On the efficiency of intra-industry information transfers: The dilution of the overreaction anomaly

Journal of Banking & Finance 60 (2015) 153–167

Contents lists available at ScienceDirect

Journal of Banking & Finance journal homepage: www.elsevier.com/locate/jbf

On the efficiency of intra-industry information transfers: The dilution of the overreaction anomaly Dennis Y. Chung ⇑, Karel Hrazdil, Kim Trottier Beedie School of Business, Simon Fraser University, Canada

a r t i c l e

i n f o

Article history: Received 5 August 2013 Accepted 8 August 2015 Available online 12 August 2015 JEL classification: G12 G14 Keywords: Information transfers Market efficiency Overreaction anomaly

a b s t r a c t We revisit the stock market anomaly documented by Thomas and Zhang (2008) and show that the apparent mispricing of information transfers has decayed over time, as the US markets experienced rapid improvements in the efficiency of the underlying price formation processes. Utilizing recent advancements in market microstructure research to estimate firm-specific proxies for market efficiency, we demonstrate that the existence of the overreaction anomaly (where stock prices of late announcers in response to the earnings reported by early announcers in the same industry are negatively related to subsequent price responses of late announcers to their own earnings reports) is specific to an earlier sample period and results from the inefficient incorporation of information into prices, largely attributable to an environment with high barriers to arbitrage. Our results indicate that the pricing efficiency of intraindustry information transfers has increased in the recent years of increased liquidity and markedly higher trading activity. Ó 2015 Elsevier B.V. All rights reserved.

1. Introduction Earnings information transfers occur when one firm’s earnings announcement provides valuation-relevant information about another firm’s earnings. Prior empirical research documents that earnings announcements provide information not only about the announcing firm but also about other firms in the same industry (e.g., Foster, 1981; Han and Wild, 1990; Freeman and Tse, 1992; Ramnath, 2002). In a more recent study, Thomas and Zhang (2008, hereafter TZ) investigate how prices of firms that have not yet announced earnings (late announcers) respond to earnings announcements of their early-announcing industry peers. TZ are the first to document an overreaction anomaly in which the stock market overestimates the intra-industry implications of early announcers’ earnings for late announcers’ earnings and the over-reaction is corrected when late announcers subsequently disclose their earnings. Specifically, TZ show that the price movements of late announcers in response to earnings reported by early announcers are significantly negatively related to the price responses to their own earnings reports. ⇑ Corresponding author. Tel.: +1 778 782 4355; fax: +1 778 782 4920. E-mail addresses: [email protected] (D.Y. (K. Hrazdil), [email protected] (K. Trottier). http://dx.doi.org/10.1016/j.jbankfin.2015.08.013 0378-4266/Ó 2015 Elsevier B.V. All rights reserved.

Chung),

[email protected]

Surprisingly, TZ also find that the investor overreaction to intraindustry information transfers is surrounded by other positive own-firm and cross-firm return relations, which imply investor underreaction to all other earnings news. The apparent mispricing raises the question of why sophisticated investors do not take advantage of the mispricing opportunities and reinforce market efficiency through arbitrage activities. The objective and contribution of our study is to re-examine the overreaction anomaly documented by TZ. We build on TZ and posit that the apparent mispricing of information transfers is related to the level of efficiency with which stock prices reflect the new information. Specifically, in the presence of inefficiencies, the documented overreaction may have been the consequence of the market’s inability to execute orders at the level of precision required to eliminate arbitrage. We expect the information transfers to be more complete for more-efficient firms that face lower arbitrage risk and to have strengthened over time as the US markets experienced rapid improvements in the efficiency of the underlying price formation processes. Recent evidence by Chordia et al. (2014) suggests that many popular return anomalies (i.e., momentum, monthly reversals, analyst dispersion, post-earnings announcement drift, and accounting accruals) have materially diminished in both strength and significance, as the pricing efficiency of equities has increased in the

154

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

recent regime of increased liquidity and markedly higher trading activity.1 Through their analysis of trends in cross-sectional predictability, Chordia et al. (2014) demonstrate that the Fama–MacBeth cross-sectional coefficient estimates and the decile-based hedge portfolio returns have attenuated from before to after decimalization for the stocks traded on major US exchanges and that many portfolio strategies have experienced markedly diminished reward-risk ratios over time. To test our market efficiency hypothesis, we use an innovative and more comprehensive measure of limits to arbitrage, than previously used, in the context of the overreaction anomaly. Recent developments in market microstructure give us a basis to explore the evolution of the price formation process and to study how efficiently information is incorporated into security prices. We rely on the market microstructure approach developed by Chordia et al. (2008, hereafter CRS), who conclude that short-horizon return predictability (SHRP) from historical order flows is an inverse indicator of market efficiency.2 The return predictability approach intends to capture the time over which market participants observe and extract information from order flows, ascertain whether there is new relevant information about firm values, take advantage of any predictable price movements, and eliminate any serial return dependence remaining after prices adjust to their new equilibrium levels. Because information is impounded in stock prices through trades, the CRS estimation of return predictability is a direct approach to assessing the efficiency of market makers, specialists, and arbitrageurs in processing current earnings information.3 This innovative approach based on market microstructure allows us to study the nature of the overreaction anomaly documented by TZ and to test whether this anomaly changed over time as the market became more efficient.4 We first replicate TZ with the same period of 1973–2005 and confirm that price movements of late announcers in response to earnings reported by early announcers (RESP; the coefficient of interest) are significantly negatively related to the price responses 1 Commonly, the periods of improved trading activity are associated with exogenous decreases in bid-ask spreads. Chordia et al. (2008) and Chung and Hrazdil (2010a, b) examine the efficiency of the price formation process across three tick size regimes identified by exogenous decreases in bid-ask spreads, and they document a substantial improvement in market efficiency for the NYSE and NASDAQ stocks. 2 The use of this methodology has become established in the literature. For instance, Aktas et al. (2008) use the order imbalance and return predictability relation to examine the effects of insider trading on market efficiency; Visaltanachoti and Yang (2010) analyze the speed of convergence to market efficiency for foreign stocks listed on the NYSE; Chordia et al. (2011) analyze the evolution of return predictability over time. Chung and Hrazdil (2011) use the return predictability approach to examine the post-earnings announcement drift, and they provide comprehensive empirical evidence that the SHRP captures factors such as volatility, information asymmetry, investor sophistication, volume, size, and trading costs that affect arbitrage activities and the extent to which information is impounded in prices. These results support our motivation to use short-horizon return predictability as a broader measure of market efficiency in the context of intra-industry information transfers. In the case of the overreaction anomaly, the stronger and more positive (negative) reaction to the good (bad) news in the earnings announcements of a firm’s peers should be associated with the less positive (negative) reaction later when the firm announces its own earnings. We expect the market’s reaction to be reflected in the changing flows of buy and sell orders and that the short-horizon predictability of returns as a function of order imbalance provides a basis for assessing the informational efficiency of the market with respect to the information contained in the earnings announcements. Further review of the market efficiency estimations is provided in the following section. 3 For financial irregularities such as the accrual anomaly that spans longer time periods, the question of whether the price process is slow to respond over a fiveminute window is not really relevant for arbitrage strategies. However, for strategies designed to exploit the intra-industry information transfers over several days, a delay in the price formation process in even a short time affects their efficiency in a notable way. 4 For further details, see Chordia et al. (2011) and Chordia et al. (2014), who explore the sharp uptrend in recent trading activity and the accompanying improvements in market efficiency.

to late announcers’ own earnings reports. We then extend the TZ analysis to a more recent sample period of 1993–2010, over which detailed trade and quote (TAQ) intraday market data are available. The use of this TAQ sample provides the data and the basis to derive SHRP as an empirical measure of market efficiency. We partition our sample into two groups based on median SHRP (an inverse indicator of market efficiency) and two equal time periods (before and after the move to decimal pricing, an exogenous event that significantly improved the efficiency of the price formation process). We document in the multivariate setting that the coefficient of interest, although negative over all time periods, is statistically significant only during the pre-decimalization period (1993–2001). Dissecting this result further, we also find that regressions based on subsamples of more-efficient firms (below median SHRP) and time periods during the post-decimalization era (2002–2010) all yield insignificant RESP coefficients. We further test the robustness of our results and estimate the standard errors of the regression coefficients in all of the regressions by clustering on both the firm and the time dimensions simultaneously, an approach developed by Petersen (2009) and Thompson (2011). The robustness tests confirm our results. Overall, our findings portray an economically intuitive picture of a strong linkage between information transfers and market efficiency, and they challenge the existence of the overreaction anomaly during the post-decimalization period. Our results provide evidence that, in the recent years of markedly higher trading activity and market efficiency, intra-industry information transfers implied by the early announcer’s earnings reports are properly incorporated into the stock prices of late announcers when the late announcers subsequently report their own earnings. Our results are also consistent with the general conclusion of Chordia et al. (2014) that many popular return anomalies have diminished in strength and significance from before to after decimalization. Under the assumption that the SHRP is a valid proxy for inefficient environments with high barriers to arbitrage and the consequent extent to which information is impounded in prices, our results show that the SHRP measure captures the inefficient processing of information and can explain the predictable stock returns of industry peers before and after their earnings announcements. The remainder of this study is organized as follows. We summarize the related research and propose our empirical predictions in Section 2. Section 3 develops measures for market efficiency, defines variables of interest, and describes the data. We discuss the main empirical results and cover robustness tests in Section 4. The final section concludes and offers opportunities for future research. 2. Background and hypothesis development 2.1. Research on intra-industry information transfers Research on intra-industry information transfers examines the association between information released by a firm and the reaction of other (peer) firms in the same industry. The extant literature explores the peer firm reaction to a variety of announcements such as management forecasts (Han et al., 1989), dividend changes (Firth, 1996), dividend omissions (Caton et al., 2003), and unexpected earnings (Han and Wild, 1990). All studies provide evidence of a positive correlation between the announcing firm’s share price movement and the returns of peer firms in the same industry. The magnitude of the peer share price reaction is explored and tested using various research approaches, although Frost (1995) warns that conclusions may be affected by the test method used. For instance, Chan et al. (2007) find that the information transfer is stronger for large companies than for small firms. Freeman and Tse (1992) find that the effect is stronger for good-

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

news industries than industries where bad news is announced, and the strength of the share price reaction varies significantly across industries. Others, such as Graham and King (1996), Lang and Lundholm (1996), and Han and Wild (1997), show that the degree of information transfer is dependent upon the amount of public information available prior to the firm’s earnings announcement. Ayers and Freeman (1997) further demonstrate that the information transfer at the time of earnings announcements is muted to the extent that investors anticipate industry-wide components of earnings prior to the release of firm-specific earnings information, whereas Elgers et al. (2008) show that the effect is due to measurement error from using realized rather than expected earnings changes. Lastly, TZ find an overreaction on the peer firms’ share price in response to an early earnings announcement and a reversal of that overreaction when the peer firm announces its own earnings. Our research extends this literature and examines the overreaction to earnings news (and its reversal) with a view to exploring whether the anomaly has decayed over time, as markets experienced increased liquidity and markedly higher trading activity. 2.2. Research on short-horizon return predictability Research on market microstructure gives us a basis to explore the price formation process and to study how information is incorporated into security prices. Recent empirical studies in the market microstructure literature show increasing interest in the relation between order flows and stock returns (e.g., Chordia and Subrahmanyam, 2004; Chordia et al., 2005; Chung and Hrazdil, 2011). These studies generally analyze the determinants and properties of market-wide order imbalances in connection with future short-horizon returns, and they provide evidence that return predictability from past order imbalances is eliminated through the trading activities of specialists and arbitrageurs (floor traders), who have the advantage of immediately observing/inferring information in order flows. In their seminal work, CRS further expand the relation between order flows and return predictability, analyze return predictability in connection with liquidity, and interpret their findings from a market efficiency perspective. The authors find that improved liquidity associated with an exogenous decrease in bid-ask spreads stimulates arbitrage activity, which reduces SHRP, an inverse indicator of market efficiency. The authors attribute the positive relationship between liquidity and market efficiency to limited riskbearing capacity and/or the facing of inventory financial constraints; if market makers cannot absorb the impact of price pressures from imbalances in buy and sell orders, temporary price deviations arise, which induces return predictability and creates arbitrage profit potential. Higher liquidity resulting from exogenous decreases in bid-ask spreads allows the market makers to reduce their excess inventories, facilitates arbitrage, and encourages trading on private information, which leads to lower return predictability and higher market efficiency. This result effectively rejects the other potential explanations that liquidity would be accompanied by greater return predictability (based on Barberis et al., 1998) and that liquidity would bear no relation with market efficiency. In their subsequent paper, Chordia et al. (2011) provide evidence that trading activity has increased sharply in recent years, where most of the increase is due to increased institutional trading and the overall improvements in market quality. Lastly, Chordia et al. (2014) further demonstrate that as trading technologies have improved and as trading activity has increased over time, the SHRP has diminished, both statistically and economically. In addition to liquidity, Chung and Hrazdil (2010a,b) further validate the SHRP as a proxy for market efficiency across a large sample of NYSE and NASDAQ stocks, and they analyze several of its additional determinants: size, volume, trading frequency, and

155

information asymmetry (proxied by a high adverse selection component of the bid-ask spread). They not only document a positive association between a continuous measure of liquidity and market efficiency but also show that this effect is amplified during periods that contain new information. Their results are consistent with the theoretical framework that public information about future returns is contained in past order flows and that it may take some time for prices to fully reflect the new information that is available in the market (Chordia et al., 2005). During periods with a high level of new information (i.e., around earnings announcements), the market tends to be less efficient in incorporating the information into prices, suggesting that stock prices incorporate information slowly, which in turn leads to trends in returns over short horizons (i.e., consistent with the underreaction hypothesis). Overall, Chung and Hrazdil (2010a, b) find that over the period January 1, 1993, to June 30, 2004, their explanatory variables account for approximately 50% of the variation in return predictability for NYSE firms and approximately 30% of the variation for NASDAQ firms. In their subsequent analysis, Chung and Hrazdil (2011) identify and analyze additional proxies for information flow, arbitrage risk, investor sophistication, volume, size, and trading costs that are expected to influence arbitrage trading activity, which in turn affect market efficiency. Consistent with the arguments developed in prior studies, the authors show that SHRP captures the impact of volatility, the number of analysts following, institutional ownership, information asymmetry, and various transaction costs that impact the efficiency with which information is impounded in prices. Of all of the analyzed variables, they find that trading activity measured by trading volume is most strongly associated with the market efficiency measure. Chung and Hrazdil (2011) conclude that, compared to specific proxies for the efficiency of price discovery used in previous literature, the SHRP measure is more effective in broadly capturing the overall degree of frictions in the market and thus provides a more comprehensive approach for assessing the efficiency of market makers, traders, and arbitrageurs in their processing of current earnings information. The CRS measure of SHRP is therefore useful because it aims to parsimoniously capture the aggregate impact of all potential frictions that operate during the price formation process in the trading of a stock. In this paper, we rely on the CRS measure and use it as a direct proxy to broadly capture the overall degree of frictions in the market. We apply this measure of market efficiency to the information environment in which the overreaction to intra-industry information transfers originates. In this case, the stronger and more positive (negative) reaction to the good (bad) news in the earnings announcements of a firm’s peers should be associated with the less positive (negative) reaction later when the firm announces its own earnings.5 The firms’ reactions to earnings news are reflected in the changing flows of buy and sell orders, and how well order flow information is utilized by the market is reflected in the SHRP measure.

5 In this paper, we focus on testing the overreaction anomaly documented by TZ (for analysis of the post-earnings announcement drift underreaction anomaly, see Chung and Hrazdil, 2011). TZ propose that the overreaction to intra-industry information transfers is based on investors not fully appreciating the positive correlation in earnings disclosed by industry peers. Consistent with the representativeness heuristic bias discussed in behavioral finance (e.g., Tversky and Kahneman, 1974; Barberis et al., 1998), the stock price of the late announcers arguably adjusts via a series of price movements that are, on average, positively related, potentially causing the late announcer’s price to overshoot the price that reflects the earnings report it eventually discloses (consider how a stock market estimates the prospects for a late-announcing firm as each early announcer reports earnings that are positively correlated with the earnings disclosed by early announcing firms). Based on this argument, the degree of overreaction/mispricing increases with each subsequent early announcer, and the overreaction corrects when the late announcer reports its own earnings. Similar to TZ, we find the overreaction to be weaker for first announcers, which further supports the overreaction explanation.

156

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

Chordia et al. (2014) consider the implications for cross-sectional return predictability in the current regime of higher trading activity and postulate that increased liquidity, reduced trading costs, and improved trading technology facilitate arbitrage activities, which in turn lead to a reduction in return predictability. We apply this argument to the overreaction anomaly of intra-industry information transfers and posit that increased arbitrage activity should help reduce the overreaction-induced predictability of information transfers. Specifically, we expect the information transfers to be less efficient for firms that face higher arbitrage risk and to have become more complete over time, as the US markets experienced rapid improvements in the efficiency of the underlying price formation processes. More formally, our hypotheses (in null form) can be stated as follows: H1. There is no difference in the overreaction anomaly to intraindustry information transfers (in the form of late announcers’ share price reaction at the earnings announcement date being negatively related to their share price response to early announcers’ earnings) between the earlier years and the more recent years of the sample period. H2. There is no difference in the overreaction anomaly to intraindustry information transfers (in the form of late announcers’ share price reaction at the earnings announcement date being negatively related to their share price response to early announcers’ earnings) between firms with a high level of market frictions (less efficient in processing information) and firms with a low level of market frictions.

3. Research design and data 3.1. Sample selection and variable definitions Our analysis is based on two different samples. First, we replicate TZ and analyze 132 quarters from 1973 to 2005 with firms that have a four-digit SIC code and data available to estimate various control variables (TZ sample). Second, we construct an extended sample and analyze all firms for which the intra-day trade and quote data are available from the NYSE TAQ database (TAQ sample). Because the earliest available date of the TAQ data is January 1, 1993, we analyze a total of 72 quarter-periods between 1993 and 2010 with our TAQ sample. Both samples have the same requisites from three sources: (1) quarterly earnings and announcement dates from quarterly Compustat files, (2) daily stock return data from the Center for Research in Security Prices (CRSP), and (3) financial data from annual Compustat files. Consistent with prior research, we focus on firms with December fiscal year-ends to ensure the earnings are reported for contemporaneous fiscal quarters. Table 1 shows the total number of available firm-quarter observations for each variable, whereas the regression analysis is based on quarterly observations with all available variables for each firm: 155,927 firm-quarter observations for the TZ sample (Model 5 in Table 3 on page 920 in TZ is based on less than 174,928 firm-quarter observations) and 179,734 firm-quarter observations for our TAQ 1993–2010 sample. We closely follow TZ and estimate two sets of excess returns for each industry quarter; one set for each late-announcing firm i and one set for its industry peers. Each set contains two three-day excess returns, one in response to that firm’s own earnings announcement and one in response to the other firm’s earnings announcement, where all excess returns are computed as the excess of raw returns over value-weighted market returns in the three-day window around earnings announcements [
1, 0, 1], where day 0 is the earnings announcement date. The first set of returns, for each late-announcing firm i, includes: the excess return around its own earnings announcement

(ARET) and its return over a similar three-day window in response to the earnings announcements of other industry peers that have already announced their quarterly earnings (RESP). Since there is typically more than one industry peer that announces its earnings earlier, the average value of RESP across those peers’ announcements is used to measure RESP. As in TZ, peer firms that report within five days of each late announcer are not considered in the return estimation. The second set of returns, for industry peers, contains: the average excess returns over three-day windows (at their own earnings announcement dates) for each of peers that announce earnings before firm i (ERLYPRARET) and the average early peers’ response to the earnings announcement for firm i (ERLYPRRESP). To illustrate, suppose that an industry has four firms (a, b, c, and i) that announce their first quarter earnings on July 1, July 4, July 9, and July 12, respectively. Firm i’s early-announcing peers include firms a and b (firm c’s earnings announcement date is less than five days of firm i’s earnings announcement date). ARET is firm i’s threeday excess returns around July 12; RESP is the average of firm i’s three-day excess returns around July 1 and July 4; ERLYPRARET is the average of the three-day excess returns of firm a around July 1 and firm b around July 4; and ERLYPRRESP is the average of the three-day excess returns for firms a and b around July 12. Financial information for each late announcing firm is represented by three constructs: the firm’s accruals (ACC), market value (MV), and the book-to-market ratio (BM).6 Four additional variables are considered for control purposes: the number of early peers (NPEER), the average number of days between the earnings announcement of each peer and each late announcer (TIME), the buy-and-hold return for six months up to one week before the earnings announcement date (RET6), and the firm’s abnormal returns surrounding the prior quarter’s earnings announcement (ARETt
1). To estimate the degree of trading frictions in the market, we collect intra-day trade and quote data from the Trade and Quote (TAQ) database for all NYSE, American Stock Exchange (AMEX) and NASDAQ firms over the period January 1, 1993, to December 31, 2010.7 Following CRS, we compute stock returns over five-minute intervals using the midpoints of bid and ask prices quoted at the end of the intervals. For order imbalance, we compute two measures for each five-minute interval t: the number of trades OIB#t and the dollar trades OIB$t, defined in the following way: OIB#t ¼ f½ðNumber of buyer initiated tradest Þ
ðNumber of seller initiated tradest Þ=ðTotal number of tradest Þg

ð1Þ OIB$t ¼ f½ðDollar tradedfrombuyer initiatedtradest Þ
ðDollar tradedfromseller initiatedtradest Þ=ðTotaldollar tradedt Þg

ð2Þ 6 Variable definitions using Compustat’s nomenclature are as follows: ACC is the change in noncash working capital [constructed as the change in noncash current assets (ACT-CHE) minus the change in the current liabilities net of debt in current liabilities (LCT-DLC)] minus depreciation expense (DP) scaled by total assets (AT); MV is the number of common shares outstanding at the end of the period (CSHO) multiplied by the share price at this time (PRCC); BM is the book value of equity (CEQ) divided by MV. 7 All three exchanges experienced the same exogenous shocks to their minimum trading tick sizes during similar time periods. Before mid 1997, firms on all three exchanges (AMEX, NASDAQ, and NYSE) traded in increments of eighths of a dollar. On May 7, 1997, the AMEX became the first of the three U.S. exchanges to adopt the minimum tick of sixteenths. The NASDAQ followed next and began its new quotation system on June 2, 1997. Finally, the NYSE officially changed to the same lower ticks on June 24, 1997. In this change to a decimal regime, the NYSE and the AMEX adopted the minimum decimal tick size on January 29, 2001; the NASDAQ immediately followed with a pilot project of changing the minimum tick size for some companies with a full adoption for remaining constituents on April 9, 2001.

157

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

Table 1 Descriptive statistics. Panel A reports the descriptive statistics for the TZ sample (January 1973–December 2005), Panel B is based on the TAQ sample (January 1993–December 2010), and Panel C splits the TAQ sample into two periods, before and after decimalization (1993–2001 and 2002–2010, respectively). SHRP (inverse indicator of market efficiency) is the adjusted R2 from Eq. (3) based on individual firm time-series regressions; ARET is firm i’s three-day earnings announcement excess returns (raw returns – valueweighted market returns) over the three-day [
1, 0, 1] period, where day 0 is the earnings announcement date; ERLYPRARET is the average of early peers’ three-day earnings announcement excess returns in the same quarter, where the peers’ earnings announcement dates are at least five days prior to firm i’s earnings announcement date; RESP is the average of firm i’s three-day excess returns around its peers’ earnings announcements, where the earnings announcement dates are at least five days prior to firm i’s earnings announcement date; ERLYPRRESP is the average of early peers’ three-day excess returns around firm i’s earnings announcements; ACC represents accruals measured as the change in noncash working capital minus depreciation expense, scaled by average total assets, where the change in non-cash working capital is equal to the change in non-cash current assets minus the change in current liabilities less short-term debt; MV is the market value at the end of the prior fiscal year (in $ millions); BM is the book-to-market ratio measured as the book value of equity divided by its market value at the end of the prior fiscal year; RET6 denotes the buy-and-hold six-month stock returns up to one week before firm i’s earnings announcement date; TIME is the average number of days between firm i’s and its peers’ earnings announcement dates; and NPEER is the number of peers that announce earnings at least five days prior to firm i’s earnings announcement date. All variables except NPEER and TIME are winsorized at 1% and 99% of the respective quarterly distributions. The t-statistics in Panel C are from the parametric two-sample t-test for equality of means between the 1993–2001 and 2002–2010 subsamples, and the normal approximation z-statistics are from the nonparametric Wilcoxon Rank-Sum two-sample test for equality of medians between the two subsamples. Test statistics that are significant at the 0.01 level are indicated by an asterisk. Variable

N

Mean

St. Dev.

Min.

Q1

Median

Q3

Max.

Panel A – TZ replication sample (1973–2005) ARET 155,927 0.0012 ERLYPRARET 155,927 0.0054 RESP 155,927 0.0004 ERLYPRRESP 155,927 0.0011 ARETt
1 155,927
0.0001 ACC 155,927
0.0428 MV 155,927 532.4 BM 155,927 0.3626 RET6 155,927 0.0903 TIME 155,820 15.7413 NPEER 155,820 12.9477

0.0440 0.0426 0.0197 0.0255 0.0459 0.3619 1,966.3 0.4182 1.1682 10.0601 23.5373


0.2093
0.1435
0.0919
0.1084
0.2227
12.3442 0.1 0.0003
25.9569 6.0000 1.0000


0.0137
0.0151
0.0067
0.0130
0.0147
0.0338 11.9 0.0904
0.1015 8.7000 2.0000

0.0000 0.0029
0.0001 0.0003
0.0001
0.0089 52.7 0.2276
0.0018 12.6667 4.0000

0.0143 0.0234 0.0063 0.0140 0.0140
0.0002 241.5 0.4785 0.1019 19.3333 12.0000

0.3118 0.2709 0.1633 0.1536 0.2737 0.3970 31,844.4 4.6566 29.5854 119.0000 268.0000

Panel B – The TAQ sample (1993–2010) SHRP 179,734 ARET 179,734 ERLYPRARET 179,734 RESP 179,734 ERLYPRRESP 179,734 ARETt
1 179,734 ACC 179,734 MV 179,734 BM 179,734 RET6 179,734 TIME 179,652 NPEER 179,652

0.0463 0.0928 0.0426 0.0388 0.0251 0.0937 0.0750 6,586.4 0.5747 2.9672 11.2908 28.3593


0.0286
0.2727
0.1182
0.1026
0.0678
0.2779
0.2146 2.5 0.0084
0.8606 6.0000 1.0000

0.0016
0.0454
0.0154
0.0170
0.0128
0.0462
0.0446 130.2 0.2197
0.2595 8.3750 3.0000

0.0104
0.0011 0.0022
0.0009 0.0001
0.0015
0.0169 385.6 0.3957 0.0183 10.6000 10.0000

0.0367 0.0451 0.0238 0.0167 0.0135 0.0441 0.0021 1,286.1 0.6637 0.3395 14.6667 31.0000

0.3905 0.5526 0.2576 0.3971 0.1583 0.5389 0.2721 70,854.0 11.3999 74.6667 194.0000 268.0000

0.0291 0.0009 0.0058 0.0021 0.0009
0.0007
0.0334 2,290.4 0.5366 0.4229 13.7567 21.9463

1993–2001 subsample Variable

N

Mean

2002–2010 subsample Median

N

Mean

Median

t-Statistic

Normal approx. z-statistic

Panel C – The TAQ (1993–2001) and (2002–2010) subsamples SHRP 65,291 0.0515 0.0299 ARET 65,291 0.0030 0.0003 ERLYPRARET 65,291 0.0098 0.0055 RESP 65,291 0.0028
0.0012 ERLYPRRESP 65,291 0.0019 0.0005 ARETt
1 65,291 0.0011 0.0000 ACC 65,291
0.0433
0.0227 MV 65,291 1,991.9 294.8 BM 65,291 0.5145 0.3646 RET6 65,291 0.6845
0.0074 TIME 65,264 13.0657 10.3333 NPEER 65,264 21.3376 8.0000

114,443 114,443 114,443 114,443 114,443 114,443 114,443 114,443 114,443 114,443 114,388 114,388

0.0163
0.0004 0.0035 0.0017 0.0003
0.0017
0.0278 2,460.7 0.5492 0.2737 14.1509 22.2935

0.0048
0.0018 0.0007
0.0008
0.0002
0.0022
0.0143 451.3 0.4125 0.0323 10.7143 12.0000

140.18⁄ 7.32⁄ 28.76⁄ 5.21⁄ 12.38⁄ 5.85⁄
40.37⁄
14.67⁄
12.21⁄ 22.65⁄
22.22⁄
6.47⁄

173.06⁄ 7.09⁄ 29.47⁄
0.21 8.97⁄ 5.62⁄
44.89⁄
46.09⁄
27.87⁄
13.27⁄
12.97⁄
34.66⁄

Table 2 Correlation matrix. The sample period spans over January 1993–December 2010, and the sample consists of all NYSE, AMEX and NASDAQ firms with trade and quote data from the TAQ database. This table shows the correlation matrix of the regression variables (defined in Table 1). Pearson (above diagonal) and Spearman (below diagonal) correlations that are significant at the 0.01 level are indicated by an asterisk. MV and BM variables are log transformed.

SHRP ARET ERLYPRARET RESP ERLYPRRESP ARETt
1 ACC MV BM RET6

SHRP

ARET

ERLY PRARET

RESP

ERLY PRRESP

ARETt
1

ACC

MV

BM

RET6

1.000 0.001 0.033⁄ 0.002 0.006⁄
0.018⁄
0.042⁄
0.387⁄ 0.047⁄
0.013⁄

0.013⁄ 1.000 0.026⁄
0.007⁄ 0.174⁄ 0.029⁄
0.006⁄ 0.039⁄
0.074⁄
0.003

0.014⁄ 0.031⁄ 1.000 0.154⁄ 0.038⁄ 0.007⁄
0.011⁄
0.035⁄ 0.012⁄
0.023⁄

0.001
0.007⁄ 0.178⁄ 1.000 0.025⁄ 0.006⁄
0.008⁄
0.017⁄
0.022⁄ 0.044⁄

0.006⁄ 0.182⁄ 0.038⁄ 0.019⁄ 1.000
0.011⁄
0.009⁄
0.018⁄
0.014⁄
0.025⁄


0.003 0.022⁄ 0.005 0.012⁄
0.009⁄ 1.000
0.003 0.050⁄ 0.005 0.017⁄


0.027⁄
0.009⁄
0.010⁄
0.010⁄
0.009⁄ 0.000 1.000 0.013⁄ 0.035⁄
0.030⁄


0.133⁄ 0.010⁄
0.015⁄
0.012⁄ 0.000 0.007⁄
0.004 1.000
0.164⁄ 0.157⁄

0.049⁄
0.075⁄ 0.030⁄
0.012⁄
0.009⁄
0.006 0.017⁄
0.083⁄ 1.000
0.300⁄

0.001
0.002 0.007⁄ 0.015⁄ 0.000 0.002
0.026⁄
0.017⁄
0.096⁄ 1.000

158

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

Table 3 Deciles based on firm i’s reaction to RESP. All variables are defined in Table 1. Each quarter, we sort firms into ten deciles based on firm i’s reaction to RESP, calculate the mean values for firms in each decile, and report in each cell above the time-series mean across the quarters in the sample. Panels A and B report the values based on the TZ sample (1973–2005) and TAQ sample (1993–2010) respectively. The remaining panels only show the extreme RESP deciles and provide partitions of Panel A based on periods before and after decimalization (Panels C and D) and based on high/low efficiency estimates (Panels E and F), respectively. Panels G-J partition Panel A into four subsamples based on two periods conditional on the efficiency of price discovery. The t-statistics in parentheses below a hedge portfolio (D1–D10, with a long position in D1 stocks and a short position in D10) are based on Fama–MacBeth (the time-series distribution of hedge portfolio returns or slope coefficients estimated for each of the 72 quarters in our sample). Consistent with TZ, portfolios with fewer than 10 stocks are eliminated, and all variables except ARET are winsorized at 1% and 99%. ARET

ERLYPRARET

ERLYPRRESP

SIZERANK

BMRANK

Panel A: TZ replication sample (1973–2005) D1 (low RESP)
0.0320 D2
0.0133 D3
0.0068 D4
0.0032 D5
0.0010 D6 0.0005 D7 0.0026 D8 0.0064 D9 0.0132 D10 (high RESP) 0.0366 D1
D10
0.0684 (
44.47)

RESP

0.0022 0.0011 0.0013 0.0017 0.0010 0.0013 0.0003 0.0007 0.0009
0.0005 0.0027 (4.87)


0.0002 0.0005 0.0018 0.0022 0.0037 0.0057 0.0066 0.0080 0.0080 0.0093
0.0097 (
10.72)

0.0011 0.0006 0.0005 0.0010 0.0009 0.0012 0.0012 0.0010 0.0015 0.0012
0.0001 (
0.28)

5.1282 5.4957 5.5726 5.6749 5.7493 5.7520 5.6219 5.6077 5.4299 4.9556

5.5831 5.5262 5.5338 5.5757 5.5512 5.5532 5.4823 5.4697 5.4204 5.3189

Panel B: The TAQ Full Sample (1993–2010) D1 (low RESP)
0.0571 D2
0.0291 D3
0.0177 D4
0.0100 D5
0.0037 D6 0.0026 D7 0.0099 D8 0.0189 D9 0.0332 D10 (high RESP) 0.0740 D1
D10
0.1311 (
24.64)

0.0052 0.0013
0.0003 0.0025 0.0002 0.0018
0.0002 0.0001 0.0008 0.0019 0.0034 (1.29)


0.0014 0.0006 0.0017 0.0033 0.0047 0.0065 0.0080 0.0089 0.0130 0.0160
0.0173 (
7.02)

0.0022 0.0005 0.0002 0.0000 0.0009 0.0017 0.0010 0.0018 0.0010 0.0010 0.0011 (0.96)

5.1838 5.4671 5.6004 5.6992 5.7031 5.7464 5.6548 5.5272 5.4262 5.0159

5.5363 5.5514 5.5947 5.5917 5.6115 5.6167 5.5073 5.3949 5.3397 5.2729

Panel C: 1993–2001 Subsample D1 (low RESP)
0.0667 D10 (high RESP) 0.0835 D1
D10
0.1502 (
23.67)

0.0090 0.0045 0.0045 (1.09)

0.0005 0.0194
0.0188 (
4.41)

0.0034 0.0024 0.0011 (0.64)

5.2139 4.8591

5.4497 5.1162

Panel D: 2002–2010 Subsample D1 (low RESP)
0.0477 D10 (high RESP) 0.0648 D1
D10
0.1125 (
15.37)

0.0016
0.0007 0.0023 (0.69)


0.0031 0.0127
0.0158 (
6.16)

0.0009
0.0003 0.0011 (0.72)

5.1546 5.1684

5.6206 5.4254

Panel E: Subsample of below median SHRP (more efficient) firms D1 (low RESP)
0.0518 0.0014 D10 (high RESP) 0.0662 0.0047 D1
D10
0.1180
0.0033 (
16.69) (
0.49)


0.0034 0.0168
0.0203 (
6.43)

0.0007 0.0015
0.0009 (
0.38)

5.2562 5.3051

5.6244 5.2571

Panel F: Subsample of above median SHRP (less efficient) firms D1 (low RESP)
0.0584 0.0066 D10 (high RESP) 0.0791
0.0033 D1
D10
0.1373 0.0100 (
32.17) (2.91)

0.0009 0.0169
0.0159 (
6.75)

0.0030 0.0001 0.0029 (1.89)

5.1621 4.8879

5.5640 5.3960

Panel G: 1993–2001 Subsample of below median SHRP (more efficient) firms D1 (low RESP)
0.0710 0.0116 D10 (high RESP) 0.0872 0.0089 D1
D10
0.1583 0.0019 (
23.86) (0.26)

0.0013 0.0213
0.0204 (
3.83)

0.0063 0.0042 0.0019 (0.75)

5.3906 5.2909

5.6193 5.1526

Panel H: 1993–2001 Subsample of above median SHRP (less efficient) firms D1 (low RESP)
0.0616 0.0071 D10 (high RESP) 0.0774 0.0013 D1
D10
0.1389 0.0055 (
25.99) (1.23)

0.0026 0.0192
0.0164 (
4.11)

0.0010 0.0009 0.0000 (0.01)

5.1381 4.5085

5.4028 5.1144

Panel I: 2002–2010 Subsample of below median SHRP (more efficient) firms D1 (low RESP)
0.0461 0.0013 D10 (high RESP) 0.0607 0.0050 D1
D10
0.1067
0.0036 (
17.44) (
0.89)


0.0035 0.0153
0.0188 (
6.00)

0.0006 0.0005 0.0001 (0.07)

5.1889 5.1365

5.5129 5.3127

Panel J: 2002–2010 Subsample of above median SHRP (less efficient) firms D1 (low RESP)
0.0487 0.0022 D10 (high RESP) 0.0660
0.0044 D1
D10
0.1149 0.0066 (
13.40) (1.77)


0.0024 0.0114
0.0141 (
5.56)

0.0008
0.0011 0.0019 (1.14)

5.1341 5.1095

5.7148 5.5499

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

We classify each trade as either a buyer-initiated or sellerinitiated trade using the well-established Lee and Ready (1991) algorithm with the same modification as implemented by CRS; we match each transaction to a bid-ask quote, which is the first quote immediately prior to the trade for years after 1998 and at least five seconds prior to the trade for years before 1999.8 To analyze the informational efficiency in connection with the over-reaction to intra-industry information transfers, we use the same approach as Chung and Hrazdil (2011) and estimate the market efficiency measure from the following time-series regression on a firm level basis.

Returnt ¼ a þ b1 OrderImbalancet
1 þ b2 ðOrderImbalancet
1  Illiquidityt Þ þ et

ð3Þ

where Returnt is the 5-min holding period return over a five-minute interval t (computed using the mid-points of the first and last quotes within each trading interval) for each firm listed on an exchange. OrderImbalancet is either OIB#t or OIB$t, and Illiquidityt is a dummy variable that is coded one if the daily average effective spread for the firm is at least one standard deviation above the detrended expected effective spread for the trading day, and zero otherwise. We include the interaction variable of OrderImbalancet and Illiquidityt to account for the effect of liquidity changes on market efficiency. Consistent with CRS, the adjusted R2 from Eq. (3) utilizing dollar trades order imbalances (OIB$t) represents the SHRP (an inverse indicator of market efficiency, where higher SHRP implies lower efficiency of price discovery).9 For these time-series regressions, we require that firms have at least 30 observations within a calendar month. The corresponding SHRP for firm i that we analyze in connection with the intra-industry information transfers is derived for each quarterly announcement by averaging the monthly SHRP measures across the three months ending in the month of the earnings announcement. 3.2. Descriptive statistics We provide detailed descriptive statistics of the variables in Table 1. Consistent with TZ, all variables except NPEER and TIME are winsorized at 1% and 99% of the respective quarterly distributions in order to mitigate the impact of outliers. We further exclude the first announcers in each industry from our sample because they do not have peers announcing earnings before them. Panel 8 Some argue that changes in short sale rules and other developments have rendered the Lee and Ready algorithm less effective for signing trades (e.g., Asquith et al., 2010). As our CRS measure relies on the Lee-Ready algorithm to sign trades, we provide additional evidence on how well or poorly the Lee-Ready algorithm performs in terms of classifying trades in recent years. Specifically, we utilize a data set obtained directly from NASDAQ, which covers two years of trades in 2008 and 2009 and one week in 2010 for a sample of 120 stocks where each trade in the data set is signed to indicate whether it is initiated by a buyer or a seller. Carrion’s (2013, p. 684) indicates that ‘‘the trade signs are high quality, and are based on records of fee and rebate payments used by the exchange.” We match the trades used in our sample with the trades in the NASDAQ dataset and compare our buy/sell classifications (which we derive using the Lee-Ready algorithm) with the actual buy/sell classifications from the NASDAQ dataset. We are able to match a total of 8,086,375 trades between the two data sets and find that the Lee-Ready algorithm is correct in classifying 83.52% of the trades. This percentage is comparable to what has been reported in the literature based on data from earlier sample periods (e.g., Finucane, 2000; Ellis et al., 2000). This evidence seems to suggest that the Lee-Ready algorithm still performs reasonably well even with all the major changes (i.e., proliferation of high frequency trading) in the market in recent years. 9 The short-horizon return predictability effectively captures the degree of market inefficiency (frictions). However, for consistency of interpretation with prior studies (e.g., Boehmer and Kelley, 2009; Chung and Hrazdil, 2010a; Visaltanachoti and Yang, 2010), we refer to the lack of short-horizon return predictability as an estimate of the extent to which information is efficiently impounded in prices (i.e., market efficiency). When we use OIB# as an explanatory variable in regression (3), the results are very similar to those reported in Section 4.

159

A reports the descriptive statistics for the TZ sample (1973– 2005), whereas Panel B is based on the TAQ sample (January 1993–December 2010). Similar to Chordia et al. (2014) who test whether many well know anomalies have attenuated from before to after decimalization, our Panel C splits the TAQ sample into two equal corresponding time periods (1993–2001 and 2002– 2010, respectively). Similar to TZ, we document that the mean earnings announcement return (ARET) is close to zero during both the 1973–2005 and 1993–2010 periods. All other variables follow a normal distribution and are comparable to those of TZ. Compared to the 1973– 2005 period, our TAQ sample in a later period (1993–2010) is characterized by larger firms (MV), value firms (BM) that have more good news in the past (RET6), and more peers in their industries (NPEER). Finally, Panel C provides sample statistics for pre- and post-decimalization and shows not only a significant increase in market efficiency (decrease in SHRP) but also significantly lower market reactions to earnings announcements (decrease in ARET) and weaker intra-industry information transfers in the postdecimalization period.10 Table 2 provides Pearson and Spearman correlations among the different variables for the period 1993–2010 for our TAQ sample. Of a total of six possible correlations among the four variables (ARET, ERLYPRARET, RESP, and ERLYPRRESP) that capture the announcement returns and responses to others’ announcements, certain variables are statistically (not necessarily economically) significantly correlated, as reported by TZ.11 Similar to TZ, we first observe that the correlations for contemporaneous combinations (ARET and ERLYPRRESP, and RESP and ERLYPRARET) are positive and significant, which is consistent with intra-industry information transfers being positive on average. We also document positive and significant correlations for the three other pairs of returns that are separated in time (ARET and ERLYPRARET, ERLYPRARET and ERLYPRRESP, and RESP and ERLYPRRESP), which is consistent with the underreaction results documented in the momentum literature.12 Worth noting are the positive and significant correlations between SHRP and ERLYPRARET and between SHRP and ERLYPRRESP, which suggest that firms with inefficient information processing experience higher returns to early peers’ earnings announcements (i.e., overreact more) and higher responses to earnings announcements of late announcers (i.e., drift in a direction of earnings news). Further confirmation of the presence of momentum or own-firm underreaction is provided by the positive and significant correlations between ARET and ARETt
1, RESP and ARETt
1, ARET and RET6, and RESP and RET6. Lastly, we document significantly negative Spearman and Pearson correlations between ARET and RESP, which indicates that the stock price for the late announcer overreacts to the information transferred from the earnings released by early announcers and is subsequently corrected when that late firm announces its earnings.

10 The weaker intra-industry information transfers during the more recent period are not unexpected. If market is more efficient in incorporating information about prices, late earnings announcers should not overreact to early industry peers’ earnings (RESP), information in earnings should seep into stock prices earlier where the market is less surprised of early peers earning announcements (ERLYPRARET), and the early announcing firms should incorporate any relevant information about their earnings into prices before late announcers reveal their earnings (ERLYPRRESP). 11 The purpose of Table 2 is to confirm that the presence of ‘an island of overreaction amidst a sea of underreaction’ documented by TZ can also be observed during our sample period for which we can estimate the SHRP. We refer the reader to Section 6 of the TZ paper, which outlines several explanations for why stock prices could overreact to intra-industry information transfers and yet underreact to all other earnings information. 12 The significant positive correlation between ARET and ERLYPRARET in our Table 2 is also consistent with the lead-lag effect from large to small firms reported by Hou (2007). TZ show the same result and emphasize this finding by pointing out that early announcers tend to be larger firms (TZ, footnote 9, page 917).

160

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

4. Tests and main empirical results 4.1. Extension of Thomas and Zhang (2008) We first provide more details on the apparent overreaction in Table 3 where we sort the TAQ sample into deciles based on firm i’s reaction to RESP. Values in Panel A are based on the sample period investigated by TZ (1973–2005), whereas Panel B shows results based on our TAQ sample (1993–2010). The remaining panels report extreme RESP deciles for various subsamples partitioned on two periods before and after decimalization and based on high/low efficiency estimates. For each sample/subsample, we sort firms into ten deciles based on firm i’s reaction to RESP in each quarter, calculate the mean values for firms in each decile, and report in each cell above the time-series mean across the quarters in the subsample. Panel A provides evidence consistent with TZ, in that, as RESP increases from D1 to D10, ARET exhibits a steady decrease from 0.22% in D1 to
0.05% in D10 with a significant hedge portfolio (long position in D1 stocks and a short position in D10 stocks) returns yielding a three-day market excess return of 0.27%. In Panel B, we are unable to observe the same declining pattern in ARET across the deciles and there is no significant three-day market excess return for a hedge portfolio returns. This finding confirms the weak correlation between ARET and RESP in Table 2 and indicates that the anomaly may not be as strong as previously documented by TZ during a more recent period. Extending TZ in a bi-variate setting, we partition our TAQ sample into different subsamples based on two periods before and after decimalization and based on high/low efficiency of price discovery estimates, and attempt to find a specific area (like a

mountain or a valley) on ‘an island of overreaction amidst a sea of underreaction’. The remaining Panels C through J report the extreme RESP deciles for various subsamples, where significant hedge returns can only be observed among the less-efficient firms (Panel F), particularly during the 2002–2010 period (Panel J). While the hedge returns are not significant in all other subsamples, it appears that the overreaction is stronger for an earlier time period (0.45% in Panel C vs 0.23% in Panel D) and is driven by the efficiency of price discovery (
0.33% in Panel E vs 1.00% in Panel F), where the hedge returns are larger for less-efficient firms during each sub-period (0.19% in Panel G vs 0.55% in Panel H and
0.36% in Panel I vs 0.66% in Panel J, respectively) which makes sense intuitively as these firms are subject to more overreaction. The return differentials between D1 and D10 in all panels appear to be unrelated to differences in size and the book-to-market ratio, as the firms-quarters across the SIZERANK and BMRANK deciles are similar in terms of their average decile values (reported in the last row of each panel in the last two columns). Extending TZ to a multivariate setting, we provide a detailed reexamination of the coefficients from regressions of ARET on RESP, which TZ found to be negative and significant and which they attributed to an overreaction to intra-industry information transfers. To follow TZ as closely as possible, we include the same set of control variables. These control variables include early peers’ announcement returns (ERLYPRARET), D as a dummy equal to one if the early announcer’s news and the late announcer’s response are in consistent directions, the late announcer’s announcement returns for the prior quarter and the same fiscal quarter in the prior year (ARETt
1 and ARETt
4), log of market capitalization (SIZE), and log of book-to-market ratio (BM). To control for the price

Table 4 Regressions of ARET on RESP and control variables (based on the Fama–MacBeth test statistics). This table shows the results of estimating regression (4) with ARET as the dependent variable. The sample period spans January 1993 to December 2010, and the sample consists of all NYSE, AMEX and NASDAQ firms with trade and quote data from the TAQ database. D is a dummy equal to 1 if the early announcer’s news and the late announcer’s response are consistent. All remaining variables are defined in Table 1. T-statistics in parentheses are based on the Fama–MacBeth test statistics. ⁄⁄⁄, ⁄⁄, and ⁄ denote 1%, 5%, and 10% significant levels, respectively, based on the Fama–MacBeth (1973) approach. Results related to the main variable of interest are highlighted in bold print.

INTERCEPT RESP D D*RESP ERLYPRARET ARETt
1 ARETt
14 SIZE LOGBM RET6 ACC Adjusted R2 (%) n

INTERCEPT RESP D D*RESP ERLYPRARET ARETt
1 ARETt
14 SIZE LOGBM RET6 ACC Adjusted R2 (%) n

TZ original analysis

I TZ replication sample (1973–2005)

II TAQ sample (1993–2010)

III TAQ subsample (1993–2001)

IV TAQ subsample (2002–2010)

0.000 (0.26)
0.084 (
7.46)⁄⁄⁄
0.000 (
0.94)
0.034 (
2.12)⁄⁄ 0.023 (3.33)⁄⁄⁄ 0.035 (10.25)⁄⁄⁄
0.004 (
1.02) 0.000 (0.42) 0.002 (5.92)⁄⁄⁄ 0.002 (1.49)
0.014 (
4.97)⁄⁄⁄ 1.30 <174,928

0.000 (
0.18)
0.039 (
3.08)⁄⁄⁄
0.000 (
1.52)
0.010 (
0.55) 0.005 (1.28) 0.030 (8.19)⁄⁄⁄ 0.007 (1.83)⁄ 0.000 (3.68)⁄⁄⁄ 0.001 (6.67)⁄⁄⁄ 0.000 (
0.36)
0.033 (
12.47)⁄⁄⁄ 1.20 155,927


0.007 (
2.38)⁄⁄
0.043 (
1.63) 0.002 (1.64) 0.005 (0.13) 0.037 (2.87)⁄⁄⁄ 0.016 (2.48)⁄⁄ 0.025 (4.37)⁄⁄⁄ 0.000 (
1.08)
0.010 (
14.45)⁄⁄⁄
0.002 (
3.52)⁄⁄⁄
0.003 (
0.40) 2.80 179,734


0.003 (
0.83)
0.083 (
3.10)⁄⁄⁄ 0.002 (1.49) 0.051 (1.08) 0.038 (2.18)⁄⁄ 0.029 (2.72)⁄⁄ 0.016 (1.82)⁄
0.001 (
2.01)⁄
0.010 (
10.95)⁄⁄⁄ 0.000 (
0.42)
0.011 (
1.15) 3.10 65,291


0.011 (
2.35)⁄⁄
0.003 (
0.08) 0.001 (0.71)
0.040 (
0.64) 0.035 (1.87)⁄ 0.004 (0.56) 0.034 (4.72)⁄⁄⁄ 0.000 (0.35)
0.010 (
9.61)⁄⁄⁄
0.004 (
3.87)⁄⁄⁄ 0.005 (0.52) 2.60 114,443

V TAQ subsample (1993–2010) Below median SHRP (more efficient)

VI TAQ subsample (1993–2010) Above median SHRP (less efficient)

VII TAQ subsample (1993–2001) Below median SHRP (more efficient)

VIII TAQ subsample (1993–2001) Above median SHRP (less efficient)

IX TAQ subsample (2002–2010) Below median SHRP (more efficient)

X TAQ subsample (2002–2010) Above median SHRP (less efficient)


0.014 (
3.57)⁄⁄⁄ 0.017 (0.37) 0.001 (0.64)
0.014 (
0.20) 0.050 (2.25)⁄⁄ 0.005 (0.41) 0.033 (3.08)⁄⁄⁄ 0.000 (
0.44)
0.014 (
10.35)⁄⁄⁄
0.004 (
2.67)⁄⁄⁄ 0.006 (0.49) 6.50 89,858


0.003 (
0.86)
0.078 (
2.21)⁄⁄ 0.001 (0.57) 0.003 (0.06) 0.031 (1.68)⁄ 0.028 (3.54)⁄⁄⁄ 0.023 (3.19)⁄⁄⁄
0.001 (
1.13)
0.008 (
9.08)⁄⁄⁄
0.001 (
2.29)⁄⁄
0.013 (
1.49) 3.70 89,876


0.009 (
1.96)⁄
0.029 (
0.67) 0.004 (1.55)
0.038 (
0.45) 0.064 (2.55)⁄⁄ 0.026 (1.76)⁄ 0.020 (1.41)
0.001 (
1.73)⁄
0.015 (
12.58)⁄⁄⁄
0.001 (
1.32) 0.005 (0.33) 5.30 32,645

0.004 (0.79)
0.086 (
2.12)⁄⁄ 0.001 (0.86) 0.056 (0.96) 0.030 (1.22) 0.033 (2.64)⁄⁄ 0.011 (0.96)
0.001 (
1.25)
0.005 (
4.43)⁄⁄⁄ 0.001 (1.25)
0.013 (
0.94) 4.30 32,646


0.013 (
2.35)⁄⁄ 0.004 (0.07) 0.004 (2.38)⁄⁄ 0.020 (0.25) 0.028 (1.17)
0.027 (
2.03)⁄⁄ 0.056 (4.77)⁄⁄⁄ 0.000 (0.07)
0.012 (
7.81)⁄⁄⁄
0.006 (
2.00)⁄ 0.016 (1.37) 5.00 57,221


0.010 (
1.62)
0.027 (
0.52)
0.001 (
0.85)
0.047 (
0.59) 0.043 (1.56) 0.017 (1.82)⁄ 0.024 (2.38)⁄⁄ 0.000 (0.15)
0.009 (
7.41)⁄⁄⁄
0.004 (
3.88)⁄⁄⁄
0.004 (
0.33) 3.30 57,222

161

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

Table 5 Regressions of ARET on RESP and control variables (based on the 2-dimension clustering procedures). This table shows the results of estimating regression (4) with ARET as the dependent variable. The sample period spans January 1993–December 2010, and the sample consists of all NYSE, AMEX and NASDAQ firms with trade and quote data from the TAQ database. D is a dummy equal to 1 if the early announcer’s news and the late announcer’s response are consistent. All remaining variables are defined in Table 1. Brackets contain the t-statistic of standard errors derived using the two-way clustering approach developed by Petersen (2009) and Thompson (2011). ⁄⁄⁄, ⁄⁄, and ⁄ denote 1%, 5%, and 10% significant levels, respectively, based on the 2-dimension clustering approach. Results related to the main variable of interest are highlighted in bold print.

INTERCEPT RESP D D*RESP ERLYPRARET ARETt
1 ARETt
14 SIZE LOGBM RET6 ACC Adjusted R2 (%) n

INTERCEPT RESP D D*RESP ERLYPRARET ARETt
1 ARETt
14 SIZE LOGBM RET6 ACC Adjusted R2 (%) n

I TZ replication sample (1973–2005)

II Our TAQ sample (1993–2010)

III TAQ subsample (1993–2001)

IV TAQ subsample (2002–2010)

0.001 (1.72)⁄
0.031 (
2.12)⁄⁄ 0.000 (
0.83)
0.004 (
0.18) 0.020 (4.40)⁄⁄⁄ 0.018 (5.01)⁄⁄⁄ 0.009 (2.87)⁄⁄⁄ 0.000 (3.51)⁄⁄⁄ 0.001 (3.92)⁄⁄⁄ 0.000 (2.47)⁄⁄
0.001 (
1.53) 0.14 155,927


0.006 (
1.49)
0.047 (
1.32) 0.002 (1.43) 0.067 (1.09) 0.067 (3.49)⁄⁄⁄ 0.013 (2.17)⁄⁄ 0.025 (3.86)⁄⁄⁄
0.001 (
1.05)
0.009 (
10.69)⁄⁄⁄
0.001 (
2.86)⁄⁄⁄
0.005 (
0.61) 0.91 179,734


0.004 (
0.96)
0.087 (
2.01)⁄⁄ 0.004 (1.91)⁄ 0.140 (1.47) 0.064 (3.13)⁄⁄⁄ 0.027 (2.87)⁄⁄⁄ 0.012 (1.22)
0.001 (
2.04)⁄⁄
0.010 (
9.41)⁄⁄⁄ 0.000 (
1.68)⁄
0.018 (
1.59) 1.22 65,291


0.007 (
1.13)
0.004 (
0.07) 0.000 (0.27)
0.009 (
0.13) 0.064 (2.29)⁄⁄ 0.003 (0.46) 0.032 (3.87)⁄⁄⁄ 0.000 (
0.13)
0.009 (
7.15)⁄⁄⁄
0.002 (
3.25)⁄⁄⁄ 0.008 (0.65) 0.86 114,443

V TAQ subsample (1993–2010) Below median SHRP (more efficient)

VI TAQ subsample (1993–2010) Above median SHRP (less efficient)

VII TAQ subsample (1993–2001) Below median SHRP (more efficient)

VIII TAQ subsample (1993–2001) Above median SHRP (less efficient)

IX TAQ subsample (2002–2010) Below median SHRP (more efficient)

X TAQ subsample (2002–2010) Above median SHRP (less efficient)


0.013 (
3.00)⁄⁄⁄
0.024 (
0.52) 0.003 (1.86)⁄ 0.140 (1.19) 0.061 (3.22)⁄⁄⁄
0.005 (
0.53) 0.028 (3.10)⁄⁄⁄ 0.000 (0.13)
0.011 (
9.94)⁄⁄⁄
0.001 (
3.29)⁄⁄⁄ 0.009 (0.80) 1.46 91,269


0.001 (
0.14)
0.072 (
1.83)⁄ 0.001 (0.46) 0.020 (0.36) 0.062 (2.44)⁄⁄ 0.032 (3.47)⁄⁄⁄ 0.021 (2.91)⁄⁄⁄
0.001 (
1.29)
0.007 (
7.15)⁄⁄⁄ 0.000 (
1.36)
0.016 (
1.82)⁄ 0.72 88,465


0.008(
1.50)
0.090 (
1.31) 0.006 (1.95)⁄ 0.165 (0.97) 0.080 (3.22)⁄⁄⁄ 0.023 (1.88)⁄ 0.014 (0.83)
0.002 (
1.98)⁄⁄
0.014 (
9.36)⁄⁄⁄
0.001 (
1.97)⁄⁄
0.013 (
0.73) 1.99 32,496

0.001 (0.20)
0.082 (
1.65)⁄ 0.002 (0.88) 0.092 (1.32) 0.047 (1.37) 0.035 (2.58)⁄⁄ 0.010 (1.12)
0.001 (
0.93)
0.005 (
4.11)⁄⁄⁄ 0.000 (
0.06)
0.023 (
1.84)⁄ 0.60 32,795


0.013 (
2.50)⁄⁄
0.009 (
0.14) 0.003 (2.24)⁄⁄ 0.059 (0.70) 0.035 (1.44)
0.015 (
1.42) 0.039 (3.42)⁄⁄⁄ 0.000 (0.66)
0.010 (
7.52)⁄⁄⁄
0.002 (
2.79)⁄⁄⁄ 0.017 (1.22) 1.12 58,490


0.001 (
0.10)
0.006 (
0.08)
0.003 (
1.42)
0.040 (
0.40) 0.085 (1.97)⁄⁄ 0.020 (1.88)⁄ 0.027 (2.36)⁄⁄
0.001 (
0.73)
0.008 (
5.02)⁄⁄⁄
0.002 (
3.03)⁄⁄⁄ 0.001 (0.09) 0.87 55,953

momentum and Sloan (1996) accrual anomalies, we include the late announcing firm’s return over the prior six months (RET6) and the level of accruals (ACC), respectively. Table 4 reports the parameter estimates and the corresponding significance levels from the following regression:

ARET ¼ a0 þ b1 RESP þ b2 D þ b3 D  RESP þ b4 ERLYPRARET þ b5 ARET t
1 þb6 ARET t
4 þ b7 SIZE þ b8 LOGBM þ b9 RET6 þ b10 ACC þ e

ð4Þ The regression output reported in the first column of our Table 4 is obtained from TZ (Table 3, page 920), as it captures their main result. TZ show that the coefficient on RESP and the associated t-statistics are not affected when the control variables are introduced and that the negative relation between RESP and ARET appears unrelated to the different pricing anomalies controlled for in the regression. To provide evidence on the validity of the overreaction to intra-industry information transfers, our coefficient of interest is RESP, whose t-statistics in Table 4 are first based on Fama and MacBeth (1973), for consistency and comparability with TZ.13 Model I reports regression results for our replication of the TZ sample; we report and confirm the general finding by TZ that stock prices of late announcers in response to earnings reported by early announcers in the same industry are significantly negatively related to the subsequent price responses of late announcers to 13 Comparable significance is documented when we implement the two-way clustering approach to control for residual autocorrelation within firms over time and across firms in each time period (detailed results discussed in the next section).

their own earnings reports (RESP coefficient of
0.039 significant at the 1% level).14 Models II through X provide an important extension to TZ’s results using our TAQ sample. First, Model II indicates that the overreaction anomaly does not seem to be present during the period 1993–2010 for which we are able to estimate the SHRP. The coefficient on RESP is
0.043 and is insignificant at conventional levels. Second, Models III and IV split the TAQ sample period into two sub-periods (before and after the move to decimal pricing, which improved the efficiency of the price formation process). The RESP coefficient is negative and significant during the 1993– 2001 period (consistent with TZ) and becomes insignificant during the post-decimalization period. This finding suggests that the

14 We note that some of the control variables in Model I have the opposite and/or significant signs compared to those reported by TZ. For example, we are unable to confirm the significantly greater late-announcer’s underreaction to earnings news by early announcers (insignificant ERLYPRARET coefficient) or the greater overreaction when the early- and late-announcers react together when the news is revealed by the early announcer (insignificant D⁄RESP coefficient); however, we find a significant presence of momentum and size effects by documenting the positive significant coefficients on ARETt
4 and SIZE, respectively. These differences are likely due to the different number of observations used by the models and our initial sample screening procedures. First, we assume that the original TZ analysis in Table 4 is based on a maximum of 174,928 observations for which the ACC is available during 1973–2005. Because TZ do not report the number of observations used in their multivariate regressions, it is difficult to compare the exact numbers used by the two models. Second, in addition to all variables analyzed by TZ, we require that our sample firms also be categorized by the Global Industry Classification (GICS) History in order to verify that the apparent mispricing of information transfers is not a phenomenon unique to information transfers derived from the SIC system, as used by TZ (the robustness of our results to industry classification is discussed in the next section). Our sample of 155,927 firm-quarter observations is therefore different from TZ.

162

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

Table 6 Regressions of ARET on RESP and tests of RESP coefficients between subsamples (Fama–MacBeth approach). This table shows the results of testing for difference in the estimated coefficients of RESP under Model I between the two time period subsamples from Models III and IV in Table 4, under Model II between the two time period subsamples for the less efficient cases from Models VIII and X in Table 4, under Model III between the more efficient and the less efficient subsamples from Models V and VI in Table 4, and under Model IV between the more efficient and the less efficient subsamples for the 1993–2001 time period from Models VII and VIII in Table 4. PERIOD is a binary variable coded with a value of one if an observation is from the 2002–2010 time period, and zero if the observation is from the 1993–2001 time period. EFF is a binary variable coded with a value of one if a sample observation is from the above medium SHRP subsample, and zero if the observation is from the below medium SHRP subsample. All other variables are defined in the same way as in Table 4. Brackets contain the test statistics derived using the modified Fama–MacBeth approach (Models I and II) or the original Fama–MacBeth procedures (Models III and IV). ⁄⁄⁄, ⁄⁄, ⁄ and + denote 1%, 5%, 10%, and 15% significant levels, respectively. Results related to the main variable of interest are highlighted in bold print.

INTERCEPT RESP Other explanatory variables Period PERIOD*RESP Other interaction variables of PERIOD with explanatory variables Test statistic in test for sum of coefficients (PERIOD + PERIOD*RESP) = 0 EFF EFF⁄RESP Other interaction variables of EFF with explanatory variables Test statistic in test for sum of coefficients (PERIOD + PERIOD*RESP) = 0

I 1993–2001 vs 2002– 2010 (full sample)

II 1993–2001 vs 2002–2010 (less efficient subsample)

III more efficient vs less efficient (full sample)

IV more efficient vs less efficient (1993–2001 subsample)

0.002 (3.85)⁄⁄⁄
0.094 (
18.81)⁄⁄⁄ Included
0.011 (
10.50)⁄⁄⁄ 0.074 (9.86)⁄⁄⁄ Included

0.009 (9.23)⁄⁄⁄
0.086 (
13.83)⁄⁄⁄ Included
0.016 (
10.83)⁄⁄⁄ 0.027 (2.75)⁄⁄⁄ Included


0.014 (
3.57)⁄⁄⁄ 0.017 (0.37) Included – – –


0.009 (
1.96)⁄
0.029 (
0.67) Included – – –

1.634⁄

1.277

–

–

– – –

– – –

⁄⁄

0.011 (2.26)
0.094 (
1.79)⁄ Included

⁄⁄

0.013 (2.22)
0.058 (
0.90) Included

–

–

3.596⁄⁄⁄

4.237⁄⁄⁄

overreaction anomaly is observed during the earlier years but not in the more recent years, rejecting our first hypothesis and providing evidence that as the efficiency of the underlying price formation processes has improved over time, the over-reaction anomaly to intra-industry information transfers has weakened considerably. Lastly, Models V and VI partition our whole sample into two subsamples of firms characterized by more (below the median SHRP value) and less (above the median SHRP value) efficient price formation processes, whereas Models VII through X provide a four-way partitioning based on the two time periods and two types of efficient price processing. The coefficients on RESP suggest that for firms with low market frictions (more efficient in incorporating news into prices), the stock price movements of late announcers in response to earnings reported by early announcers are not significantly related to the subsequent price responses of late announcers to their own earnings reports. For less-efficient firms, the RESP coefficient is more negative in the earlier period (Model VIII) than the more recent period (Model X), and the coefficient is only significant in the early period. Overall, the overreaction anomaly appears to be driven by the pre-decimalization period and by firms with inefficient processing of information, which is consistent with our predictions. 4.2. Correction for cross-sectional and time-series dependence Recent developments in the literature have suggested that the Fama–MacBeth standard errors do not adequately correct for the effects of both cross-sectional and time-series dependence in regression analysis (Gow et al., 2010). We implement the twoway clustering approach proposed by Petersen (2009) and Thompson (2011) to control for the residual autocorrelation within firms over time (the firm effect) and residual correlation across firms in each time period (the time effect). Below each coefficient presented in the regression analysis in Table 5, we show the t-statistic of the double-clustered standard error. The results reported in Table 5 further support our predictions that the overreaction anomaly significantly attenuated over time as the market became more efficient in processing information in earnings news. The implementation of the two-way clustering

approach causes the coefficient of interest to lose some of its statistical significance; however, it remains significant at conventional levels. We observe similar signs of the coefficients in Table 5 as in Table 4; however, with the exception of Models III, VI, and VIII (corresponding to earlier period and/or to inefficient firms), none of the coefficients is statistically significant at conventional levels. In order to provide additional evidence on the magnitude of changes in the overreaction anomaly over time, we proceed to test for the difference in the RESP coefficient between various sub-samples. 4.3. Tests of RESP coefficients between subsamples To test the difference in the estimated coefficients of RESP between the two time period subsamples of Models III and IV in Tables 4 and 5, we combine the two models into a single model and apply the interaction variable approach allowing for both intercept shift and slope shifts for all independent variables in a fully unrestricted model setting (Jobson, 1991, page 314–320). We use a binary variable, PERIOD, and code it with a value of one if an observation is from the 2002–2010 time period, and zero if the observation is from the 1993–2001 time period. We add to the regression the main interaction variable, PERIOD*RESP, along with all of the other interaction variables created by multiplying PERIOD with each of the other explanatory variables in the model.15 In this model, the estimated coefficient of PERIOD*RESP provides the basis for us to test for significant difference in the overreaction effect between the 1993–2001 and the 2002–2010 subsample time periods, respectively. Similarly, we apply the interaction variable approach to test the difference in the estimated coefficients 15 An alternative is to specify a restricted model using PERIOD*RESP as the only PERIOD-based interaction variable in the model. By not allowing the interaction of PERIOD with each of the other explanatory variables, this restricted model forces the coefficient of each of these explanatory variables to be the same across the two time periods. We decide against imposing such a restriction (especially given significant differences in regression variables documented in Table 1, Panel C) and use the more general unrestricted model to allow for possible slope shifts for these variables even though inclusion of the additional interaction variables will likely have the effect of reducing the statistical significance of our main interaction variable PERIOD*RESP.

163

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

Table 7 Regressions of ARET on RESP and tests of RESP coefficients between subsamples (2-dimension clustering procedures). This table shows the results of testing for difference in the estimated coefficients of RESP under Model I between the two time period subsamples from Models III and IV in Table 5, under Model II between the two time period subsamples for the less efficient cases from Models VIII and X in Table 5, under Model III between the more efficient and the less efficient subsamples from Models V and VI in Table 5, and under Model IV between the more efficient and the less efficient subsamples for the 1993–2001 time period from Models VII and VIII in Table 5. PERIOD is a binary variable coded with a value of one if an observation is from the 2002–2010 time period, and zero if the observation is from the 1993–2001 time period. EFF is a binary variable coded with a value of one if a sample observation is from the above medium SHRP subsample, and zero if the observation is from the below medium SHRP subsample. All other variables are defined in the same way as in Table 5. Brackets contain the t-statistic of standard errors derived using the two-way clustering approach developed by Petersen (2009) and Thompson (2011). ⁄⁄⁄, ⁄⁄, ⁄ and + denote 1%, 5%, 10%, and 15% significant levels, respectively. Results related to the main variable of interest are highlighted in bold print.

INTERCEPT RESP Other explanatory variables PERIOD PERIOD⁄RESP Other interaction variables of PERIOD with explanatory variables F-statistic in test for sum of coefficients (PERIOD + PERIOD*RESP) = 0 EFF EFF*RESP Other interaction variables of EFF with explanatory variables F-statistic in test for sum of coefficients (EFF + EFF*RESP) = 0 Model R2 (%) n

I 1993–2001 vs 2002– 2010 (full sample)

II 1993–2001 vs 2002–2010 (less efficient subsample)

III more efficient vs less efficient (full sample)

IV more efficient vs less efficient (1993–2001 subsample)


0.006 (
1.79)⁄
0.097 (
2.17)⁄⁄ Included
0.004 (
0.63) 0.104 (1.56)+ Included


0.004 (
0.97)
0.122 (
3.02)⁄⁄⁄ Included 0.003 (0.31) 0.100 (1.47)+ Included


0.015 (
3.90)⁄⁄⁄
0.013 (
0.26) Included – – –


0.011 (
2.03)⁄⁄
0.074 (
1.01) Included – – –

0.020

0.160

–

–

– – –

– – –

⁄⁄

0.011 (2.38)
0.080 (
1.61)+ Included

0.11 (1.55)
0.013 (
0.13) Included

–

–

6.620⁄⁄

2.410

1.07 179,734

0.95 89,876

1.06 179,734

1.36 65,291

of RESP for the less efficient (i.e., above median SHRP) subsample between the two time periods.16 We also use the interaction variable approach to test for the difference in the RESP coefficients between the more efficient (i.e., below medium SHRP) and the less efficient (i.e., above medium SHRP) subsamples from Models V and VI in Tables 4 and 5. We introduce a binary variable, EFF, and code it with a value of one if a sample observation is from the above medium SHRP subsample, and zero if the observation is from the below medium SHRP subsample. Similarly, we apply the interaction variable approach to test the difference in the estimated coefficients of RESP for the 1993–2001 time period between the more efficient and the less efficient subsamples.17 We implement the interaction variable models applying the framework of the Fama–MacBeth approach as well as the 2dimensional double-clustering procedures. We discuss these approaches in the next two sections and present our test results in Tables 6 and 7, respectively. Results of our tests on the difference in the RESP coefficient between 1993–2001 and 2002–2010 are presented as Model I; on the difference in RESP for the less efficient (i.e., above median SHRP) subsample between the two time periods are presented as Model II; on the difference in RESP between the more efficient (i.e., below medium SHRP) and the less efficient (i.e., above medium SHRP) are presented as Model III; and on the difference in RESP for the 1993–2001 time period between the more efficient and the less efficient subsamples are presented as Model IV.

4.3.1. Tests of RESP coefficients between subsamples: the Fama–MacBeth approach One complication created by the interaction variable approach is that the Fama–MacBeth estimates can no longer be 16 The interaction variable approach is not applied to test for difference in RESP for the more efficient (i.e., below median SHRP) subsample between the two time periods because RESP is not significant in both time period regressions of Models VII and IX in Tables 4 and 5. 17 The interaction variable approach is not applied to test for difference in RESP for the 2002–2010 time period between the more efficient and the less efficient subsamples because RESP is not significant in both the below and above SHRP regressions of Models IX and X in Tables 4 and 5.

obtained in the usual manner when we are testing for the difference in RESP between 1993–2001 and 2002–2010.18 To solve this problem, we derive an alternate approach by randomly pairing earnings announcement quarters from the 1993–2001 and 2002– 2010 subsample time periods so that the full interaction model can be estimated with all observations from each quarter-pair in a modified cross-sectional setting.19 We generate 1,000 such random quarter-pairs, average the estimated coefficients across the modified cross- sectional regressions, and obtain the observed test statistic for each variable, consistent with the spirit of the original Fama–MacBeth framework. We apply the bootstrap method and utilize the sample data as a ‘surrogate population’ to approximate the sampling distribution of the test statistic. For Models I and II in our Table 6, we obtain the bootstrap samples by drawing the quarter-pairs from the combined 1993–2001 and 2002–2010 samples to simulate the condition of the null hypothesis distribution. The relevant test statistics under the null hypothesis are derived and the process is repeated 10,000 times to generate a bootstrap distribution for each test statistic. For hypothesis testing, we obtain a bootstrap p-value as the proportion of the test statistics from the null hypothesis distribution that are more extreme than the observed test statistic (Davidson and MacKinnon, 2006; MacKinnon, 2009). Results in Table 6 provide further support for our previous Table 4 results confirming the significant difference in the coefficients of RESP between the 1993–2001 and the 18 The Fama–MacBeth procedure requires running cross-sectional regressions separately for each time period and averaging the estimated coefficients across the time periods. In our setting, the interaction variables will all have the same value of either one or zero for each earnings announcement quarter, and estimation of the cross-sectional regressions on a quarter-by-quarter basis will fail to generate valid estimated coefficients for the interaction variables. 19 Our modified Fama–MacBeth procedures allow the valid estimation of these interaction variables on a quarter-pair by quarter-pair basis. To illustrate, we randomly generate quarter-pairs with one quarter from the early period and the other quarter from the recent period. For example, with a quarter-pair 1993Q2 and 2008Q3, we code the PERIOD variable accordingly, estimate the interaction variable model using all quarterly observations from 1993Q2 and 2008Q3, and include all the interaction variables in the model. We repeat the process with other randomly generated quarter-pairs, collect the estimated coefficients from each regression, and apply the final averaging procedures across all the quarter-pairs to obtain our modified Fama–MacBeth t-statistics.

164

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

Table 8 Regressions of ARET on RESP and tests of RESP coefficients between subsamples for cases of low institutional blockholding (2-dimension clustering procedures). This table shows the results for the same models as in Table 7 except that cases of high institutional blockholding are excluded from the analyses. All variables are defined in the same way as in Table 7. Brackets contain the t-statistic of standard errors derived using the two-way clustering approach developed by Petersen (2009) and Thompson (2011). ⁄⁄⁄, ⁄⁄, and ⁄ denote 1%, 5%, and 10% significant levels, respectively. Results related to the main variable of interest are highlighted in bold print.

INTERCEPT RESP Other explanatory variables PERIOD PERIOD*RESP Other interaction variables of PERIOD with explanatory variables F-statistic in test for sum of coefficients (PERIOD + PERIOD*RESP) = 0 EFF EFF*RESP Other interaction variables of EFF with explanatory variables F-statistic in test for sum of coefficients (EFF + EFF*RESP) = 0 Model R2 (%) n

I 1993–2001 vs 2002– 2010 (full sample)

II 1993–2001 vs 2002–2010 (less efficient subsample)

III more efficient vs less efficient (full sample)

IV more efficient vs less efficient (1993–2001 subsample)


0.001 (
0.25)
0.097 (
2.02)⁄⁄ Included
0.001 (
1.13) 0.165 (2.13)⁄⁄ Included

0.001 (0.15)
0.128 (
2.81)⁄⁄⁄ Included
0.006 (
0.46) 0.154 (2.11)⁄⁄ Included


0.013 (
2.31)⁄⁄ 0.023 (0.35) Included – – –


0.009 (
1.45)
0.061 (
0.72) Included – – –

1.340

0.220

–

–

– – –

– – –

0.010 (1.46)
0.119 (
1.66)⁄ Included

0.017 (2.00)⁄⁄
0.041 (
0.36) Included

–

–

6.170⁄⁄

2.920⁄

1.51 83,498

1.32 48,373

1.47 83,498

1.97 40,050

2002–2010 time period subsamples (significant PERIOD*RESP coefficient in Model I), between the two period subsamples for the less efficient cases (significant PERIOD*RESP coefficient in Model II), and between the more efficient and the less efficient subsamples (significant EFF*RESP coefficient in Model III). 4.3.2. Tests of RESP coefficients between subsamples: the double-clustered approach Table 7 shows only weak evidence of difference in the coefficients of RESP across the subsamples. The Petersen (2009) double-clustered t-statistics for PERIOD*RESP in Models I and II, and for EFF*RESP in Models III, and IV are 1.56, 1.47,
1.61, and
0.13 respectively. The first three statistics are only marginally significant at the 0.15 level and the last one is not significant at any conventional level. Other results in Table 7 are consistent with and support the findings reported in Table 5. Specifically, the statistically significant coefficients of RESP in Models I and II reiterate the previous results for RESP in Table 5 from the 1993–2001 subsample (Model III in Table 5) and from the 1993–2001 less efficient subsample (Model VIII in Table 5), and the insignificant F-statistics for the sum of coefficients (PERIOD + PERIOD*RESP) reaffirm the previous Table 5 results for RESP from the 2002–2010 subsample (Model IV in Table 5) and from the 2002–2010 less efficient subsample (Model X in Table 5). Similarly, the statistically insignificant coefficients of RESP in Models III and IV of Table 7 reiterate the previous results for RESP in Table 5 from the more efficient full period subsample (Model V in Table 5) and from the more efficient 1993–2001 subsample (Model VII in Table 5), and the significant F-statistic for the sum of coefficients (EFF + EFF*RESP) in Model III reassert the previous Table 5 results from the less efficient full period subsample (Model VI in Table 5). However, the insignificant Fstatistic for the sum of coefficients (EFF + EFF*RESP) in Model IV of Table 7 is too weak to corroborate the previous Table 5 results for RESP from the less efficient 1993–2001 subsample (Model VIII in Table 5). 4.3.3. Additional evidence on the RESP coefficients between subsamples To provide additional insights into our findings and in light of the various weak results reported in Table 7, we consider the possibility that the coarse partition of our sample may have

contributed to the apparent lack of statistical power limiting our ability to decisively reject the null in our hypothesis testing. Our sample is partitioned on the basis of decimalization and SHRP, and it is possible that other factors such as increased hedge fund activity and trading by sophisticated investors also have important effects on the attenuation of the overreaction anomaly. For example, Chordia et al. (2014) study twelve wellknown equity market anomalies and report empirical evidence linking decimalization, improvements in trading technology, and increased arbitrage and hedge fund activity to the attenuation of the anomalies. We aim to identify a factor not already captured by SHRP or any of the other variables in our current regression models. Our expectation is that a finer partitioning of our sample based on this factor should improve the statistical significance of the weak Table 7 results.20 One factor we consider is institutional blockholding.21 Prior research shows that institutional investors are able to exploit their informational advantages and adjust their investment preferences in pursuit of greater arbitrage profits (Bennett et al., 2003). Institutional blockholders are also perceived by market makers to have superior information, and the evidence suggests that blockholders have the ability to access or develop private, value-relevant information (Heflin and Shaw, 2000; Rubin, 2007). Our conjecture is that companies with large institutional holdings already enjoy a more efficient market for their equity and thus would be less subject to an overreaction anomaly in the first place. Therefore, these companies would be less likely to show an attenuation of an already small or nonexistent overreaction anomaly. Excluding these companies from the sample focuses attention on those companies whose equity markets will benefit most from decimalization. For each stock in our sample, we collect data on institutional blockholders from the Thomson Reuters Institutional (13f) Holdings database, and partition all observations into the below and above median groups representing cases of low and high institutional blockholding respectively. We repeat the estimation of our Table 7 regressions

20 We thank an anonymous referee for suggesting this direction of additional analysis. 21 Another factor we consider is hedge fund activity. However, due to the lack of databases available for measuring the level of hedge funds activity in the trading of individual stocks, we decide to concentrate on using institutional blockholding as a proxy for measuring the effects of these and other sophisticated investors.

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

this time excluding the high institutional blockholding cases and present the results in Table 8. Results in Table 8 are remarkably stronger compared to the previous results presented in Table 7. By removing cases that are more susceptible to the attenuation of the overreaction anomaly, we are able to refine our sample partitions and accentuate the detectability of the overreaction effect. Notable improvements include the double-clustered t-statistics for PERIOD*RESP in Models I and II which are now 2.13 and 2.11, both significant at the 0.05 level, confirming the difference in the coefficients of RESP across the 1993–2001 and the 2002–2010 subsamples (Models III vs IV and Models VIII vs X in Table 5). The double-clustered t-statistics for EFF*RESP in Model III is now 1.66, significant at the 0.10 level, providing support for the difference in the coefficients of RESP between the more efficient and the less efficient subsamples (Models V vs VI in Table 5). The coefficient of EFF*RESP in Model IV remains insignificant confirming that the RESP coefficients for both the more efficient and the less efficient subsamples in the 1993– 2001 time period are close and not significantly different (Models VII vs VIII in Table 5). All the F-statistics in Table 8 on the sum of coefficients further confirm and strengthen the previous findings in Table 7. In particular, the F-statistic for the sum of coefficients (EFF + EFF*RESP) in Model IV is now 2.92, significant at the 0.10 level, providing support to corroborate the significant RESP coefficient in the less efficient 1993–2001 subsample (Model VIII in Table 5). Overall, the Table 8 results on the coefficients of RESP and PERIOD*RESP, and the sum of coefficients (PERIOD + PERIOD*RESP) confirm our main results in Table 5, and that especially for cases involving low institutional blockholding, the coefficient of RESP is more negative in the 1993–2001 period compared to the 2002– 2010 period, and also more negative for the less-efficient firms compared to the more-efficient firms. The evidence supports our overall finding that the overreaction anomaly was significant in the pre-decimalization period especially for the less-efficient firms, but has since significantly attenuated.

4.4. Tests of robustness We conduct several robustness tests to provide further evidence for the importance of our results. First, in a recent study, Hao et al. (2011) show that the returns of less-profitable firms are more sensitive to industrylevel news and that this sensitivity is stronger with positive news and weaker with negative news. We carry out additional analyses and introduce a binary variable, PROFIT, to control for the effects of relative profitability of the late announcing firms, and another binary variable, GNEWS, for the positive vs negative nature of the message released by the early announcing industry peers. We code PROFIT with a value of one if the net income of the announcing firm is positive, and zero otherwise, and GNEWS with a value of one for good news from the industry peers in the form of a positive ERLYPRARET, and zero otherwise. To control for the effects of relative profitability of the late announcing firms, we add PROFIT and an interaction variable PROFIT*RESP to our regression models in Table 5. To control for the effects of the positive vs negative nature of the message released by the early announcing industry peers, we add GNEWS and the interaction variable GNEWS*RESP to our regression models. The analyses confirm our major findings in Table 5. An important result is that, after the introduction of the PROFIT*RESP and GNEWS*RESP interaction variables, our main variable RESP is always negative and remains statistically significant in Models III, VI and VIII (corresponding to the earlier time periods and

165

the less-efficient firms) even with the Petersen (2009) doubleclustering adjustment.22 Second, we carry out an additional analysis and further confirm that the overreaction anomaly we identified for the earlier time periods and for the less-efficient firms is not driven entirely by the positive lead or lag cross-correlations of returns across firms. Lo and MacKinlay (1990) show that abnormal returns from a portfolio can be the consequence of positive cross-correlations in returns across securities even though the returns of each security are serially independent. In our setting, the concern is that the negative correlation between returns during the two windows for lateannouncing firms (i.e., ARET and RESP for firm i) observed in crosssectional regressions can be due entirely to the positive correlation between returns for early announcers during the first window (i.e., ERLYPRARET for early peers) and returns for late announcers during the second window (i.e., ARET for firm i). This is possible even in the case where there is absolutely no negative autocorrelation in individual security returns or no stock market overreaction. TZ carry out additional procedures in their study to rule out this possibility (TZ, footnote 5, page 911). We follow TZ and document the same negative correlation between returns for late announcers (i. e., firm i) during the two windows (i.e., ARET and RESP) in timeseries regressions.23 Third, recent studies demonstrate that the Global Industry Classification System (GICS), developed jointly by Standard & Poor’s and Morgan Stanley Capital International, is empirically superior to the SIC in creating more homogenous industry groupings that in turn produce better industry-related measures (e.g., Bhojraj et al., 2003; Hrazdil et al., 2013, 2014).24 We re-examine the association between ARET and RESP and test whether the apparent mispricing of information transfers is a phenomenon unique to information transfers derived from the SIC system. Specifically, we re-estimate the information transfers based on the complete GICS History during 1985–2005 and confirm that the information transfers estimated based on the GICS provide consistent evidence, as in Model I. Untabulated results also provide confirmatory evidence in support of our two hypotheses. Fourth, to address the issue of how our results may be affected by the exact date-split, we repeat our analyses on Models III, VI and VIII in Table 5 that involve the subsample period 1993–2001 and check for the sensitivity of our major results by ending the period two years earlier. We decide to remove the 2000–2001 period for two reasons. First, NYSE and NASDAQ firms adopted decimalization during different months within this period (the decimal 22 The Petersen (2009) double-clustering adjusted t-statistics for RESP are
2.01,
2.84 and
1.73 in Models III, VI and VIII respectively, all significant at p < 0.10. Our main results in Table 5 remain unchanged. We also repeat the analyses with a full interaction model including all the interaction variables of PROFIT and GNEWS with each of the other explanatory variables in the regressions. However, with the Petersen (2009) double-clustering adjustment in this full interaction setting, our main variable RESP remains significant only in Model VI (corresponding to the less-efficient firms). The weaker results in general could be due to multicollinearity and the large number of explanatory variables included in many of these models. Our results show only weak support for the relative profitability effect and the nature of the news effect as reported by Hao et al. (2011). For example, the coefficient of PROFIT*RESP is positive and significant in Model VI when this interaction variable is introduced but is no longer significant after all the other interaction variables are added to form the full interaction model. Similarly, the coefficient of GNEWS*RESP is significant in Model III when this interaction variable is introduced but is no longer significant in the full interaction model. Further inclusion of the more complicated interaction variable PROFIT*GNEWS*RESP to test the sensitivity of PROFIT*RESP in relation to GNEWS also yields no significant results. 23 Specifically, we estimate the time-series regressions and find an overall negative RESP coefficient (with a t-statistic of
1.66, significant at p < 0.10). We implement the procedures in a way similar to the Fama–MacBeth approach except that the crosssectional and time-series operations are performed in the opposite order. 24 The GICS History provides firms’ industry classifications as of a given historical date starting in 1985 and is only available from Standard and Poor’s on a subscription basis.

166

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167

regime began on January 29, 2001, after a five-month test period for NYSE and on April 9, 2001, for NASDAQ firms). Second, the 2000–2001 period includes the peak and collapse of the hightech bubble, representing events that could have had a significant impact on trading and market efficiency. The sensitivity results confirm that our findings in Table 5 are robust and unaffected by the shortened 1993–1999 sample period.25 Lastly, for completeness, we control for information transfers around firm i’s earnings announcements by including ERLYPRRESP as an additional explanatory variable in Eq. (4). Repeating the regressions in Table 5 with ERLYPRRESP as an additional explanatory variable, we find this variable to be positive and highly significant (p-value <0.001) across all models, confirming that intraindustry information transfers are positive on average. However, this variable has very little effect on the magnitude and significance of all of the other variables, and the results reported in Table 5 are virtually unaffected by the addition of ERLYPRRESP in the multivariate setting.26 5. Conclusion Thomas and Zhang (2008) empirically demonstrate that the stock market overestimates the intra-industry implications of early announcers’ earnings for late announcers’ earnings and that the overestimation is corrected when late announcers disclose their earnings. In this paper, we revisit the analysis by TZ and show that the overreaction anomaly appears to have decayed in U.S. stock markets to the point that it is no longer statistically and economically significant. The key innovation of our paper is the use of short-horizon return predictability (a recently established market microstructure proxy for market efficiency, based on Chordia et al., 2008), to examine whether the anomaly is more pronounced among firms with a higher degree of market frictions, which hinders market makers, traders, and arbitrageurs in their processing of current earnings information. Our study contributes to the literature by demonstrating that the existence of the overreaction anomaly is specific to the earlier time period and results from the inefficient incorporation of information into prices, largely attributable to an environment with high barriers to arbitrage. We further document that the magnitude and significance of the anomaly is affected by the cross-sectional and time-series dependence of the residuals, where the implementation of the two-way clustering approach causes the coefficients of interest to decrease in significance. Overall, our results indicate that the pricing efficiency of intraindustry information transfers has increased in the regime of markedly higher trading activity. Given that short-horizon return predictability is a valid proxy for capturing an inefficient environment with high barriers to arbitrage, our findings show that the CRS measure of market efficiency better predicts and explains stock returns surrounding earnings announcements and their impact on other firms’ returns. Whether this and similar market microstructure measures of market efficiency can also help explain other financial anomalies (accrual, cash flow, or momentum), or whether portfolio managers can utilize these measures further in designing profitable trading strategies, remains a subject for future research. 25 Specifically, with the last two years, 2000–2001, removed from the sample period, the corresponding Table 5 Petersen (2009) double-clustering adjusted tstatistics for RESP are
1.95,
2.86 and
2.02 in Models III, VI and VIII respectively, all significant with p-value <0.05. Our overall results in Table 5 remain unchanged. 26 For example, with ERLYPRRESP as an additional explanatory variable, the corresponding Table 5 Petersen (2009) double-clustering adjusted t-statistics for RESP are
1.85,
2.19 and
1.65 in Models III, VI and VIII respectively, all significant with p-value <0.10.

Acknowledgements We are grateful for helpful comments and suggestions from two anonymous referees, M. Brennan, R. Lundholm, T. Scott, K. Lo, J. Begley, N. Massoud, and seminar participants at the University of British Columbia, the University of Alberta, the Canadian Academic Accounting Association conference, the International Conference on Finance & Banking, and the Global Finance Association conference. We also acknowledge editorial help from the English Language Editing service from Elsevier and the financial support of the Social Sciences and Humanities Research Council of Canada, the Canadian Institute of Chartered Accountants, and the Institute of Chartered Accountants of British Columbia. All of the data were obtained from the publicly available sources cited in the study. Any errors are ours.

References Aktas, N., De Bodt, E., Van Oppens, V., 2008. Legal insider trading and market efficiency. Journal of Banking & Finance 32, 1379–1392. Asquith, P., Oman, R., Safaya, C., 2010. Short sales and trade classification algorithms. Journal of Financial Markets 13, 157–173. Ayers, B., Freeman, R., 1997. Market assessment of industry and firm earnings information. Journal of Accounting & Economics 24, 205–218. Barberis, N., Shleifer, A., Vishny, R., 1998. A model of investor sentiment. Journal of Financial Economics 49, 307–343. Bennett, K., Sias, R., Starks, L., 2003. Greener pastures and the impact of dynamic institutional preferences. Review of Financial Studies 16, 1203–1238. Bhojraj, S., Lee, C., Oler, Z., 2003. What’s my line? A comparison of industry classification schemes for capital market research. Journal of Accounting Research 41, 745–769. Boehmer, E., Kelley, E., 2009. Institutional investors and the informational efficiency of prices. Review of Financial Studies 22, 3563–3594. Carrion, A., 2013. Very fast money: high-frequency trading on the NASDAQ. Journal of Financial Markets 16, 680–711. Caton, G., Goh, J., Kohers, N., 2003. Dividend omissions and intraindustry information transfers. Journal of Financial Research 26, 51–64. Chan, L., Lakonishok, J., Swaminathan, B., 2007. Industry classifications and return comovement. Financial Analyst Journal 63, 56–70. Chordia, T., Subrahmanyam, A., 2004. Order imbalance and individual stock returns: theory and evidence. Journal of Financial Economics 72, 485–518. Chordia, T., Roll, R., Subrahmanyam, A., 2005. Evidence on the speed of convergence to market efficiency. Journal of Financial Economics 76, 271–292. Chordia, T., Roll, R., Subrahmanyam, A., 2008. Liquidity and market efficiency? Journal of Financial Economics 87, 429–468. Chordia, T., Roll, R., Subrahmanyam, A., 2011. Recent trends in trading activity and market quality. Journal of Financial Economics 101, 243–263. Chordia, T., Subrahmanyam, A., Tong, Q., 2014. Have capital market anomalies attenuated in the recent era of high liquidity and trading activity? Journal of Accounting & Economics 58, 41–58. Chung, D., Hrazdil, K., 2010a. Liquidity and market efficiency: a large sample study. Journal of Banking & Finance 24, 2346–2357. Chung, D., Hrazdil, K., 2010b. Liquidity and market efficiency: analysis of Nasdaq firms. Global Finance Journal 21, 262–274. Chung, D., Hrazdil, K., 2011. Market efficiency and the post-earnings announcement drift. Contemporary Accounting Research 28, 926–956. Davidson, R., MacKinnon, J., 2006. The power of bootstrap and asymptotic tests. Journal of Econometrics 133, 421–441. Elgers, P., Porter, S., Xu, L., 2008. The timing of industry and firm earnings information in securities prices: a re-evaluation. Journal of Accounting & Economics 45, 78–93. Ellis, K., Michaely, R., O’Hara, M., 2000. The accuracy of trade classification rules: evidence from Nasdaq. Journal of Financial and Quantitative Analysis 35, 529– 551. Fama, E., MacBeth, J., 1973. Risk, return, and equilibrium: empirical tests. Journal of Political Economy 81, 607–636. Finucane, T., 2000. A direct test of methods for inferring trade direction from intraday data. Journal of Financial and Quantitative Analysis 35, 553–576. Firth, M., 1996. Dividend changes, abnormal returns, and intra-industry firm valuations. Journal of Financial and Quantitative Analysis 31, 189–211. Foster, G., 1981. Intra-industry information transfers associated with earnings releases. Journal of Accounting & Economics 3, 201–232. Freeman, R., Tse, S., 1992. An earnings prediction approach to examining intercompany information transfers. Journal of Accounting & Economics 15, 509– 523. Frost, C., 1995. Intraindustry information transfer: an analysis of research methods and additional evidence. Review of Quantitative Finance and Accounting 5, 111–126. Gow, I., Ormazabal, G., Taylor, D., 2010. Correcting for cross-sectional and timeseries dependence in accounting research. The Accounting Review 85, 483–512.

D.Y. Chung et al. / Journal of Banking & Finance 60 (2015) 153–167 Graham, R., King, R., 1996. Industry information transfers: the effect of information environment. Journal of Business Finance and Accounting 23, 1289–1306. Han, J., Wild, J., 1990. Unexpected earnings and intraindustry information transfers: further evidence. Journal of Accounting Research 28, 211–219. Han, J., Wild, J., 1997. Timeliness of reporting and earnings information transfers. Journal of Business Finance and Accounting 24, 527–540. Han, J., Wild, J., Ramesh, K., 1989. Managers’ earnings forecasts and intra-industry information transfers. Journal of Accounting & Economics 11, 3–33. Hao, S., Jin, Q., Zhang, G., 2011. Relative firm profitability and stock return sensitivity to industry-level news. The Accounting Review 86, 1321–1347. Heflin, F., Shaw, K., 2000. Blockholder ownership and market liquidity. Journal of Financial and Quantitative Analysis 35, 621–633. Hou, K., 2007. Industry information diffusion and the lead-lag effect in stock returns. Review of Financial Studies 20, 1113–1138. Hrazdil, K., Trottier, K., Zhang, R., 2013. A comparison of industry classification schemes: a large sample study. Economics Letters 118, 77–80. Hrazdil, K., Trottier, K., Zhang, R., 2014. An intra- and inter-industry evaluation of three classification scheme common in capital market research. Applied Economics 46, 2021–2033. Jobson, D., 1991. Applied Multivariate Data Analysis, Volume I: Regression and Experimental Design, Springer-Verlag, New York. Lang, M., Lundholm, R., 1996. The relation between security returns, firm earnings, and industry earnings. Contemporary Accounting Research 13, 607–629.

167

Lee, C., Ready, M., 1991. Inferring trade directions from intra-day data. Journal of Finance 46, 733–747. Lo, A., MacKinlay, A., 1990. When are Contrarian profits due to stock market overreaction? Review of Financial Studies 3, 175–205. MacKinnon, J., 2009. Bootstrap hypothesis testing. In: Belsley, D., Kontoghiorghes, E. (Eds.), Handbook of Computational Econometrics. John Wiley & Sons Ltd, Chichester UK. Petersen, M., 2009. Estimating standard errors in finance panel data sets: Comparing approaches. Review of Financial Studies 22, 435–480. Ramnath, S., 2002. Investor and analyst reactions to earnings announcements of related firms: an empirical analysis. Journal of Accounting Research 40, 1351– 1376. Rubin, A., 2007. Ownership level, ownership concentration and liquidity. Journal of Financial Markets 10, 219–248. Sloan, R., 1996. Do stock prices fully reflect information in accruals and cash flows about future earnings? The Accounting Review 71, 289–315. Thomas, J., Zhang, F., 2008. Overreaction to intra-industry information transfers? Journal of Accounting Research 46, 909–940. Thompson, S., 2011. Simple formulas for standard errors that cluster by both firm and time. Journal of Financial Economics 99, 1–10. Tversky, A., Kahneman, D., 1974. Judgment under uncertainty: heuristics and biases. Science 185, 1124–1130. Visaltanachoti, N., Yang, T., 2010. Speed of convergence to efficiency for NYSE-listed foreign stocks. Journal of Banking & Finance 34, 594–605.