data set depending on the surplus requirements ( $ N\ge \mathrm{1,2,3,4,5} $ ) of an information-driven insider trading signal. Graphs A and B clearly show that, for both purchase and sale signals, the absolute threshold of 1, that is, an excess of one purchase ( $ N\ge 1 $ ) or sale ( $ N\le -1 $ ) signal, does not lead to a meaningful signal reduction compared to the number of unconditional signals. As surplus requirements increase, correspondingly clearer signals, that is, larger absolute surpluses of individual information-driven purchase or sale signals, are generated. However, a clearer signal also substantially decreases the sample size. Considering this trade-off, we use a $ CID $ greater than or equal to 2 ( $ {CID}^2 $ ) as the baseline approach. This choice is also motivated by the heterogeneous appearance of individual information-driven signals both across measures and across countries (see Tables OA.5 and OA.6 of the Supplementary Material). As such, a $ CID $ greater than or equal to 2 prevents the composite signal from being driven by the more prevalent individual information-driven purchase or sale signals. Consequently, combinations of less common signals can also generate a valid composite information-driven signal, thus enabling the use of the full breadth of individual measures. A $ CID $ greater than or equal to 3, 4, or 5 is increasingly driven by the most prevalent signals and could potentially not be implementable in countries with low or poorer signal appearances. We report the results of surplus requirements ( $ N $ ) greater than or equal to 3 ( $ {CID}^3 $ ) and $ 4 $ ( $ {CID}^4 $ ) in Section V.A.

FIGURE 1Composite Firm-Month Signals

Figure 1 shows the total number of firm-month signals over the sample period from 2000 to 2021 for the pooled country sample. We show unconditional trade firm-month signals, that is, all months in which at least one insider traded. Furthermore, we report composite firm-month signals greater than or equal to ( $ N\ge $ ) 1, 2, 3, 4, 5, respectively. A composite firm-month signal (see equation (1)) greater than or equal to 2 shows the number of firm-month with a surplus, that is, absolute value of individual information-driven purchases—individual information-driven sale firm-month signals being greater than or equal to 2 ( $ N\ge 2 $ ). Graphs A and B show purchases and sales, respectively. We only consider firm-month signals with a minimum absolute surplus of 2 as information-driven (the area to the right of the dashed red line).

FIGURE 1 Long description

Two vertically stacked bar charts.

Graph A, titled Purchases, has a Y-axis labeled Firm Month Signals in 1,000 ranging from 0 to 500. The X-axis lists levels 1, 2, 3, 4, and 5. Each level features a stacked bar where the light gray portion represents C I D Buy Level greater than or equal to N and the dark gray portion represents Unconditional Buy. As the level N increases from 1 to 5, the light gray portion decreases significantly from nearly 450 to approximately 50. A vertical dashed red line is positioned between level 1 and level 2, indicating the threshold for information-driven signals.

Graph B, titled Sales, has a Y-axis labeled Firm Month Signals in 1,000 ranging from 0 to 600. The X-axis lists levels 1, 2, 3, 4, and 5. The light gray portion represents C I D Sell Level greater than or equal to N and the dark gray portion represents Unconditional Sell. Similar to Graph A, the light gray portion decreases as N increases, starting near 580 at level 1 and dropping to approximately 150 at level 5. A vertical dashed red line is also placed between level 1 and level 2.

A. Unconditional Insider Trading

We start by analyzing unconditional (for brevity henceforth $ UNC $ ) insider trades as a benchmark. More precisely, we buy stocks in month $ t $ that were bought by (any) insiders in the previous month $ t-1 $ and short-sell stocks that were sold, respectively. We use the actual trade date to determine whether an unconditional signal was generated in month $ t-1 $ , disregarding the reporting lag of the transaction. Thus, the $ UNC $ strategy mimics unconditional insider trades with some delay for insider trades that occurred at the beginning of month $ t-1 $ . The $ UNC $ strategy considers all individual insider signals before generating a firm-month signal. Thus, we hold a firm in both the long and short legs in month $ t $ if at least one insider in that firm was a net buyer and another insider a net seller in month $ t-1 $ .

Table 2 shows the raw and abnormal returns (6-factor alphas following Fama and French (Reference Fama and French2018), henceforth FF6 alphas) of the $ UNC $ strategy as well as the performance of the long and short legs separately for each country. The t-statistics are Newey and West (Reference Newey and West1987) adjusted.Footnote ⁶

TABLE 2Unconditional Performance

TABLE 2 Long description

The table is structured with three main performance categories: Long-Short, Long, and Short portfolios. Each category is further divided into Raw Returns (V W and E W), F F 6 alpha (V W and E W), the number of months N, and the average number of firms F.

Key data points include:

* Australia: Long-Short E W Raw Return of 0.973 with a t-statistic of 3.43, significant at the 1% level.

* Canada: Highest average number of firms in the Long-Short portfolio at 722.

* United States: Long-Short E W alpha of 0.912 with a t-statistic of 6.03, significant at the 1% level; average of 1,666 firms.

* Turkey: High Long-Short E W Raw Return of 2.400 and alpha of 2.220, both highly significant.

* South Korea: Long-Short E W Raw Return of 2.032 and alpha of 1.703, both highly significant.

The bottom rows summarize totals:

* Total greater than 0 [Significant]: For Long-Short E W alpha, 34 countries were positive, with 28 being statistically significant.

* Total less than 0 [Significant]: For Short E W alpha, 18 countries were negative, with 6 being statistically significant.

t-statistics are provided in parentheses below each return value. Asterisks indicate significance levels: * for 10%, ** for 5%, and *** for 1%.

Only eight (eight) countries exhibit significantly positive value-weighted raw (abnormal) returns, respectively. The only countries whose raw and abnormal returns are both significantly greater than 0 at the 5% level are France, Germany, Malaysia, Poland, South Korea, Spain, and Turkey. The results are substantially stronger for equal-weighted $ UNC $ portfolios, where 26 (raw) and 28 (abnormal) of the $ UNC $ equal-weighted long–short returns are positively significant at the 5% level, respectively. The only countries that exhibit insignificant equal-weighted abnormal returns are Denmark, Egypt, Finland, Indonesia, Israel, and the Philippines. In general, the return dispersion is large, both across value-weighted and equal-weighted strategies.

Almost all long legs have positive raw returns and alphas. The performance is especially strong among equal-weighted $ UNC $ portfolios, where almost all alphas are statistically significantly positive. The patterns in the short leg are less pronounced. The returns and alphas tend to be more dispersed, but they are statistically insignificant in the majority of cases.

Overall, the profitability of the $ UNC $ strategies appears to be limited for value-weighted portfolios. However, there seems to be predictive power for equal-weighted unconditional insider trades in most countries.

B. Composite Measures of Information-Driven Insider Trading

We use the composite approach developed in Section III.B to identify information-driven insider trades. Following the same timing and portfolio construction as described in Section IV.A for the $ UNC $ strategy, we construct the $ CID $ strategy by buying (selling) a stock if the difference between information-driven insider buying and selling signals in the previous month is greater than or equal to 2 (less than or equal to −2). Table 3 shows the monthly raw returns and FF6 alphas for the $ CID $ trading strategy separately for each country.Footnote ⁷

TABLE 3Composite Performance

TABLE 3 Long description

The table reports summary performance statistics for a monthly rebalanced insider trading strategy based on a composite measure C I D.

Header Structure:

* The table is divided into three main portfolio sections: Long-Short, Long, and Short.

* Each section includes columns for Raw Returns (V W and E W), F F 6 alpha (V W and E W), number of months N, and average number of firms F.

* V W stands for Value-Weighted and E W stands for Equal-Weighted.

Data Rows:

* The table lists 34 countries including Australia, Belgium, Brazil, Canada, Chile, China, Denmark, Egypt, Finland, France, Germany, Greece, Hong Kong, India, Indonesia, Israel, Italy, Malaysia, Netherlands, Norway, Pakistan, Philippines, Poland, Singapore, South Africa, South Korea, Spain, Sri Lanka, Sweden, Switzerland, Thailand, Turkey, United Kingdom, and United States.

* For each country, two rows of data are provided: the top row contains the return and alpha values with significance markers (asterisks), and the bottom row contains Newey and West adjusted t-statistics in parentheses.

* Statistical significance is indicated by * for 10 percent, ** for 5 percent, and *** for 1 percent levels.

Summary Statistics at the Bottom:

* Total greater than 0 [Sig.]: Reports the count of positive performances and the number of those that are statistically significant. For example, in the Long-Short E W Raw Returns column, 33 countries had positive returns, 22 of which were significant.

* Total less than 0 [Sig.]: Reports the count of negative performances and the number of those that are statistically significant. For example, in the Short E W F F 6 alpha column, 21 countries had negative alphas, 7 of which were significant.

In total, 9 (10) of the value-weighted raw (abnormal) $ CID $ returns are significantly positive at the 5% level. For the equal-weighted $ CID $ portfolios, we observe a clearer picture with 22 and 23 statistically significant positive raw and abnormal returns at the 5% level, respectively.

To identify the performance drivers of the $ CID $ strategy, we analyze the long and short legs separately. Almost all of the value-weighted raw and abnormal returns of the long legs are positive. Furthermore, 15 and 16 of these value-weighted raw and abnormal $ CID $ returns are significant at the 5% level. For equal-weighted portfolios, we observe greater economic magnitude and significance of abnormal returns. In total, 29 countries show significantly positive (at the 5% level) abnormal returns in the long leg. Looking at the short legs, there are only five and seven statistically significant negative value- and equal-weighted abnormal returns at the 5% level, respectively.

Table OA.7 of the Supplementary Material provides an overview of the performance differences at the country level by investing in a portfolio that goes long the respective composite $ CID $ approach and short the unconditional $ UNC $ benchmark strategy. The comparison at the country level does not provide a conclusive picture. Noticeably, at least three-quarters of the countries show positive improvements (point estimates) in value- and equal-weighted raw and abnormal performance. However, only one-third of the countries show significant positive improvements for equal-weighted raw and abnormal returns, while seven and six countries show significant positive improvements for value-weighted raw and abnormal returns.

In the following, to provide a more general and potentially more meaningful perspective on the improvements produced by the $ CID $ strategy, we analyze the performance of all countries in aggregate regions. In Table 4, we report value- and equal-weighted FF6 alphas for the $ UNC $ and $ CID $ strategies for the long–short (Panel A), long (Panel B), and short (Panel C) portfolios, respectively. Furthermore, we report alphas for a long–short strategy ( $ CID- UNC $ ) that goes long the respective $ CID $ long–short, long, and short strategies and short the respective $ UNC $ long–short, long, and short strategies. To account for structural differences between countries, we distinguish between developed markets (DM) and emerging markets (EM). We use the following two aggregation methods: We aggregate all firm-months across all countries in the respective region (DM or EM) to create a regional monthly sample that includes all information-driven (or unconditional) insider trade firm-month signals. This “pooled” sample is dominated by countries with a greater number of information-driven (unconditional) firm-month observations. To weaken the influence of countries with many monthly signals, we additionally calculate “country-neutral” regional results by averaging the respective $ UNC $ , $ CID $ , and $ CID- UNC $ monthly country strategy return time series in each month. Thus, in the “country-neutral” approach, each country with a valid monthly return receives the same weight in the month under consideration. For both approaches, we require at least 15 stocks per month for a valid long (short) portfolio construction. All abnormal returns for the “pooled” and “country-neutral” regions are calculated using the regional U.S. dollar Fama and French (Reference Fama and French2018) 6-factor model obtained from Kenneth French’s data library.

TABLE 4 CID − UNC Performance Difference Alphas

TABLE 4 Long description

The table is divided into three sections: Panel A for Long-Short, Panel B for Long, and Panel C for Short strategies. Each panel compares U N C F F 6 alpha, C I D F F 6 alpha, and the difference C I D minus U N C F F 6 alpha.

* Panel A (Long-Short): For Pooled D M, the C I D minus U N C difference is 0.275 for V W and 0.286 for E W, both significant at the 1 percent level. For Pooled E M, the difference is 0.323 for V W and 0.383 for E W. Under Country-Neutral implementations, D M shows a difference of 0.136 for V W and 0.282 for E W, while E M shows 0.332 for V W and 0.269 for E W.

* Panel B (Long): For Pooled D M, the C I D minus U N C difference is 0.255 for V W and 0.210 for E W. For Pooled E M, the difference is 0.181 for V W and 0.109 for E W. Under Country-Neutral, D M shows 0.126 for V W and 0.130 for E W, while E M shows 0.282 for V W and 0.010 for E W.

* Panel C (Short): For Pooled D M, the C I D minus U N C difference is minus 0.019 for V W and minus 0.076 for E W. For Pooled E M, the difference is minus 0.142 for V W and minus 0.275 for E W. Under Country-Neutral, D M shows minus 0.009 for V W and minus 0.173 for E W, while E M shows minus 0.139 for V W and minus 0.273 for E W.

All panels report N equals 227 months. T-statistics are provided in parentheses below each alpha value, with asterisks indicating statistical significance levels at 10 percent, 5 percent, and 1 percent.

Panel A of Table 4 reports the FF6 long–short alphas for the value- and equal-weighted $ UNC $ and $ CID $ strategies, respectively. In the pooled DM sample, the long–short returns for value- and equal-weighted $ UNC $ portfolios are −0.05% and 0.72%, respectively. The corresponding returns for the $ CID $ DM strategies are 0.23% and 1.01%, respectively. In both cases, only equal-weighted portfolios show significant alphas. The value-weighted (0.28%) and equal-weighted (0.29%) $ CID- UNC $ alphas are significantly positive, indicating an outperformance of the $ CID $ strategies in the pooled DM samples. For the pooled EM samples, we observe similar results with higher absolute alphas. The value- and equal-weighted $ UNC $ alphas are 0.47% and 1.00%, respectively. The corresponding EM $ CID $ alphas are higher with 0.79% and 1.39% for the value- and equal-weighted portfolios, respectively. The value-weighted and equal-weighted $ CID- UNC $ alphas are 0.32% and 0.38%, statistically significant at the 5% and 1% levels, respectively. In sum, the pooled DM and EM samples reveal a significant outperformance of the $ CID $ strategies.

In the country-neutral DM samples, the value- and equal-weighted long–short $ UNC $ ( $ CID $ ) alphas are 0.28% and 0.88% (0.42% and 1.17%), all of which are significant at the 1% level. Furthermore, the $ CID- UNC $ differences are significant for both equal- (0.28%) and value-weighted (0.14%) DM portfolios. Similarly, for EM, both the country-neutral value-weighted (0.33%) and equal-weighted (0.27%) $ CID- UNC $ alpha differences are significant. In sum, the alternative aggregation method, that is, country-neutral DM and EM portfolios, also suggests a significant outperformance of the $ CID $ strategies.

Relative to the more mixed country-level results in Tables 2 and 3 and OA.7 of the Supplementary Material, the high statistical significance of the pooled and country-neutral regional strategies can be explained by international diversification benefits. More precisely, most country-level $ UNC $ , $ CID $ , and $ CID- UNC $ strategies exhibit low correlation between countries. These findings suggest that informed insider trading tends to be based rather on firm-specific (or country-specific or industry-specific) information as opposed to market or regional information.

Panels B and C of Table 4 report the FF6 alphas for the long and short legs separately. Noticeably, for equal-weighted portfolios, we observe significant $ CID- UNC $ alphas for almost all pooled and country-neutral long–short, long, and short portfolios, respectively. This finding suggests that among small firms, the outperformance of $ CID $ relative to the unconditional strategy stems from its ability to identify both information-driven buys and sells. Overall, however, the significant $ UNC $ , $ CID $ , and $ CID- UNC $ long–short alphas are often driven by their long legs, suggesting that insider buying generally has more return predictive ability. Table OA.8 of the Supplementary Material shows similar results for raw returns. In Table OA.9 of the Supplementary Material, we benchmark the $ CID $ strategy against the performance obtained from averaging the returns of the individual presumably information-driven strategies outlined in Section III.A.Footnote ⁸ For this high-hurdle robustness check, we find a statistically significant and economically relevant (often several percent p.a.) superiority of the $ CID $ strategy only for equal-weighted portfolios. Notably, the outperformance can be found in all specifications with equal-weighted returns, and it increases with the extent of the signal surplus of the composite strategy (see also Section V.A). In summary, the advantages of signal synthesis materialize in both developed markets and emerging markets but are predominantly found among smaller stocks. In the following, we examine potential statistical and economic mechanisms behind the performance of the $ CID $ strategy in more detail.

A. More Restrictive Composition Approaches

We test more restrictive versions of the ( $ {CID}^2 $ ) measure, that is, signal surplus requirements larger than $ N=2 $ . Specifically, we set $ N=3 $ or $ N=4 $ (see equation (1)). Note that even higher surplus requirements, such as $ N=5 $ , would not be consistently possible due to the often sparse and heterogeneous signal appearances across countries (see Figure 1).

Table 5 shows the resulting performance statistics for pooled and country-neutral developed (DM) and emerging market (EM) samples. The results show that the more restrictive versions ( $ {CID}^3 $ and $ {CID}^4 $ ) of the composite measure of information-driven trades uniformly, that is, regardless of the country aggregation scheme, show significant positive abnormal long–short returns for equal-weighted portfolios. However, pooled DM value-weighted $ {CID}^3 $ and $ {CID}^4 $ portfolios and the pooled EM value-weighted $ {CID}^4 $ portfolio show insignificant (but still positive) long–short FF6 alpha estimates, which is largely consistent with the baseline measure $ {CID}^2 $ in Panel A of Table 4. Evaluating the effectiveness of $ {CID}^3 $ and $ {CID}^4 $ to improve on the unconditional predictability by looking at the $ {CID}^3- UNC $ and $ {CID}^4- UNC $ alphas shows substantial differences across samples. These findings are broadly comparable to the baseline measure $ {CID}^2 $ in Panel A of Table 4.

TABLE 5More Restrictive Composition Approaches

TABLE 5 Long description

The table is divided into two main horizontal sections: Pooled and Country-Neutral implementations. Each section contains data for D M (Developed Markets) and E M (Emerging Markets) across four primary metric groups: C I D super 3 F F 6 alpha, C I D super 3 minus U N C F F 6 alpha, C I D super 4 F F 6 alpha, and C I D super 4 minus U N C F F 6 alpha. Metrics include V W (Value-Weighted) and E W (Equal-Weighted) alphas, with t-statistics in parentheses and month counts (N).

Key data points from the Pooled section:

* D M: C I D super 3 V W alpha is 0.334 (1.86) and E W is 1.086 (5.35). C I D super 4 V W alpha is 0.240 (0.95) and E W is 1.269 (5.12).

* E M: C I D super 3 V W alpha is 0.909 (2.79) and E W is 1.430 (8.98). C I D super 4 V W alpha is 0.745 (1.57) and E W is 1.497 (10.63).

Key data points from the Country-Neutral section:

* D M: C I D super 3 V W alpha is 0.565 (4.76) and E W is 1.371 (12.83). C I D super 4 V W alpha is 0.674 (3.60) and E W is 1.769 (6.15).

* E M: C I D super 3 V W alpha is 0.965 (4.46) and E W is 1.359 (8.06). C I D super 4 V W alpha is 1.328 (4.50) and E W is 1.601 (9.44).

Statistical significance is marked by asterisks: * for 10 percent, ** for 5 percent, and *** for 1 percent.

Graph A of Figure 2 shows the difference between the unconditional ( $ UNC $ ) and all versions of $ CID $ ( $ {CID}^2 $ , $ {CID}^3 $ , and $ {CID}^4 $ ) alphas for the pooled DM and EM samples. Graph A shows that a more restrictive $ CID $ monotonically increases the benefits of our $ CID $ strategy relative to the $ UNC $ strategy for the equal-weighted pooled DM sample. However, the equal-weighted pooled EM sample shows increased benefits from $ {CID}^2 $ to $ {CID}^3 $ , but not additional benefits for the $ {CID}^4 $ strategy. The value-weighted pooled EM and DM samples show similar increased benefits from $ {CID}^2 $ to $ {CID}^3 $ , but reversals back to comparable $ {CID}^2 $ alphas for the $ {CID}^4 $ strategy.

FIGURE 2Composition Benefits

Figure 2 shows the benefits of our composite measures by displaying the monthly value- (VW) and equal-weighted (EW) FF6 alphas of long–short strategies that go long the long–short composition measure ( $ {CID}^2 $ , $ {CID}^3 $ , $ {CID}^4 $ ) and short the long–short unconditional strategy ( $ UNC $ ) for pooled developed (DM) and emerging market (EM) samples in Graph A. Graph B shows the potential benefits of a change in holding periods for our baseline composite measure $ {CID}^2 $ by displaying the monthly equal- and value-weighted FF6 alphas of long–short strategies that go long the long–short composition measure ( $ {CID}^2 $ ) and short the long–short unconditional strategy ( $ UNC $ ) for various holding periods (1, 2, 3, 6, and 12 months) for pooled developed (DM) and emerging market (EM) samples. Statistical significance at the 10%, 5%, and 1% level is indicated by *, **, and ***, respectively.

FIGURE 2 Long description

The figure consists of two vertically stacked panels, Graph A and Graph B. Both panels use a Y-axis labeled Average Monthly alpha in percentage and share a legend where dark gray bars represent V W alpha and light gray bars represent E W alpha.

Graph A, titled Composition Approaches, compares three methods: C I D super 2, C I D super 3, and C I D super 4.

- In the D M (Developed Markets) section, alpha values increase as the composition power increases, with C I D super 4 reaching approximately 0.55 percent for E W.

- In the E M (Emerging Markets) section, alpha values are more stable across methods, peaking around 0.45 percent for C I D super 3.

Graph B, titled Holding Periods, focuses on the C I D super 2 method across five timeframes: 1 M, 2 M, 3 M, 6 M, and 12 M.

- In both D M and E M, there is a clear downward trend where alpha decreases as the holding period lengthens.

- For D M, the 1 M alpha starts near 0.3 percent and drops to below 0.1 percent at 12 M.

- For E M, the 1 M alpha starts higher at nearly 0.4 percent and similarly declines toward 0.1 percent at 12 M.

Asterisks above bars indicate statistical significance: three asterisks for 1 percent, two for 5 percent, and one for 10 percent.

In contrast, the value- and equal-weighted DM and EM country-neutral portfolios show that more restrictive $ CID $ strategies outperform more or less monotonically. We conjecture that the different pooled versus country-neutral results indicate that increasing the $ CID $ restrictions works better on a country-level basis and has a somewhat limited ability to increase the benefits in our value-weighted pooled portfolios, which are United States (DM) and China/India (EM) dominated. Overall, the results indicate that the baseline results are robust to the choice of the surplus requirement of the composite measure ( $ N $ in equation (1)).

B. Composition Approaches Using Machine Learning

Can more complex methods enhance return predictability? Following Gu, Kelly, and Xiu (Reference Gu, Kelly and Xiu2020), Drobetz and Otto (Reference Drobetz and Otto2021), Leippold, Wang, and Zhou (Reference Leippold, Wang and Zhou2022), and Hanauer, Soebhag, Stam, and Hoogteijling (Reference Hanauer, Soebhag, Stam and Hoogteijling2025), among others, we implement a variety of machine learning techniques. Specifically, we employ ordinary least squares ( $ OLS $ ), least absolute shrinkage and selection operator ( $ LASSO $ ), elastic net ( $ ENET $ ), random forest ( $ RF $ ), gradient boosting regression trees ( $ GBR $ ), and neural networks with three layers ( $ NN3 $ ). For all algorithms, we apply an expanding-window approach with disjoint training and testing samples. The initial training period spans 5 years and expands annually. Stock portfolios are formed monthly by sorting stocks into terciles based on their predicted returns. Unlike most of the machine learning literature (e.g., Rasekhschaffe and Jones (Reference Rasekhschaffe and Jones2019), Freyberger, Neuhierl, and Weber (Reference Freyberger, Neuhierl and Weber2020), Tobek and Hronec (Reference Tobek and Hronec2021), and Hanauer and Kalsbach (Reference Hanauer and Kalsbach2023)), we do not use quintile or decile splits in the baseline analysis. Tercile splits result in similar portfolio sizes as in the $ CID $ approaches and thus allow for better comparability.Footnote ⁹ We report quintile split results in Table OA.11 of the Supplementary Material. Further details on the implementation of the machine learning models are provided in Section OA.2 of the Supplementary Material.

Table 6 shows the value- and equal-weighted FF6 alphas for the machine learning-based long–short portfolios ( $ ML $ ), along with alphas for long–short strategies ( $ ML- UNC $ ) that go long the long–short machine learning strategies and short the long–short unconditional strategies. We implement the strategies separately for pooled developed (DM) and emerging markets (EM) samples. We do not consider country-neutral approaches, as the limited data, especially in the smaller countries, often do not satisfy the requirements for adequate training of the respective models. Due to the (at least) 5-year training period, the sample period starts in 2008 instead of 2003. This sample period adjustment also affects $ CID $ and $ CID- UNC $ alphas (see Table 4) that act as a benchmark. These values are therefore shown at the bottom of Table 6. For DM (EM), the value- and equal-weighted FF6 $ CID $ alphas change to −0.02% and 0.78% (0.46% and 1.28%). The $ CID- UNC $ alphas remain statistically significant at the 1% level for equal-weighted portfolios, amounting to 0.21% in DM and 0.25% in EM. For value-weighted portfolios, however, the $ CID- UNC $ alpha spreads become statistically insignificant, at 0.12% (DM) and 0.14% (EM), respectively.

TABLE 6Alternative Composition Approaches Using Machine Learning

TABLE 6 Long description

The table reports monthly value-weighted (V W) and equal-weighted (E W) F F 6 factor alphas and Newey-West adjusted t-statistics for pooled Developed Market (D M) and Emerging Market (E M) implementations. All models use N = 168 months.

* O L S: For D M, E W alpha is 0.927 (t = 4.98). For E M, E W alpha is 0.835 (t = 5.17). O L S minus U N C shows a D M E W alpha of 0.365 (t = 1.99).

* LASSO: For D M, E W alpha is 0.945 (t = 5.04). For E M, E W alpha is 0.840 (t = 5.68). LASSO minus U N C shows a D M E W alpha of 0.383 (t = 2.17).

* E N E T: For D M, E W alpha is 0.715 (t = 4.48). For E M, E W alpha is 0.459 (t = 2.72). E N E T minus U N C shows an E M E W alpha of minus 0.575 (t = minus 2.89).

* R F: For D M, E W alpha is 0.929 (t = 4.42). For E M, V W alpha is 0.574 (t = 1.76) and E W alpha is 0.958 (t = 4.29).

* G B R: For D M, E W alpha is 0.897 (t = 4.72). For E M, V W alpha is 0.553 (t = 1.76) and E W alpha is 0.868 (t = 4.76).

* N N 3: For D M, E W alpha is 0.838 (t = 4.48). For E M, E W alpha is 0.750 (t = 5.38). N N 3 minus U N C shows an E M E W alpha of minus 0.284 (t = minus 1.71).

* C I D: For D M, E W alpha is 0.775 (t = 3.39). For E M, V W alpha is 0.460 (t = 1.75) and E W alpha is 1.284 (t = 8.13). C I D minus U N C shows significant E W alphas for both D M (0.213, t = 5.55) and E M (0.250, t = 4.30).

Turning to the machine learning approaches in Table 6, we observe highly significant equal-weighted alphas around 0.9% in DM across most models, while value-weighted alphas remain insignificant around 0%. In EM, equal-weighted alphas remain significant but are slightly lower at around 0.8% for most models, while value-weighted alphas in EM are around 0.4%, but statistically insignificant. Hence, the machine learning models produce comparable alphas to the $ CID $ strategy for DM samples, but much lower alphas for EM samples.

When comparing the machine learning $ ML- UNC $ strategies with the $ CID- UNC $ strategy, we find only significant and comparable alphas in equal-weighted DM samples for the linear regression ( $ OLS $ ), $ LASSO $ , and random forest ( $ RF $ ) models. Notably, the equal-weighted $ ML- UNC $ alphas in EM are uniformly negative. Value-weighted $ ML- UNC $ alphas are mostly positive, but statistically insignificant in all models and in both DM and EM, ranging between −0.16% and 0.29%. Therefore, the machine learning models produce $ ML- UNC $ alphas comparable to those of the $ CID- UNC $ strategies for value-weighted DM and EM samples and equal-weighted DM samples, but underperform significantly in equal-weighted EM samples. We report statistical tests for the $ ML- CID $ return differences in Table OA.23 of the Supplementary Material, which show that none of the machine learning models significantly outperform the straightforward baseline strategy across regions and weighting schemes. Furthermore, a direct comparison of the machine learning models reveals no definitive superior performer.

Notably, the machine learning alpha spreads increase materially for more restrictive quintile splits in the pooled DM sample, but still remain insignificant or negative for the pooled EM sample (see Table OA.11 of the Supplementary Material). Overall, under comparable conditions, more complex methodsFootnote ¹⁰ do not outperform the baseline composite strategy ( $ CID- UNC $ ), which we therefore concentrate on in the remainder of the article.

C. Long Versus Short-Term Information

Our results so far concentrate on one-month ahead performance. However, an information advantage could also be of a longer-term nature. To test this possibility, we examine the profitability of $ CID $ insider trades based on longer holding periods ranging from 2 to 12 months. In other words, if a stock has at least two excess buy or sell signals in a month, it is assigned to the long or short portfolio accordingly and remains there for 2, 3, 6, or 12 months. We also implement an analogous portfolio approach for $ UNC $ trades, which allows us to study the $ CID- UNC $ performance difference.

In Table 7, we report value- and equal-weighted FF6 alphas separately for pooled and country-neutral developed (DM) and emerging market (EM) $ CID $ (Panel A) and $ CID- UNC $ strategies (Panel B) for different holding periods. We report the $ UNC $ alphas for different holding periods in Table OA.12 of the Supplementary Material.

TABLE 7Longer Holding Periods

TABLE 7 Long description

The table is divided into two main sections: Panel A and Panel B. Both panels share a common header structure with columns for holding periods of 1 Month, 2 Months, 3 Months, 6 Months, and 12 Months. Each period is subdivided into V W (Value-Weighted) and E W (Equal-Weighted) metrics. The final column N indicates 227 available months.

Panel A: C I D F F 6 alpha (Long-Short)

* Pooled Strategy:

- D M: E W alpha starts at 1.011 (t-stat 4.92) at 1 month and decreases to 0.337 (t-stat 3.35) at 12 months. V W alpha starts at 0.226 and turns negative to -0.053 at 12 months.

- E M: E W alpha starts at 1.387 (t-stat 8.80) and decreases to 0.477 (t-stat 5.62) at 12 months. V W alpha starts at 0.789 and decreases to 0.143.

* Country-Neutral Strategy:

- D M: E W alpha starts at 1.165 and decreases to 0.394. V W alpha starts at 0.418 and decreases to 0.033.

- E M: E W alpha starts at 1.313 and decreases to 0.480. V W alpha starts at 0.794 and decreases to 0.205.

Panel B: C I D minus U N C F F 6 alpha (Long-Short)

* Pooled Strategy:

- D M: E W alpha starts at 0.286 and decreases to 0.086. V W alpha starts at 0.275 and decreases to 0.017.

- E M: E W alpha starts at 0.383 and decreases to 0.072. V W alpha starts at 0.323 and decreases to 0.078.

* Country-Neutral Strategy:

- D M: E W alpha starts at 0.282 and decreases to 0.105. V W alpha starts at 0.136 and decreases to 0.044.

- E M: E W alpha starts at 0.269 and decreases to 0.010. V W alpha starts at 0.332 and decreases to 0.062.

Statistical significance is noted with asterisks, and t-statistics are provided in parentheses below each alpha value.

The main finding is that for both $ CID $ and $ CID- UNC $ strategies, irrespective of the region or country aggregation scheme, equal-weighted alphas decrease with longer time horizons by half or more, but largely remain significant. For example, the equal-weighted alphas of $ CID $ trades in the pooled DM and EM portfolios fall from highly significant 1.01% and 1.39%, respectively, in the 1-month holding period, to still highly significant 0.34% and 0.48%, respectively, in the 12-month holding period. Value-weighted $ CID $ alphas are mostly significant with the exception of pooled DM samples and also decrease with longer time horizons. Value-weighted $ CID- UNC $ alphas are often significant, but also decrease with longer time horizons, the exception being the country-neutral DM alphas.

Collectively, the economic difference between $ CID- UNC $ alphas tends to become smaller over time, which suggests that $ CID $ insider trades capture partly a short-term information advantage. Graph B of Figure 2 illustrates this negative relation between time horizon and $ CID- UNC $ alphas for the pooled DM and EM portfolios. The results of the country-neutral portfolios are qualitatively comparable, with the exception of the country-neutral DM alphas. Nevertheless, the equal-weighted $ CID- UNC $ alphas remain statistically significant at long time horizons, indicating that the $ CID $ measure also outperforms the equal-weighted $ UNC $ strategies at longer time horizons. In summary, we can confirm the hypothesis that $ CID $ trades contain an important short-term information component.

D. Real-Time Perspective and Earnings Months

In this section, we analyze two further aspects regarding the timing of insider transactions. In Panel A of Table 8, we first report the performance of a real-time implementable version of our baseline strategy to outside investors ( $ {CID}_O $ ), which includes only trades that have been reported to the respective authorities before the end of the previous month, that is, before the portfolio construction. Furthermore, we report in Panel A the performance of a pure insider version of our baseline strategy ( $ {CID}_I $ ), which includes only trades that have been reported to the respective authorities after the end of the previous month. These trades are therefore not available to outside investors at the time of portfolio formation.Footnote ¹¹

TABLE 8Real-Time Perspective and Earnings Months

TABLE 8 Long description

The table is divided into two main sections.

Panel A, Outsider versus Insider Perspective (Long-Short), compares C I D sub O and C I D sub I strategies.

- For Pooled D M, C I D sub O shows a V W alpha of 0.285 and E W alpha of 0.842. The difference C I D sub O minus U N C sub O is 0.319 (V W) and 0.262 (E W) over 227 months.

- For Pooled E M, C I D sub O shows a V W alpha of 0.654 and E W alpha of 1.274.

- Country-Neutral D M shows C I D sub O alphas of 0.473 (V W) and 1.048 (E W).

- Country-Neutral E M shows C I D sub O alphas of 0.791 (V W) and 1.253 (E W).

- Insider perspective (C I D sub I) generally shows higher E W alphas, reaching 2.222 for Country-Neutral D M.

Panel B, Q E A (Long-Short), compares months with quarterly earnings announcements (C I D sub Q) and without (C I D sub n Q).

- For Pooled D M, C I D sub Q shows a V W alpha of 0.231 and E W alpha of 0.803 over 167 months.

- For Pooled E M, C I D sub Q shows a negative V W alpha of -0.185 but a positive E W alpha of 1.350.

- In months without announcements (C I D sub n Q), Pooled E M shows a V W alpha of 0.707 and E W alpha of 1.436 over 221 months.

- Country-Neutral D M for C I D sub n Q shows a V W alpha of 0.559 and E W alpha of 1.181.

T-statistics are provided in parentheses below each alpha value, with asterisks indicating statistical significance at the 10%, 5%, and 1% levels.

The $ {CID}_O $ alpha (averaged across all EM and DM pooled and country-neutral specifications) decreases by approximately 6 bps, while the $ {CID}_O-{UNC}_O $ alpha spreads closely resemble the baseline results in Table 4 across both regions and weighting schemes. The $ {CID}_I $ alpha (averaged across all EM and DM pooled and country-neutral specifications) increases often significantly by approximately 43 bps,Footnote ¹² while the $ {CID}_I-{UNC}_I $ alpha spreads again closely resemble the baseline results in Table 4, with the exception of the value-weighted DM alpha spread becoming insignificant. In Table OA.22 of the Supplementary Material, we formally test whether the pure outsider strategy and pure insider strategy differ from one another and from the baseline strategy. For the $ CID $ alphas, we find mainly significant outperformance in equal-weighted DM samples, with the pure insider strategy outperforming the baseline ( $ {CID}_I- CID $ ) and the outsider strategy ( $ {CID}_I-{CID}_O $ ),Footnote ¹³ and the baseline strategy outperforming the outsider strategy ( $ CID-{CID}_O $ ). For the $ CID- UNC $ alpha spreads, we find no significant differences, except for significantly lower returns of the value-weighted pure insider strategy relative to the baseline strategy in developed markets. Overall, the slightly decreased outsider results again point to a short-term information advantage of insiders. Although the pure insider results might suggest a similar interpretation, we cannot directly make this link as the baseline and the pure insider strategy are by design far less comparable than the baseline to the outsider strategy.

In Panel B of Table 8, we perform a similar variation of the baseline strategy to investigate the influence of QEAs. Specifically, we rerun the analysis separately for trades in earnings announcement months on the one hand ( $ {CID}_Q $ ) and all other months on the other hand ( $ {CID}_{nQ} $ ). The alphas of trades initiated in months without an earnings announcement are at least as large as, if not larger than, the alphas in earnings months (all respective tests for significant differences in Table OA.22 of the Supplementary Material are insignificant), suggesting that the latter are not a major driver of the success of the baseline analysis.

E. Developed Versus Emerging Markets

The information environment in emerging markets (EM) is likely to be different from that in developed markets (DM) (see also Section V.F). For example, Brochet (Reference Brochet2019) shows that institutions in EM demand less transparency regarding insider trades, which, in turn, increases the predictability of returns. Bhattacharya et al. (Reference Bhattacharya, Daouk, Jorgenson and Kehr2000) find that stock prices in Mexico hardly reacted to usually value-relevant corporate news in the period between 1994 and 1997. Griffin et al. (Reference Griffin, Hirschey and Kelly2011) show that stock prices in DM move more strongly on public news, especially on earnings announcements. These results suggest that insiders in EM, in particular, may have private information advantages. In this respect, the question arises whether the outperformance of $ CID $ trades compared to $ UNC $ trades is also driven by this mechanism. However, Table 9, which reports tests for differences in abnormal returns between pooled and country-neutral EM and DM portfolios for the $ UNC $ , $ CID $ , and $ CID- UNC $ strategies, shows that this is unlikely to be the case.

TABLE 9Developed Versus Emerging Markets

TABLE 9 Long description

The table is divided into three vertical sections: Panel A (Long minus Short), Panel B (Long), and Panel C (Short). Each panel compares Emerging Markets (E M) minus Developed Markets (D M) across three strategy categories: Unconditional (U N C), Information-driven (C I D), and the difference (C I D minus U N C).

* Panel A (Long minus Short): In the Pooled implementation, the U N C V W alpha is 0.514 (z-stat 2.33) and C I D V W alpha is 0.563 (z-stat 1.77). In Country-Neutral, the C I D V W alpha is 0.376 (z-stat 2.23).

* Panel B (Long): In the Pooled implementation, U N C V W alpha is 0.289 and C I D V W is 0.215. In Country-Neutral, the U N C E W alpha is 0.484 (z-stat 2.42) and C I D E W is 0.364 (z-stat 1.69).

* Panel C (Short): In the Pooled implementation, U N C V W alpha is minus 0.225 and C I D V W is minus 0.348. The C I D minus U N C E W alpha is minus 0.199 (z-stat minus 2.41).

Columns for each strategy include Value-Weighted (V W), Equal-Weighted (E W), average number of firms in E M (F E M), and average number of firms in D M (F D M). The number of months (N) is 227 for all implementations. Statistical significance is marked with asterisks.

Panel A reveals that insiders in emerging markets achieve higher long–short alphas, sometimes significantly, than in developed markets for both unconditional insider trades ( $ UNC $ ) and the composite measure ( $ CID $ ). However, all $ CID- UNC $ alphas show small and insignificant differences between EM and DM, suggesting that differences in the information environment between EM and DM are not the main performance driver of the increased predictability of the $ CID $ strategies.

For example, in the pooled analysis with value-weighted returns, the alpha of $ UNC $ trades is 0.51% higher in EM than in DM, which is almost equally driven by the long and short legs (see Panels B and C). The respective estimate for $ CID $ trades is 0.56%, which is again driven by both the long and short legs (see Panels B and C), leading to an insignificant $ CID- UNC $ alpha of 0.05%. In the country-neutral equal-weighted strategy, the alpha difference is 0.16% for $ UNC $ trades and 0.15% for $ CID $ trades, again leading to an insignificant $ CID- UNC $ alpha of −0.01%.

Overall, higher informational advantages in EM manifest themselves already in the performance of unconditional insider trading, not additionally through the composite measure. In summary, there is limited to no evidence for the conjecture that stronger private information advantages, or other possible differences in the information environment between EM and DM, lead to stronger return predictability for information-driven $ CID $ trades.

F. Cross-Country Analysis

2. Analysis Based on Insider Trading Restrictions

The cross-country analysis aims to identify cross-country differences that are able to explain the varying abnormal performance of the unconditional ( $ UNC $ ) and composite ( $ CID $ ) information-driven insider trading strategies across countries. Therefore, the dependent variables are the country-level long–short raw returns obtained from Tables 2 and 3. Consequently, the independent variables are country characteristics comprising the time-invariant AS, CLASSA, and ITR, the time-averaged BLACKOUT, and the composite IRI (a standardized index of the preceding four proxies). Following Kolasinski and Yang (Reference Kolasinski and Yang2018) and Cochrane (Reference Cochrane2005), we estimate a one-step panel regression framework that jointly identifies the country-level risk-adjusted returns and their relation to country characteristics, using country-specific FF6 factor loadings (implemented via country-factor interactions):

(2) $$ {R}_{c,t}^{LS}={b}_0+{\boldsymbol{\gamma}}^{\prime }{\mathbf{X}}_c+{\boldsymbol{\beta}}_c^{\prime }{\mathbf{F}}_{c,t}+{\varepsilon}_{c,t}, $$

where $ {R}_{c,t}^{LS} $ is the value-weighted (equal-weighted) monthly long–short return in country $ c $ and month $ t $ , based on the country-level portfolios reported in Tables 2 and 3, respectively. The vector $ {\mathbf{X}}_c $ contains the time-invariant country-level insider regulation proxies and the vector $ {\mathbf{F}}_{c,t} $ contains the country-specific monthly FF6 factor returns. The factor-loading vector $ {\boldsymbol{\beta}}_c $ is allowed to vary by country, so that each country has its own exposure to the FF6 factors. The coefficient vector $ \boldsymbol{\gamma} $ captures the relation between the insider regulation proxies and cross-country differences in risk-adjusted long–short returns. Standard errors are double-clustered by country and month.

Table 10 reports the empirical results with respect to unconditional ( $ UNC $ , in Panel A) or composite ( $ CID $ , in Panel B) information-driven long–short insider trading strategies. In Table OA.15 of the Supplementary Material, we provide the same analysis separately for the long and short portfolios.

TABLE 10Cross-Country Analysis

TABLE 10 Long description

The table is divided into Panel A for the U N C (unconditional) strategy and Panel B for the C I D (composite information-driven) strategy. Both panels report value-weighted (V W) and equal-weighted (E W) long-short returns across 34 countries.

Panel A (U N C):

- AS (anti-self-dealing index): V W coefficients range from -0.530 to -0.507; E W coefficients range from -0.644 (significant at 1 percent) to -0.397 (significant at 10 percent).

- CLASSA (class action dummy): V W coefficients are -0.256 and -0.196; E W coefficients are -0.114 and -0.052.

- BLACKOUT (blackout periods): V W coefficients are -0.210 and 0.161; E W coefficients are -0.979 and -0.887 (both significant at 5 percent).

- ITR (insider trading restriction index): V W coefficients are -0.047 and -0.032; E W coefficients are -0.053 and 0.045.

- IRI (combined index): V W coefficient is -0.199 (significant at 10 percent); E W coefficient is -0.260 (significant at 5 percent).

- N (observations): Ranges from 6,029 to 6,210.

- R-squared: Ranges from 0.097 to 0.101.

Panel B (C I D):

- AS: V W coefficients are -0.743 and -0.715; E W coefficients are -0.541 and -0.257.

- CLASSA: V W coefficients are -0.225 and -0.123; E W coefficients are -0.066 and -0.004.

- BLACKOUT: V W coefficients are -0.502 and 0.064; E W coefficients are -0.962 and -1.148 (both significant at 10 percent).

- ITR: V W coefficients are -0.103 and -0.084; E W coefficients are -0.014 and 0.102.

- IRI: V W coefficient is -0.281 (significant at 10 percent); E W coefficient is -0.209.

- N (observations): Ranges from 6,029 to 6,210.

- R-squared: Ranges from 0.089 to 0.102.

T-statistics are provided in parentheses below each coefficient. Asterisks indicate statistical significance levels: * for 10 percent, ** for 5 percent, and *** for 1 percent.

Assuming that the monitoring hypothesis is correct, one would expect AS, CLASSA, BLACKOUT, ITR, and IRI to take negative values, as an increase in all of them indicates higher insider trading restrictions.

Panels A and B of Table 10 report the results of the regressions for the value- and equal-weighted long–short implementations of the unconditional ( $ UNC $ ) and composite ( $ CID $ ) information-driven insider trading strategies, respectively. The univariate regressions of Panel A show consistent results in line with the monitoring hypothesis, that is, consistently negative although only sometimes significant coefficients.Footnote ¹⁶ We only find a significant (at the 5% level) relationship between the equal-weighted $ UNC $ strategy and the anti-self-dealing index (AS), the occurrence of blackout periods (BLACKOUT), and the composite insider regulation index (IRI), respectively. This suggests that the abnormal performance of the equal-weighted $ UNC $ strategy is greater in countries with lower insider trading restrictions. Judging the explanatory power of insider trading restrictions based on the multiple regression framework which jointly includes all four original proxies reveals no significant relation with respect to the value-weighted $ UNC $ strategy and a negative relation (at the 5% level) between the equal-weighted $ UNC $ strategy and BLACKOUT only. Nevertheless, at least most of the signs remain negative across proxies and specifications.

Assuming that the relative outperformance of the $ CID $ strategy originates from insiders using nonpublic information, we conjecture that the explanatory power of insider trading restriction proxies is stronger for these trading strategies. Contradicting this conjecture, Panel B shows fewer (i.e., no) significant coefficients (at the 5% level) for the insider regulation proxies. This is also confirmed by the insignificant coefficients in all specifications of the cross-country analysis for the long–short $ CID- UNC $ strategy in Table OA.16 of the Supplementary Material.Footnote ¹⁷

Finally, as an alternative test to the regression-based cross-country analysis, we use the composite insider regulation index (IRI) to create high (HIRI) and low (LIRI) regulation regions via a tercile split to replicate the main DM versus EM analysis of Tables 4 and 9 for the high and low regulation samples. The results are reported in Table OA.21 of the Supplementary Material.Footnote ¹⁸ Overall, the high versus low insider regulation sample tests only show limited evidence for a greater outperformance in low regulation regions, as only the value-weighted country-neutral $ CID- UNC $ spreads are statistically significant.

Overall, the comparatively weaker findings for the $ CID $ strategies in Panel B of Table 10 and the insignificant results for the performance differences between high and low regulation samples may imply that regulatory differences of insider trading restrictions across countries are not a key driver of the strategy performance. Rather than exploiting nonpublic information, informative insiders in the $ CID $ strategies may achieve their outperformance primarily through a superior interpretation of public information compared to outside investors.