Director discretion and insider trading profitability

Pacific-Basin Finance Journal 39 (2016) 28–43

Contents lists available at ScienceDirect

Pacific-Basin Finance Journal journal homepage: www.elsevier.com/locate/pacfin

Director discretion and insider trading profitability Sean Foley a, Amy Kwan a, Thomas H. McInish b,⁎, Richard Philip a a b

University of Sydney, Australia University of Memphis, United States

a r t i c l e

i n f o

Article history: Received 8 November 2015 Received in revised form 17 February 2016 Accepted 22 May 2016 Available online 26 May 2016 Keywords: Insider trading Director discretion

a b s t r a c t Using a machine-learning algorithm, we classify over 60,000 director transactions into discretionary and non-discretionary purchases and sales based on the trading motive provided by the insider. We find that discretionary trades by company insiders are more informed than non-discretionary trades. Further, discretionary purchases generate higher abnormal returns (1) for larger purchases, or when the purchase is for (2) the stock of a smaller firm, or (3) a firm with greater information asymmetry. © 2016 Elsevier B.V. All rights reserved.

1. Introduction Corporate insiders trade in their company's stock for many reasons, including for diversification or liquidity needs, to take advantage of discount plans, or to cash in vested options (see Cohen et al., 2012). Moreover, insiders have private information and can profit by trading prior to takeover announcements (Meulbroek, 1992; Acharya and Johnson, 2010; Heitzman and Klasa, 2015), earnings news (Piotroski and Roulstone, 2005; Huddart et al., 2007), dividend announcements (John and Lang, 1991), or significant accounting disclosures (Ke et al., 2003). These differences in motivation are typically difficult to differentiate. Additionally, insiders may simply be better market timers (Lakonishok and Lee, 2001). In this study, we analyse the profitability of directors' trades using a new dataset that identifies the directors' trading motives. Uniquely, we are able to classify directors' transactions into discretionary and non-discretionary trades. Similar to U.S. markets where corporate insiders are required to file monthly reports to the SEC on open market trades, the Australian stock exchange (ASX) requires directors to disclose any changes to director interests within 5 business days of the transaction. However, in contrast to the U.S., the ASX also requires that the director specify the reason for the trade. We use the additional granularity in our data to provide deeper insights into the profitability of insider trading. Deciphering the reasons provided in director filings is not straightforward. With over 85,000 director filings submitted to the ASX between January 2005 and December 2014, hand classifying the reasons provided by the director is not feasible. Drawing from the computer science literature, we use recent developments in machine-learning algorithms, which can be applied to many areas of financial research, to classify the trading motive provided in each filing into 12 broad categories. We find that directors change their holdings for a large variety of reasons. For example, directors may participate in a rights offering or dividend reinvestment plan. Similarly, insiders may increase their shareholdings as part of a performance bonus or remuneration package. Importantly, we find that in approximately 65% of the filings, the nature of the change in holdings is given as either ‘on-market’ or ‘off-market.’ Our initial analysis indicates that these trades without a clear motive are more predictive of future returns than trades for which a specific reason is provided.

⁎ Corresponding author at: 108 Fogelman Executive Center Memphis, TN 38152-3120, United States. E-mail address: [email protected] (T.H. McInish).

http://dx.doi.org/10.1016/j.pacfin.2016.05.005 0927-538X/© 2016 Elsevier B.V. All rights reserved.

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

29

Based on our 12 categories, we further categorise transactions into discretionary and non-discretionary buys and sells and find that discretionary transactions are more profitable. Our results show that insider discretionary purchases have a cumulative abnormal return of 4.6% in the 200 trading days after the trade while non-discretionary purchases return −4.7%. Similarly, abnormal returns are on average 15% lower in the 200 trading days after a discretionary sell while abnormal returns following nondiscretionary sell transactions are not statistically different from zero. For discretionary purchases, abnormal returns are also larger (1) the bigger the size of the purchase, (2) when the purchase is for the stock of a smaller firm, and (3) when the purchase is for the stock of a firm in an industry with greater information asymmetry. Our study contributes to the large body of academic literature on insider trading. The early literature concludes that insiders can earn significant abnormal profits by trading in their own firms (Rogoff, 1964; Lorie and Niederhoffer, 1968; Jaffe, 1974; Finnerty, 1976; Seyhun, 1986, 1988; Rozeff and Zaman, 1988; Lin and Howe, 1990; Lakonishok and Lee, 2001, and Marin and Olivier, 2008). These findings have been largely supported by studies using data for the U.K. (Pope et al., 1990 and Friederich et al., 2002), Sweden (Kallunki et al., 2009), Australia (Aspris et al., 2014, Chang and Watson, 2015), and New Zealand (Etebari et al., 2004). On the other hand, Eckbo and Smith's (1998) study of the Oslo Stock Exchange and Chiang et al.'s (2004) study of the Taiwan Stock Exchange find no abnormal performance for insider trades. While many previous studies find that insider purchases are highly informative, only weak evidence of the profitability of insider sales is reported. For example, Jeng et al. (2003) find that insider purchases earn abnormal returns of more than 6% per year while insider sales do not earn significant abnormal returns. Similarly, Lakonishok and Lee (2001) find that insider purchases predict future market movements whilst insider sales have no predictive ability. To investigate the reasons driving these information differences more closely, recent studies develop methods to identify reasons behind the insiders' trading decision. Cohen et al. (2012) infer the nature of an insider trade based on historical patterns in director trading behaviour. Specifically, they conjecture that ‘routine’ insider trades (i.e., those that occur in the same month every year), likely represent trading for diversification or liquidity reasons while ‘opportunistic’ trades (i.e. those that are less predictable), are likely to be more informed. Using these trade classifications, these authors find that ‘routine’ trades contain less information than ‘opportunistic’ trades. Similarly, Tirapat and Visaltanachoti (2013) distinguish between opportunistic and liquidity trades by insiders on the Stock Exchange of Thailand based on the probability of informed trading and report higher average returns for opportunistic trades. Other studies use incidences of delayed disclosure to infer the reason behind insiders' trading decisions. Cheng et al. (2007) exploit an exception to the SEC insider-trading-disclosure rules to distinguish between liquidity- and private-informationdriven insider sales. Prior to the Sarbanes-Oxley Act of 2002, the SEC required the majority of open-market sales to be disclosed to the SEC by the tenth day of the next month. However, some sales are exempt from this requirement and can be reported within 45 days of the fiscal year end. These authors find that in large firms, late-disclosed insider sales, which are more likely to be information driven, are more predictive of negative future returns than liquidity driven, quickly disclosed sales. Betzer et al. (2015) extend this work and find that the delayed reporting of stealth trades (i.e., a trade that is followed by another trade before the first trade is reported) outperforms the returns of similar non-stealth trades. Etebari et al. (2004) distinguish between immediate disclosures and delayed disclosures by corporate insiders from companies listed on the New Zealand Stock Exchange and report similar findings. Specifically, transactions involving delayed disclosure earn large abnormal returns while transactions that are immediately disclosed to the market earn insignificant returns. Similarly, Chang and Watson (2015) show that delayed disclosures of insider transactions are more predictive of future firm performance than timely disclosures in Australian firms. To reiterate, in contrast to previous work, because of the unique information in the ASX filings, we do not need to infer the reason behind insiders' trades. Using text recognition techniques, we generate a unique dataset containing detailed information on the motivations behind each insiders' trade. 2. Research Design 2.1. Data and Text Processing Our data on the changes to directors' holdings are obtained from the Securities Industry Research Centre of Asia-Pacific Australian Company Announcement database. ASX listing rule 3.19A.2 requires an entity to notify the ASX when there is a ‘change to a notifiable interest of a director of the entity’ within 5 business days of the change. The rule requires the director to lodge with the ASX an Appendix 3Y ‘Change in Director's Interest Notice’ that includes the name of the entity and the director, the date of the change, the number of securities held prior to and after the change, and the number of shares acquired or disposed.1 Additionally, the director is required to outline the nature of the change in holdings. While the required explanation for the change is unrestricted, the ASX also provides examples of common reasons for trading, including: on-market trade, off-market trade, exercise of options, issue of securities under a dividend reinvestment plan, or participation in a buy-back. We download Appendix 3Y forms for all ASX listed companies from January 2005 to December 2014.2 However, we are unable to analyse the raw data for two reasons. First, the forms are filed to the ASX as a pdf document containing embedded text (see Appendix 1) that needs to be converted to a machine readable format. Second, the motivations provided in the filings is at the

1 2

Appendix 1 provides an example of the Appendix 3Y form. Prior to January 2005, Appendix 3Y forms were faxed and as such were not machine readable.

30

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

discretion of the directors, and, thus, cannot be easily classified.3 Given that there are approximately 85,000 director filings over our sample period, we develop a method to systematically classify each reason provided in the filings. We convert the pdf files into html format to extract from each field the information provided by the director. For fields in which a number is expected, only the numerical component is extracted. For example, if the response to the field ‘Number acquired’ is ‘527,449 Ordinary shares’, we recognise ‘527,449’ as the number of shares acquired. Appendix 2 provides an example of the text conversion process using the filing from Appendix 1. We then classify each of the unique reasons provided by the directors into 12 broad categories. Because it is infeasible to hand classify over 85,000 filings, we apply machine-learning algorithms from the computer science literature for the text classification. Generally, text classification via machine-learning involves two main steps. First, the text string is converted to a suitable format for the learning algorithm. Typically, text strings are transformed into a document-term matrix that describes the frequency of key words in documents. In our study, each row represents a unique document (one for each director filing) and the columns represent the frequency of key words appearing in the document. Next, a text classification algorithm learns from a ‘training dataset’ so that the remaining observations can be automatically classified. We use a training dataset that comprises 2000 manually-classified observations. While many text classification algorithms exist, we use the support vector machine (SVM) algorithm as it yields fewer errors when conducting experiments on binary text classification (see Joachims, 1998; Burges, 1998; and Yang and Liu, 1999).4 Specifically, the SVM model represents the document-term matrix as points in space, and maps each of the 2000 observations in the training dataset into 12 separate categories that are divided by a clear gap that is as wide as possible. Next, the algorithm maps the document-term matrix for the remaining observations into the same space and individually classifies the observations into the 12 categories based on their location.5 In Appendix 3, we show various phrases that the SVM assigns to the 12 categories. To ensure the accuracy of the technique in our application, we use the algorithm to re-assign 100 observations from our original training dataset. Comparing the SVM classifications to our manual classifications, we find that the SVM accurately classifies all but one of the 100 observations. We formally test the stability of the SVM model in Appendix 4. We then merge the classified director trading data with firm-level data from Bloomberg, including daily adjusted closing prices, market capitalization and book-to-market ratios. Table 1 presents summary statistics for our final sample of Appendix 3Y filings, which contains 63,607 director filings, representing 74% of the filings downloaded. There are several major reasons for the reduction in sample size. First, Appendix 3Y forms that are faxed to the ASX, scanned from another document, or are hand-written cannot be converted into html format and are excluded from the sample. Approximately 10% of the original sample is excluded for these reasons. Second, filings that do not conform to the standard format, are incorrectly filled out by the director, or are corrections to an earlier submission are excluded. Finally, companies with missing price data are excluded from the sample.6 Overall, we find that the majority of the data issues arise in the earliest year of our sample. In the later years, we successfully process approximately 85% of the raw Appendix 3Y forms submitted to the ASX. Table 2 lists our 12 categories and their associated frequencies. We find that a large proportion of the filings relate to capital raising initiatives. In combination, dividend reinvestment plans, rights issues, and underwriting activities are the stated motive for trading in approximately 13% of filings. Director holdings in the company's securities also change due to share and option based remuneration packages. We find that remuneration or option exercise/expiry is the reason provided in over 15% of insider filings. Importantly, in approximately 65% of filings, directors buy or sell shares without a stated reason. It is likely that these trades, whether on-market or off-market, contain the most director discretion. 2.2. Calculating Abnormal Returns We assess the informativeness of a trade using abnormal returns. We first estimate the market model for a 200 day period ending 10 trading days before the director trade: Rit ¼ α i þ βi Rmt þ εit ;

where Rit is the return for firm i at time t, Rmt is the return of the return of the ASX200 accumulation index at time t, αi and βi are parameters to be estimated, and εit is an error term. Returns are calculated from daily closing prices adjusted for dividends and stock splits for each day in the estimation period.

3 For example, the reasons provided for a dividend reinvestment plan vary widely, including ‘DRP’, ‘Dividend Reinvestment Plan Allotment’ and ‘Plan in lieu of cash dividends’. 4 Yang and Liu (1999) show that more computationally intensive supervised learning techniques such as naïve Bayes and k-nearest neighbours make little improvement over the SVM. Joachims (1998) also notes that SVM achieves substantial improvements over other algorithms when there is a high dimensional input space. The SVM algorithm is available through the RTextTools package in R. 5 Because the algorithm classifies observations based on the observation's location, the SVM model is highly stable and identical strings are always allocated to the same category. Similarly, if provided with the same training dataset, SVM always produces the same classifications. 6 Scanned documents are more common in earlier years.

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

31

Table 1 Distribution of announcements by year. This table reports the distribution of director filings (Columns 2 and 5) by year. We exclude announcements that are not machine-readable (perhaps because the announcement is hand written or a scanned copy/fax rather than an embedded text PDF document). We lose more observations in the earlier years. Embedded text PDFs are very rare prior to 2005, which is why we do not collect earlier data. The initial sample reports all announcements and the final sample reports the useable announcements. Columns 3 and 6 (4 and 7) show the percentage (number) of firms that report directors' changes in ownership during the year. Year

Initial sample

2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Total

Final sample

Announcements

%

Firms

Announcements

%

Firms

7517 8311 8409 11,040 9120 8570 9185 8255 7937 7836 86,180

8.72 9.64 9.76 12.82 10.58 9.94 10.66 9.58 9.21 9.09 100.00

1296 1420 1572 1666 1601 1610 1603 1574 1545 1552 15,439

1534 5750 5924 8085 7419 6811 7388 7026 7045 6625 63,607

2.41 9.04 9.31 12.70 11.66 10.71 11.62 11.05 11.08 10.42 100.00

530 910 1026 1177 1234 1256 1264 1328 1347 1355 11,427

We use the index return and the αi and βi estimates from the estimation period to estimate expected returns for the event window [0,200]. Abnormal returns are calculated as: ARit ¼ r it −Rit ; where rit is the actual return of stock i, on day t and Rit is the expected return for firm i, on day t. The abnormal returns are cumulated over the 200-day event window to produce a cumulative abnormal return (CAR). To ensure the robustness of our results, we also calculate CARs over a number of alternative event windows. 3. Results 3.1. Discretionary vs. Non-discretionary Director Trades Recent studies find that the motive behind an insider's trade can have a large effect on the trade's profitability (Etebari et al., 2004; Cheng et al., 2007; Cohen et al., 2012; Tirapat and Visaltanachoti, 2013 and Chang and Watson, 2015). Each of the 12 transaction categories outlined in Table 2 reflect different levels of trading discretion. For example, transactions that are due to option expiry, remuneration or a rights issue are likely to contain less discretion than an on-market trade, which is likely to reflect an active choice by the director. Trades due to director underwriting are also likely to be discretionary in nature if the director chooses to underwrite a rights issue or placement. For these reasons, we classify on-market, off-market and underwriting trades as discretionary transactions and designate the remaining categories as non-discretionary.7 Fig. 1 shows the CAR from the day of the insider transaction to 200 trading days after the transaction for discretionary and non-discretionary purchases and sales. We find that discretionary purchases and sales contain more information than nondiscretionary purchases and sales. Specifically, the results show an average CAR of approximately 4.6% (t-stat = 8.74) and − 15% (t-stat = − 15.4) in the 200 days following discretionary purchases and sales, respectively. In contrast, for nondiscretionary purchases, we report a CAR of −5% (t-stat = 7.43) while non-discretionary sales return −3%, although this result is insignificant at conventional levels. Thus, while non-discretionary transactions do not contain information, discretionary purchases and sales appear to be highly informed. To test the information content of the trades more formally, Table 3 compares summary statistics of discretionary and nondiscretionary trades. In Table 3, Panel A, we show that discretionary purchases outperform non-discretionary purchases in both shorter-term and longer-term event windows after the trade. We also find that discretionary purchases are smaller than nondiscretionary purchases and that they are more likely to occur in firms with lower profitability and lower book-to-market ratios. For sales in Table 3, Panel B, we find that discretionary sales outperform non-discretionary sales in both the 100 and 200 trading day intervals. We conduct OLS regressions to test the relationship between CAR[0,+200] and indicator variables for whether a purchase or sale is discretionary or non-discretionary. We control for differences in firm size (Ln(MktCapi)) and value (Ln(BooktoMarketi)) in

7

The results are robust to excluding underwriting trades from the analysis.

32

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

Table 2 Distribution of announcements across our 12 categories of motives. This table reports the distributions of useable director trading filings by category. Column 2 (Column 3) shows the total number (percentage) of directors' filings for each category. Column 4 presents the average AUD dollar value per filing for each category. The last three columns indicate the percentage of each row for which there was only an acquisition of stock, only a disposal, or both an acquisition and disposal, respectively. The categories are changes in holdings due to: share splits or consolidations, Consolidation; death of a director, Death; dividend re-investment plans, DRP; expiry of escrow periods, Escrow; exercise or expiry of listed or unlisted options/warrants, Exercise/expiry; purchases between related entities or transfers occurring due to the settlement of put/call options, Off-market; no reason is given, or the reason given is simply “on-market trade”, On-market; issued for services rendered to the company, Remuneration; the uptake of a rights issue, Rights; transactions associated with a firm takeover, Takeover; transfers between parties, often between direct and indirect ownership, Transfer; the director acquiring shares by choosing to become the underwriter of a rights issue, Underwriting.

Consolidation Death DRP Escrow Exercise/expiry Off-market On-market Remuneration Rights Takeover Transfer Underwriting Total

Announcements

%

Consideration

Buy %

Sell %

Both %

242 74 3314 240 3722 1864 39,900 6760 5004 507 1707 273 63,607

0.38 0.12 5.21 0.38 5.85 2.93 62.72 10.63 7.87 0.8 2.68 0.43 100.00

118,280 53,250 7964 111,256 125,000 72,459 24,999 17,087 16,669 376,388 80,238 98,979

11.57 28.38 95.32 50.00 35.81 45.33 58.11 92.38 90.35 30.57 25.48 94.87

11.57 37.84 0.81 5.00 28.05 32.03 16.07 1.76 2.40 57.99 20.80 0.37

76.86 33.78 3.86 45.00 36.14 22.64 25.82 5.86 7.25 11.44 53.72 4.76

our sample of firms. To control for market run-ups prior to the directors trade, we include the stock returns in the previous month (PastMonthReti) and year (PastYearReti), relative to the time of the transaction. Specifically, we estimate Eq. (1) as follows for each trade: CAR½0; 200i ¼ β1 Buyi þ β2 DiscretionaryBuyi þ β3 Selli þ β4 DiscretionarySelli þ β5 LnðMktCapi Þ þ β6 LnðBooktoMarket i Þ þ β7 PastMonthRet i þ β8 PastYearRet i þ ϵi

ð1Þ

where Buyi (Selli) is an indicator variable equal to 1 if transaction i is a purchase (sale) and 0 otherwise. DiscretionaryBuyi (DiscretionarySelli) is equal to 1 if transaction i is a purchase (sale) classified as ‘on-market’, ‘off-market’ or ‘underwriting’, and zero for the other transaction categories. In Table 4, we present the results separately for director purchases, sales and for both purchases and sales together. Consistent with our hypothesis, we find that discretionary purchases, which are more likely to contain the director's private information, result in higher abnormal returns than non-discretionary purchases. Specifically, we find that DiscretionaryBuyi is positive and significant, indicating that discretionary purchases outperform non-discretionary purchases by an additional 5% (Columns 1 and 3). On the other hand, we do not detect a significant difference in abnormal returns between discretionary and non-discretionary sales after including the control variables and month fixed effects.8 Our results are qualitatively similar when we replace CAR[0,200] with CAR[0,25] or CAR[0,100] as the dependent variable. However, our results are economically insignificant when we measure CARs over shorter horizon event windows of [0,1], [0,5] or [−10,10].9 Given that insider trading is illegal, directors are unlikely to trade on short-term information, such as information contained in upcoming earnings and/or takeover announcements. Further, ASX Listing Rule 12.9 requires listed companies to implement a trading policy which details the entity's black-out periods. For example, according to ASX Guidance Note 27, the ASX expects entities to impose trading black-out periods prior to the release of half- and full-year financial results and around the release of market-sensitive announcements. For these reasons, it is likely that directors are trading on longer-term information related to the future outlook of the company and as such, their trading is only profitable over longer horizons. 3.2. Large vs. Small Director Trades We investigate the effect of trade size on the relation between the motivation of an insider's trade and future abnormal returns. Directors who are more informed about the future stock price performance of the company are likely to commit more of their own personal wealth to the firm. Conversely, a director may attempt to hide their information through a series of smaller trades (Betzer et al., 2015). We define a large trade as a director trade in which the total consideration is above the median

8 In Appendix 5, we split the full sample into three distinct sub periods: pre-crisis (January 2003 to March 2007), crisis (April 2007 to December 2008) and post-crisis (January 2009 onwards). Consistent with our main results, we find that discretionary purchases result in higher abnormal returns than non-discretionary purchases across all sub periods while discretionary sales do not outperform non-discretionary sales. 9 These results are available upon request from the authors.

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

33

Fig. 1. The profitability of insider trades. This figure shows the cumulative abnormal returns (CARs) for discretionary and non-discretionary purchases and sales from the day of the transaction to 200 trading days after the transaction. CARs are estimated based on a market model for a 200 day estimation period ending 10 trading days before the director trade.

consideration (25,000 AUD) for the full sample of transactions. We augment our baseline regressions to: CAR½0; 200i ¼ β1 Buyi þ β2 Buyi  LargeTradei þ β3 DiscretionaryBuyi þ β4 DiscretionaryBuyi  LargeTradei þ β5 Selli þ β6 Selli  LargeTradei þ β7 DiscretionarySelli þ β8 DiscretionarySelli  LargeTradei þ ∑βk Controlsi þ ϵi

ð2Þ

where LargeTradei is an indicator variable equal to 1 if trade i is greater than $25,000 and zero otherwise. The control variables for this equation and all subsequent regressions are the same as those from Eq. (1). Table 3 Descriptive statistics for discretionary and non-discretionary trades. This table reports descriptive statistics for director purchases (Panel A) and sales (Panel B) separated by whether trades are discretionary or non-discretionary. Discretionary trades—Off-Market, On-Market and Underwriting—are those where the director directly determines the timing of the trade. The remaining categories are nondiscretionary. CAR[0,25], CAR[0,100], and CAR[0,200] provide the CARs (in percent) for directors' trades from the day of trade (0) to the 25th, 100th, and 200th trading day, respectively. Consideration indicates the average amount of each transaction, in thousands of AUD. MktCap is the market capitalization in millions of AUD of the firm 30 days prior to the director's trade. EBIT is the earnings before interest and tax for each firm and BooktoMarket is the book-to-market ratio for the firm. Both are measured at the end of the most recently completed financial year. Non-Discretionary Mean

Discretionary

Diff

STD

Mean

Panel A. Director purchases CAR[0,25] −0.430 CAR[0,100] −1.599 CAR[0,200] −4.742 CAR[−10,10] −0.277 Consideration 136.3 MktCap 1210 EBIT 90.05 BooktoMarket 2.800

−0.630 −1.261 −3.553 −0.445 17.87 91.76 −0.560 1.590

23.45 48.69 77.26 23.18 379.2 4190 268.0 4.860

0.785 2.797 4.638 0.592 82.30 1178 49.79 2.460

−0.057 0.826 1.579 −0.149 19.89 42.78 −0.62 1.330

21.66 49.49 82.72 21.34 262.4 20,647 254.8 4.650

1.215 4.396 9.380 0.870 −54.01 −31.54 −40.26 −0.340

5.14⁎⁎ 8.65† 11.3† 1.80⁎ −6.47† −0.05 −2.28⁎ −2.36⁎

Panel B. Director sales CAR[0,25] −1.049 CAR[0,100] −1.603 CAR[0,200] −2.677 CAR[−10,10] −0.775 Consideration 446.0 MktCap 1298 EBIT 83.49 BooktoMarket 3.820

−1.770 −1.113 −4.809 −1.195 133.0 146.3 8.610 2.300

26.20 55.25 87.88 14.68 721.2 5556 179.34 6.770

−1.806 −6.849 −14.43 −0.249 536.5 760.0 76.54 3.760

−1.517 −5.576 −11.21 −1.095 166.5 113.2 3.340 2.380

19.65 46.74 78.26 17.68 816.9 1819 198.2 5.270

−0.756 −5.246 −11.75 0.525 90.48 −538.1 −6.950 −0.060

−1.10 −3.54† −4.90† 0.42 1.53 −1.48 −0.29 −0.11

† Statistical significance at 1% level. ⁎⁎ Statistical significance at 5% level. ⁎ Statistical significance at 10% level.

Median

t-Stat

Median

STD

34

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

Table 4 Returns to director trades. This table reports estimates of CARs for directors' trades over the 200 days following the trade. Buy (Sell) is a dummy variable that equals 1 when shares are solely acquired (disposed) and 0 otherwise. DiscretionaryBuy (DiscretionarySell) is a dummy variable that equals 1 for discretionary director trades where shares are solely acquired (disposed) and 0 otherwise. MktCap is the market capitalization in millions of AUD of the firm 30 days prior to the director's trade. BooktoMarket is the book-tomarket ratio for the firm as of the last financial year prior to the trade. PastMonthRet is the CAR from trading day −25 to 0, and PastYearRet is the CAR from day −200 to −26. The parentheses report t-statistics based on standard errors double clustered by GICS industry and year (Thompson, 2011). Variable

Buys only (1)

Buy

0.68† (6.36) 0.05† (2.83)

DiscretionaryBuy Sell DiscretionarySell Ln(MktCap) Ln(BooktoMarket) PastMonthRet PastYearRet Observations Adjusted R2 Fixed effects

−0.03† (−4.91) −0.11† (−5.34) 0.07 (0.79) 0.36† (7.24) 33,572 17% Year

Sells only (2)

0.50† (2.74)⁎ −0.02 (−0.52) −0.03⁎⁎ (−2.15) −0.12† (−5.45) −0.53⁎⁎ (−2.11) −0.15 (−0.92) 7231 19% Year

Both buys and sells (3) 0.65† (5.54) 0.05† (2.88) 0.62† (4.98) −0.03 (−0.82) −0.03† (−4.37) −0.11† (−5.50) −0.03 (−0.31) 0.28† (5.83) 40,803 17% Year

† Statistical significance at 1% level. ⁎⁎ Statistical significance at 5% level. ⁎ Statistical significance at 10% level.

Table 5 shows that large purchases contain less information than smaller purchases, indicated by the interaction term between Buyi and LargeTradei for both the buys only subsample (Column 1) and the full sample of trades (Column 3). This is likely driven by participation in rights issues and the exercise of outstanding options. On the other hand, comparing between large and small discretionary purchases, we find that the interaction term between DiscretionaryBuyi and LargeTradei is positive and significant indicating that large discretionary purchases are more informative than smaller discretionary purchases. This finding on discretionary purchases differs from the findings in Betzer et al. (2015), possibly due to the continuous reporting requirements on the ASX. During the pre-SOX sample period analysed in Betzer et al. (2015), company directors can report trades up to 40 days after the original transaction. However, the ASX requires directors to report trades within 5 days of a trade. Thus, in comparison to the U.S. markets, there are fewer incentives to hide information through a series of smaller trades on the ASX. For sales, no difference is detected between the information content of discretionary and non-discretionary sales (Columns 2 and 3). 3.3. Direct vs. Indirect Director Trades Next, we investigate the information content of direct and indirect insider transactions. Direct stock holdings are held in the insider's own name while indirect holdings are held through the accounts of a relative or related entity. Investigating transactions before earnings and takeover announcements, Berkman et al. (2014) find that guardians channel their most profitable trades through the accounts of their children, rather than trading in their own accounts. Based on their findings, we expect indirect insider transactions to be more informative than direct insider trades. Similarly, Del Brio et al. (2002) find that indirect insider trades contain more information than direct trades in a sample of Spanish firms, suggesting that insiders could be camouflaging their trades by transacting through a third party. ASX listing rules require directors to report whether they are holding a direct or indirect interest. In Table 6, we report the results from regressions that incorporate an indicator variable for indirect trades (IndirectTrade), which is equal to 1 if the trade is disclosed as an indirect holding, and zero otherwise: CAR½0; 200i ¼ β1 Buyi þ β2 Buyi  IndirectTradei þ β3 DiscretionaryBuyi þ β4 DiscretionaryBuyi  IndirectTradei þ β5 Selli þ β6 Selli  IndirectTradei þ β7 DiscretionarySelli þ β8 DiscretionarySelli  IndirectTradei þ ∑βk Controlsi þ ϵi ð3Þ For purchases in Columns 1 and 3, our results do not support the results reported in Berkman et al. (2014) and Del Brio et al. (2002). We find a negative and significant coefficient on Buyi × IndirectTradei indicating that director purchases through the account of a third party underperform trades in the director's own name. Similarly, dividing director purchases into discretionary and non-discretionary purchases, we do not find that indirect discretionary purchases are more informed than discretionary

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

35

Table 5 Returns to large vs small director trades. This table reports estimates of CARs for directors' trades over the 200 days following the trade based on whether the trade is a large trade (larger than 25,000 AUD). LargeTrade is a dummy variable that equals 1 for large trades and 0 otherwise. Buy (Sell) is a dummy variable that equals 1 when shares are solely acquired (disposed) and 0 otherwise. DiscretionaryBuy (DiscretionarySell) is a dummy variable that equals 1 for discretionary director trades where shares are solely acquired (disposed) and 0 otherwise. MktCap is the market capitalization in millions of AUD of the firm 30 days prior to the director's trade. BooktoMarket is the book-to-market ratio for the firm as of the last financial year prior to the trade. PastMonthRet is the CAR from trading day −25 to 0, and PastYearRet is the CAR from day −200 to −26. The parentheses report t-statistics based on standard errors double clustered by GICS industry and year (Thompson, 2011). Variable

Buys only (1)

Buy

0.90† (10.80) −0.08⁎⁎ (−2.57) −0.00 (−0.16) 0.08⁎⁎

Buy × LargeTrade DiscretionaryBuy DiscretionaryBuy × LargeTrade

(2.14) Sell Sell × LargeTrade DiscretionarySell DiscretionarySell × LargeTrade Ln(MktCap) Ln(BooktoMarket) PastMonthRet PastYearRet Observations Adjusted R2 Fixed effects

−0.04† (−5.67) −0.13† (−5.42) −0.17 (−0.95) 0.20 (1.45) 16,080 20% Year

Sells only (2)

Both buys and sells (3) 0.90† (7.99) −0.08⁎⁎ (−2.51) −0.01 (−0.19) 0.08⁎⁎

0.83⁎ (1.90) 0.02 (0.08) 0.03 (0.32) −0.06 (−0.33) −0.04⁎ (−1.73) −0.11† (−3.56) −0.82⁎⁎ (−2.10) −0.38 (−1.24) 2206 26% Year

(2.03) 0.81† (6.40) −0.05 (−0.30) −0.02 (−0.29) −0.00 (−0.01) −0.04† (−4.72) −0.12† (−5.76) −0.22 (−1.26) 0.15 (1.11) 18,286 21% Year

† Statistical significance at 1% level. ⁎⁎ Statistical significance at 5% level. ⁎ Statistical significance at 10% level.

transactions involving the director's own accounts. These findings support the efficacy of the ASX disclosure requirements for director trading. Because the ASX requires directors to report both direct and indirect trades in Appendix 3Y reports, there are fewer incentives to disguise informed trading through accounts that are held indirectly by the director. For sales, we find some evidence that indirect director sales underperform direct sales (Columns 2 and 3). However, the interaction term between DiscretionarySelli and IndirectTradei is insignificant, indicating that the underperformance of indirect sales is unlikely to be information driven. 3.4. Director Trades in Small vs. Large Firms Previous studies also report that the profitability of insider transactions is dependent on the firm's market capitalization. Lakonishok and Lee (2001) show that insiders are more successful in predicting returns for smaller firms. Similarly, Etebari et al. (2004) find that most of the abnormal returns in their sample are driven by insider transactions from smaller, less researched companies. Chang and Watson (2015) show that delayed reporting of purchases in small firms signals positive future returns while only director sales in large firms predict future negative returns. To test the effects of firm size on the relation between the motivation of the trade and future abnormal returns, we regress: CAR½0; 200i ¼ β1 Buyi þ β2 Buyi  SmallCapi þ β3 DiscretionaryBuyi þ β4 DiscretionaryBuyi  SmallCapi þ β5 Selli þ β6 Selli  SmallCapi þ β7 DiscretionarySelli þ β8 DiscretionarySelli  SmallCapi þ ∑βk Controlsi þ ϵi

ð4Þ

where SmallCapi is an indicator variable equal to 1 if the firm's market capitalisation is below the mean and zero otherwise. In contrast to the previous literature, we find that director purchases in small firms underperform purchases in large firms (Table 7, Column 1). However, conditioning on whether the trade is discretionary or non-discretionary, we find that discretionary

36

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

Table 6 Returns to indirect vs direct director trades. This table reports estimates of CARs for directors' trades over the 200 days following the trade based on whether the trade is a direct holding in the directors own name or indirect holding held through a trust, company or in a related party's name, such as a wife or child. IndirectTrade is a dummy variable equals 1 for indirect holdings and 0 otherwise. Buy (Sell) is a dummy variable that equals 1 when shares are solely acquired (disposed) and 0 otherwise. DiscretionaryBuy (DiscretionarySell) is a dummy variable that equals 1 for discretionary director trades where shares are solely acquired (disposed) and 0 otherwise. MktCap is the market capitalization in millions of AUD of the firm 30 days prior to the directors' trades. BooktoMarket is the book-to-market ratio for the firm as of the last financial year prior to the trade. PastMonthRet is the CAR from trading day −25 to 0, and PastYearRet is the CAR from day −200 to −26. The parentheses report t-statistics based on standard errors double clustered by GICS industry and year (Thompson, 2011). Variable

Buys only (1)

Buy

0.70† (6.30) −0.02 (−1.28) 0.04† (2.71) 0.01 (0.70)

Buy × IndirectTrade DiscretionaryBuy DiscretionaryBuy × IndirectTrade Sell Sell × IndirectTrade DiscretionarySell DiscretionarySell × IndirectTrade Ln(MktCap) Ln(BooktoMarket) PastMonthRet PastYearRet Observations Adjusted R2 Fixed effects

−0.03† (−4.91) −0.11† (−5.33) 0.06 (0.75) 0.36† (7.22) 33,572 17% Year

Sells only (2)

Both buys and sells (3) 0.67† (5.60) −0.02 (−1.30) 0.04† (2.76) 0.02 (0.73) 0.66† (5.13) −0.09⁎⁎

0.54† (2.97) −0.08⁎⁎ (−2.19) −0.02 (−0.67) 0.02 (0.47) −0.03⁎⁎ (−2.22) −0.12† (−5.48) −0.52⁎⁎

(−2.19) −0.03 (−1.04) 0.02 (0.47) −0.03† (−4.41) −0.11† (−5.50) −0.03⁎

(−1.97) −0.14 (−0.83) 7231 19% Year

(−0.32) 0.28† (6.07) 40,803 18% Year

† Statistical significance at 1% level. ⁎⁎ Statistical significance at 5% level. ⁎ Statistical significance at 10% level.

purchases in small capitalisation firms generate higher abnormal returns than discretionary purchases in large cap firms (Columns 1 and 3). This finding is consistent with inside information being more valuable in smaller, less researched firms. For sales, we find that insider selling in smaller firms outperforms insider sales in larger firms, but we find no differences in the information content between discretionary sales in small or large firms (Columns 2 and 3). 3.5. Director Trades in High vs. Low Information Asymmetry Firms The firm's market capitalisation reflects the information environment in which the firm operates. For example, small firms tend to have fewer analysts following them and are over represented in growth industries. In Table 8, we use an alternative proxy, the firm's industry classification, to distinguish between high- and low-information-asymmetry sectors. Based on GICS codes, we classify firms with primary businesses belonging to the Energy, Materials, Pharmaceuticals and Biotechnology, or Software and Service industries as firms with high information asymmetry. Firms belonging to other industry sectors are defined as low information asymmetry. We estimate the following equation: CAR½0; 200i ¼ β1 Buyi þ β2 Buyi  HighInfoAsymmetryi þ β3 DiscretionaryBuyi þ β4 DiscretionaryBuyi  HighInfoAsymmetryi þ β5 Selli þ β6 Selli  HighInfoAsymmetryi þ β7 DiscretionarySelli þ β8 DiscretionarySelli  HighInfoAsymmetryi þ ∑βk Controlsi þ ϵi ð5Þ where HighInfoAsymmetryi is an indicator variable equal to 1 if the trade is in a firm that is in a high information asymmetry industry, and zero otherwise. In Table 8, Column 1, we find a negative coefficient on the Buyi × HighInfoAsymmetryi interaction variable indicating that purchases (both discretionary and non-discretionary) in high-information-asymmetry firms do not outperform purchases in lowinformation-asymmetry firms. On the other hand, discretionary purchases in firms with high-information-asymmetry generate

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

37

Table 7 Returns to directors trades in small vs large firms. This table reports estimates of CARs for directors' trades over the 200 days following the trade based on whether the trade is in a small firm (smaller than the mean firm size across our sample) or a large firm (larger than the mean firm size). SmallCap is a dummy variable that equals 1 for small firms and 0 otherwise. Buy (Sell) is a dummy variable which takes 1 when shares are solely acquired (disposed) and 0 otherwise. DiscretionaryBuy (DiscretionarySell) is a dummy variable that equals 1 for discretionary director trades where shares are solely acquired (disposed) and 0 otherwise. MktCap is the market capitalization in millions of AUD of the firm 30 days prior to directors' trades. BooktoMarket is the book-to-market ratio for the firm as of the last financial year prior to the trade. PastMonthRet is the CAR from trading day −25 to 0, and PastYearRet is the CAR from day −200 to −26. The parentheses report t-statistics based on standard errors double clustered by GICS industry and year (Thompson, 2011). Variable

Buys only (1)

Buy

1.08† (5.60) −0.19† (−3.15) −0.01 (−0.38) 0.07⁎ (1.82)

Buy × SmallCap DiscretionaryBuy DiscretionaryBuy × SmallCap Sell Sell × SmallCap DiscretionarySell DiscretionarySell × SmallCap Ln(MktCap) Ln(BooktoMarket) PastMonthRet PastYearRet Observations Adjusted R2 Fixed effects

−0.04† (−5.21) −0.11† (−5.31) 0.05 (0.59) 0.36† (7.03) 33,572 18% Year

Sells only (2)

0.92† (2.75) −0.15 (−1.18) 0.05 (0.38) −0.07 (−0.53) −0.04⁎⁎ (−2.57) −0.11† (−5.32) −0.54⁎⁎ (−2.15) −0.15 (−0.88) 7231 19% Year

Both buys and sells (3) 1.07† (5.14) −0.19† (−2.98) −0.01 (−0.44) 0.07⁎⁎ (1.98) 1.00† (5.80) −0.15 (−1.22) 0.05 (0.35) −0.08 (−0.52) −0.04† (−4.81) −0.11† (−5.48) −0.04 (−0.49) 0.28† (5.85) 40,803 18% Year

higher abnormal returns than discretionary purchases in firms with low information asymmetry (Column 1). Supporting this finding, the interaction term between DiscretionaryBuyi and the HighInfoAsymmetryi indicator remains positive and significant in the sample of both director purchases and sales (Column 3). This result indicates that insiders tend to be more informed when trading in the securities of firms that have more information asymmetry, but only when they are trading with discretion. Firms in low-information-asymmetry industries are easier to value, which reduces the profitability of inside information. Turning to the director sales, we do not find that discretionary sales in firms with high information asymmetry underperform discretionary sales in low-information-asymmetry firms (Column 3).

4. Conclusion Previous studies document that the profitability of insider transactions depend largely on the motivation behind the trade (Etebari et al., 2004; Cheng et al., 2007; Cohen et al., 2012; Tirapat and Visaltanachoti, 2013; Chang and Watson, 2015). However, these studies can only distinguish between liquidity- and information-motivated trading using patterns in director trading behaviour or the timeliness of trade report filings and are likely to misclassify non-discretionary but non-recurring events such as rights issues. In this study, we use director filings from the ASX that uniquely require directors to state a motivation for their trade. We apply machine-learning algorithms to classify trading motives provided in more than 60,000 Appendix 3Y filings submitted to the ASX. We find that discretionary purchases generate higher abnormal returns than non-discretionary purchases. Further, for discretionary purchases, we report higher abnormal returns for large trades, and for purchases of stock in small firms or firms in industries with greater information asymmetry. In contrast, we do not detect a significant difference between the information content of discretionary and non-discretionary insider sales. In many financial markets around the world, corporate insiders are not required to disclose the nature of their holdings. Our results support the need for increased regulation on the disclosure of insider transactions to ensure that corporate insiders are not trading on private information. Also, the machine-learning algorithms we use have a wide range of applications to other fields of finance. For example, many recent studies in corporate finance rely on manually coding information from text documents, which introduces error and can involve subjective judgement. The application of machine-learning algorithms can significantly improve the efficiency and accuracy of the data collection process.

38

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

Table 8 Returns to directors' trades in firms with high and low asymmetric information. This table reports estimates of CARs for directors' trades over the 200 days following the trade based on whether the trade is in a firm with high-asymmetric information (firms whose GICS identifies them as operating in the energy, materials, biotechnology or technology industries) or low information asymmetry (all remaining industries). HighInfoAsymmetry is a dummy variable that equals 1 for firms with high information and 0 otherwise. Buy (Sell) is a dummy variable that equals 1 when shares are solely acquired (disposed) and 0 otherwise. DiscretionaryBuy (DiscretionarySell) is a dummy variable that equals 1 for discretionary director trades where shares are solely acquired (disposed) and 0 otherwise. MktCap is the market capitalization in millions of AUD of the firm 30 days prior to the director's trade. BooktoMarket is the book-to-market ratio for the firm as of the last financial year prior to the trade. PastMonthRet is the CAR from trading day −25 to 0, and PastYearRet is the CAR from day −200 to −26. The parentheses report t-statistics based on standard errors double clustered by GICS industry and year (Thompson, 2011). Variable

Buys only (1)

Buy

0.74† (6.93) −0.06† (−2.94) 0.02 (0.75) 0.06⁎ (1.75)

Buy × HighInfoAsymmetry DiscretionaryBuy DiscretionaryBuy × HighInfoAsymmetry Sell Sell × HighInfoAsymmetry DiscretionarySell DiscretionarySell × HighInfoAsymmetry Ln(MktCap) Ln(BooktoMarket) PastMonthRet PastYearRet Observations Adjusted R2 Fixed effects † Statistical significance at 1% level. ⁎⁎ Statistical significance at 5% level. ⁎ Statistical significance at 10% level.

Appendix 1. Appendix

Fig. A1 Example of completed ASX Appendix 3Y form.

−0.03† (−5.37) −0.11† (−5.32) 0.06 (0.63) 0.36† (5.47) 32,854 18% Year

Sells only (2)

0.58† (3.16) −0.07 (−0.84) −0.04 (−0.60) 0.03 (0.35) −0.03⁎⁎ (−2.44) −0.11† (−5.23) −0.53⁎⁎ (−2.31) −0.15 (−1.06) 7119 19% Year

Both buys and sells (3) 0.71† (6.03) −0.06† (−3.04) 0.02 (0.78) 0.06⁎ (1.88) 0.68† (4.64) −0.05 (−0.67) −0.04 (−0.60) 0.02 (0.22) −0.03† (−4.77) −0.11† (−5.43) −0.03 (−0.37) 0.28† (4.57) 39,973 18% Year

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

39

Appendix 2. Appendix

Table A2 Data extracted from ASX form shown in Appendix 1. Field

Raw data extracted

Processed data

Name of Director Date of last notice Nature of indirect interest Date of change No. of securities held prior to change Class Number acquired Number disposed Value/Consideration No. of securities held after change Nature of change

Dr John Stanley Keniry 28 December 2004 Direct 20 January 2005 527,499 ordinary shares (RIC) – 1786 ordinary shares – $1.40 per share 529,285 ordinary shares (RIC) Monthly acquisition under Employee Share Acquisition Plan

Dr John Stanley Keniry 28 December 2004 Direct 20 January 2005 527,499 – 1786 0 $2500.40 529,285 Remuneration

Appendix 3. Appendix

Table A3 Transaction categories. We classify 63,607 director transactions into 12 transaction categories based on the trading motive provided by the insider. The table also provides examples of text strings from the original filing. Category

Examples of text from original filing

Consolidation

‘Consolidation of capital’ ‘Share consolidation on a basis of 1 new share for every 5 shares held’ ‘Executor in respect of the Estate’ ‘Off-market inheritance from deceased estate’ ‘Dividend Reinvestment Plan Allotment’ ‘Plan in lieu of cash dividends’ ‘Release from escrow’ ‘Vesting’ ‘Conversion of 1 cent options into ordinary shares.’ ‘Cancellation of share options’ ‘Off market trade’ ‘Off market purchase’ ‘Acquired on market’ ‘Purchase of shares on ASX’ ‘Participation in Employee share purchase plan’ ‘Fully paid ordinary shares issued as a result of the satisfaction of the Class 1 and Class 2 performance shares milestone’ ‘Subscription for shares pursuant to shareholder approval’ ‘Purchase of shares under share purchase plan’ ‘Acceptance of an off-market takeover bid’ ‘Sale of share pursuant to the takeover offer’ ‘Change of registered holder’ ‘Off market transfer between family entities’ ‘Participation in entitlement placement of shortfall offer’ ‘Underwriting of rights issue’

Death Dividend reinvestment plan (DRP) Escrow Exercise/expiry Off-market On-market Remuneration Rights/SPP Takeover Transfer Underwriting

40

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

Appendix 4. Stability of the SVM Model We formally test the stability of the SVM model using a cross-validated sample. Specifically, we compute two measures, precision and recall, which are commonly used in the computer science literature. Precision refers to the percentage of time the algorithm correctly predicts the category. In contrast, recall refers to the percentage of a category that the algorithm correctly classifies. Following the procedure described in Sokolova et al. (2006), we calculate the harmonic mean of recall and precision values to arrive at a combined F-score, which is bounded between 0 (lowest performance) and 1 (highest performance). Table A4, which presents the precision, recall, and F-score values for the 12 categories, shows that the SVM classifications are highly accurate across the full sample. Excluding ‘death’ and ‘off-market’ categories, we find that the F-scores range from 0.87 to 0.97. For the ‘death’ and ‘off-market’ categories, we find that the F-scores are relatively low at 0.32 and 0.64, respectively. Because ‘death’ and ‘off-market’ comprises less than 0.15% and 5% of the sample, respectively, these misclassifications are unlikely to affect our results. Further, we find that the low F-scores are driven by low recall, rather than low precision. For our analysis, precision is more important because we are more concerned about the accuracy of classifying each observation correctly (precision), rather than capturing all observations in a given category (recall). Additionally, recall values tend to be lower for small categories.10 Table A4 Stability of the SVM model. This table presents the precision, recall, and F-score values for our 12 transaction categories following the procedure of Sokolova et al. (2006). The harmonic mean of recall and precision values is used to generate the combined F-score, which is bounded between 0 (lowest performance) and 1 (highest performance). Category

Precision

Recall

F-score

Consolidation Death Dividend reinvestment plan (DRP) Escrow Exercise/expiry Off-market On-market Remuneration Rights/SPP Takeover Transfer Underwriting

0.88 1 0.98 0.97 0.96 0.98 0.95 0.97 0.97 0.99 0.96 0.97

0.9 0.19 0.96 0.82 0.97 0.48 1 0.94 0.96 0.96 0.97 0.79

0.89 0.32 0.97 0.89 0.96 0.64 0.97 0.95 0.96 0.97 0.96 0.87

Appendix 5. Director Trading Over Different Sub-periods To investigate whether the profitability of discretionary director trades changes over time, we divide our sample into three distinct periods: pre-crisis (January 2003 to March 2007), crisis (April 2007 to December 2008) and post-crisis (January 2009 onwards). Figs. A5.1 and A5.2 show the cumulative abnormal returns (CARs) for discretionary purchases and sales, from the day of the transaction to 200 trading days after the transaction for the three time periods, in turn. While there is some evidence that the profitability of discretionary trades is disappearing over time, discretionary trades remain profitable towards the end of the sample period. For purchases in Fig. A5.1, we see that discretionary buys are highly profitable in the pre-crisis period (200-day CAR = 12.0%) and remain moderately profitable over the post-crisis period (200-day CAR = 4.2%). Interestingly, during the crisis, we find that discretionary buys are initially unprofitable but become slightly profitable over the longer term horizon (200-day CAR = 1.7%). Together, these results indicate that purchases during the crisis may not be completely information driven. For example, directors could be purchasing shares to support a falling stock price, which is only marginally profitable after a longer time horizon. Turning to the discretionary sales in Fig. A5.2, we also find that discretionary sales are less profitable over time. The 200-day CARs for discretionary sales in the pre- and post-crisis periods are −16.2% and −12.4%, respectively. During the crisis period, directors are reluctant to sell shares, which could further depress a falling stock price. Thus, directors are only selling when there is a very negative outlook on firm performance, which is reflected in a large negative CAR (−20.4%) during the crisis period. To formally investigate whether the profitability of discretionary trades changes through time, Table A5 reports our main results based on the pre-crisis, crisis, and post-crisis sub-periods. Consistent with our main results from Table 4, we find that discretionary purchases result in higher abnormal returns than non-discretionary purchases across all sub-periods. Again, we do not find that discretionary sales outperform non-discretionary sales after including the control variables. In unreported results, we form interaction variables to test whether the DiscretionaryBuyi indicator variable changes between the different subperiods. Confirming our observations from Fig. A5.1, we find some evidence that the profitability of discretionary purchases

10 Assume there are 10 observations, of which 2 should be classified as ‘type A’ and 8 as ‘type B’. Further assume that we classify 90% of the observations correctly: 1 observation as ‘type A’ (correct) and the other 9 observations as ‘type B’ (8 correct, 1 incorrect). For precision, ‘type A’ and ‘type B’ have values of 1 and 0.89, respectively. For recall however, the smaller category, ‘type A’, has a value of only 0.5 while ‘type B’ has a value of 1.

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

41

decreases over time. Specifically, our results show that the profitability of discretionary purchases is lower in the post-crisis period, relative to both the pre-crisis and crisis periods (p-value = 0.001 and 0.057, respectively). However, we do not find a significant difference between the pre-crisis and crisis CARs for discretionary purchases.

Fig. A5.1. The profitability of discretionary insider purchases over different sub-periods. This figure shows the cumulative abnormal returns (CARs) for discretionary purchases from the day of the transaction to 200 trading days after the transaction. CARs are estimated based on a market model for a 200 day estimation period ending 10 trading days before the director trade. The sample is divided into three distinct periods: pre-crisis (January 2003 to March 2007), crisis (April 2007 to December 2008) and post-crisis (January 2009 onwards).

Fig. A5.2. The profitability of discretionary insider sales over different sub-periods. This figure shows the cumulative abnormal returns (CARs) for discretionary sales from the day of the transaction to 200 trading days after the transaction. CARs are estimated based on a market model for a 200 day estimation period ending 10 trading days before the director trade. The sample is divided into three distinct periods: pre-crisis (January 2003 to March 2007), crisis (April 2007 to December 2008) and postcrisis (January 2009 onwards).

42

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

Table A5 Returns to directors' trades around the global financial crisis. This table reports estimates our model for CARS for directors' trades over the 200 days following the trade based on whether the trade is prior to the GFC (1/1/2005–1/4/ 2007), during (1/4/2007–1/1/2009) or after (post 1/1/2009). The model specification is identical to Eq. (1), with both buy and sell trades included. Buyi (Selli) is a dummy variable that equals 1 when shares are solely acquired (disposed) and 0 otherwise. DiscretionaryBuyi (DiscretionarySelli) is a dummy variable that equals 1 for discretionary director trades where shares are solely acquired (disposed) and 0 otherwise. MktCapi is the market capitalization in millions of AUD of the firm 30 days prior to the director's trade. BooktoMarketi is the book-to-market ratio for the firm as of the last financial year prior to the trade. PastMonthReti is the CAR from trading day −25 to 0, and PastYearReti is the CAR from day −200 to −26. The parentheses report t-statistics based on standard errors double clustered by GICS industry and year (Thompson, 2011). Pre-GFC

During GFC

Post-GFC

PastYearRet

0.52† (3.36) 0.11† (3.03) 0.43⁎⁎ (2.46) −0.04 (−0.61) −0.03† (−4.40) −0.05 (−1.56) −0.02 (−0.19) 0.09⁎

Observations Adjusted R2 Fixed effects

(1.66) 5884 6% Year

1.56† (10.67) 0.08† (2.63) 1.53† (11.34) −0.03 (−1.18) −0.04† (−3.78) −0.10⁎⁎ (−2.04) −0.39⁎⁎ (−2.04) 0.14 (1.33) 13,369 20% Year

0.56† (4.05) 0.03⁎ (1.91) 0.52† (3.36) −0.02 (−0.36) −0.02† (−2.89) −0.13† (−9.42) 0.15⁎ (1.68) 0.34† (4.65) 21,550 17% Year

Buy DiscretionaryBuy Sell DiscretionarySell Ln(MktCap) Ln(BooktoMarket) PastMonthRet

† Statistical significance at 1% level. ⁎⁎ Statistical significance at 5% level. ⁎ Statistical significance at 10% level.

References Acharya, V.V., Johnson, T.C., 2010. More insiders, more insider trading: evidence from private-equity buyouts. J. Financ. Econ. 98 (3), 500–523. Aspris, A., Foley, S., Frino, A., 2014. Does insider trading explain price run-up ahead of takeover announcements? Account. Finance 54 (1), 25–45. Betzer, A., Gider, J., Metzger, D., Theissen, E., 2015. Stealth trading and trade reporting by corporate insiders. Eur. Finan. Rev. 19 (2), 865–905. Burges, C.C., 1998. A tutorial on support vector machines for pattern recognition. Data Min. Knowl. Disc. 2 (2), 121–167. Chang, M., Watson, I., 2015. Delayed disclosure of insider trades: incentives for and indicators of future performance? Pac. Basin Financ. J. (forthcoming). Cheng, S., Nagar, V., Rajan, M.V., 2007. Insider trades and private information: the special case of delayed-disclosure trades. Rev. Financ. Stud. 20 (6), 1833–1864. Chiang, M.-H., Hwang, L.-J., Wu, Y.-C., 2004. Insider trading performance in the Taiwan stock market. Int. J. Bus. Econ. 3 (3), 239–256. Cohen, L., Malloy, C., Pomorski, L., 2012. Decoding inside information. J. Financ. 67 (3), 1009–1043. Eckbo, B.E., Smith, D.C., 1998. The conditional performance of insider trades. J. Financ. 53 (2), 467–498. Etebari, A., Tourani-Rad, A., Gilbert, A., 2004. Disclosure regulation and the profitability of insider trading: evidence from New Zealand. Pac. Basin Financ. J. 12 (5), 479–502. Finnerty, J.E., 1976. Insiders and market efficiency. J. Financ. 31 (4), 1141–1148. Friederich, S., Gregory, A., Matatko, J., Tonks, I., 2002. Short-run returns around the trades of corporate insiders on the London Stock Exchange. European Financial Management 8 (1), 7–30. Heitzman, S., Klasa, S., 2015. Trading Activity and Abnormal Stock Returns Surrounding Private Merger Negotiations. Huddart, S., Ke, B., Shi, C., 2007. Jeopardy, non-public information, and insider trading around sec 10-K and 10-Q filings. J. Account. Econ. 43 (1), 3–36. Jaffe, J.F., 1974. Special information and insider trading. J. Bus. 47 (3), 410–428. Jeng, L.A., Metrick, A., Zeckhauser, R., 2003. Estimating the returns to insider trading: a performance-evaluation perspective. Rev. Econ. Stat. 85 (2), 453–471. Joachims, T., 1998. Text categorization with support vector machines: learning with many relevant features. In: Nédellec, C., Rouveirol, C. (Eds.), Machine Learning: Ecml-98. Springer, Berlin Heidelberg, pp. 137–142. John, K., Lang, L.H.P., 1991. Insider trading around dividend announcements: theory and evidence. J. Financ. 46 (4), 1361–1389. Kallunki, J.-P., Nilsson, H., Hellström, J., 2009. Why do insiders trade? Evidence based on unique data on Swedish insiders. J. Account. Econ. 48 (1), 37–53. Ke, B., Huddart, S., Petroni, K., 2003. What insiders know about future earnings and how they use it: evidence from insider trades. J. Account. Econ. 35 (3), 315–346. Lakonishok, J., Lee, I., 2001. Are insider trades informative? Rev. Financ. Stud. 14 (1), 79–111. Lin, J.-C., Howe, J.S., 1990. Insider trading in the OTC market. J. Financ. 45 (4), 1273–1284. Lorie, J.H., Niederhoffer, V., 1968. Predictive and statistical properties of insider trading. J. Law Econ. 11 (1), 35–53. Marin, J.M., Olivier, J.P., 2008. The dog that did not bark: insider trading and crashes. J. Financ. 63 (5), 2429–2476. Meulbroek, L.K., 1992. An empirical analysis of illegal insider trading. J. Financ. 47 (5), 1661–1699. Piotroski, J.D., Roulstone, D.T., 2005. Do insider trades reflect both contrarian beliefs and superior knowledge about future cash flow realizations? J. Account. Econ. 39 (1), 55–81. Pope, P.F., Morris, R.C., Peel, D.A., 1990. Insider trading: some evidence on market efficiency and directors' share dealings in Great Britain. Journal of Business Finance & Accounting 17 (3), 359–380. Rogoff, D.L., 1964. The forecasting properties of insiders' transactions. J. Financ. 19 (4), 697–698. Rozeff, M.S., Zaman, M.A., 1988. Market efficiency and insider trading: new evidence. J. Bus. 61 (1), 25–44. Seyhun, H.N., 1986. Insiders' profits, costs of trading, and market efficiency. J. Financ. Econ. 16 (2), 189–212. Seyhun, H.N., 1988. The information content of aggregate insider trading. J. Bus. 61 (1), 1–24.

S. Foley et al. / Pacific-Basin Finance Journal 39 (2016) 28–43

43

Sokolova, M., Japkowicz, N., Szpakowicz, S., 2006. Beyond accuracy, F-score and ROC: a family of discriminant measures for performance evaluation. Proceedings of the 19th Australian Joint Conference on Artificial Intelligence: Advances in Artificial Intelligence. Springer-Verlag, Hobart, Australia. Thompson, S.B., 2011. Simple formulas for standard errors that cluster by both firm and time. J. Financ. Econ. 99, 1–10. Tirapat, S., Visaltanachoti, N., 2013. Opportunistic insider trading. Pac. Basin Financ. J. 21 (1), 1046–1061. Yang, Y., Liu, X., 1999. A re-examination of text categorization methods. Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, Berkeley, California, USA, 312647, pp. 42–49.