本章导读 本章罗列股票市场中清楚拒绝 EMH 的诸多证据（异象）。§9.1 对弱式 EMH 的违背：9.1.1 日历效应（一月效应 Keim 1983、九月效应、周末效应、假日效应、月末效应、季节性 Heston & Sadka 2008，图 9.1–9.2）；9.1.2 动量 Jegadeesh & Titman (1993)（买赢家卖输家，2–12 月正自相关，图 9.3）；9.1.3 长期反转 De Bondt & Thaler (1985)（3–8 年负自相关、市场过度反应，图 9.4）；注 9.1 短期反应不足 + 长期过度反应。§9.2 对半强式 EMH 的违背：9.2.1 特征效应（规模效应、B/M 效应，Fama-French 1992）；9.2.2 股价对错的东西反应 Cutler et al. (1988)（新闻 vs 非新闻，图 9.5–9.6）、Huberman & Regev (2001)（EntreMed，图 9.7–9.8）；9.2.3 盈余公告溢价 Frazzini & Lamont (2007)、Barber et al. (2013)（图 9.9–9.10）；9.2.4 盈余公告后漂移 (PEAD) Bernard & Thomas (1989)（图 9.11）。§9.3 对 EMH 不利证据的若干顾虑：联合假设、经济显著性、统计显著性、异象是否仍存在；注 9.2。图 9.1–9.11 已转述。

各种日历效应 / Various calendar effects 一月效应 (January effect)：Keim (1983) 用 1963–1979 年 NYSE 与 AMEX 上市公司数据表明，小市值股票（最小十分位）在一月赚取最高的异常收益（脚注：每只股票的异常收益取自 CRSP，为该股收益减去按 Scholes-Williams beta 排序构造的控制组合的等权日收益；再以该规模十分位内所有股票异常收益的等权平均作为该十分位的异常收益），见图 9.1。超过 50% 的一月溢价发生在第一周、尤其一月第一个交易日。人们用组合再平衡（一月因投资年终奖金而需求更高）与税损出售（12 月需求下降以实现更少资本利得/更多资本损失用于抵税，使一月卖压更小）来合理化这一发现。九月效应：在美国，按月平均收益看九月是迄今最差的月份。周末效应：Lakonishok & Smidt (1988) 用道指 90 年日度数据发现股票周五收益好、周一收益差。假日效应：Lakonishok & Smidt (1988) 还发现股票在假日前收益好。月末效应：股票在月末前后往往收益好。January effect: Keim (1983) uses data on firms on NYSE and AMEX between 1963 and 1979 to show that small-size stocks (smallest decile) earn the highest abnormal return in January (footnote: the abnormal return of each stock comes from the CRSP database, the stock return minus the equal-weighted daily return of a control portfolio formed on ranking of Scholes-Williams betas; then the author computes the size-decile abnormal return as the equal-weighted average of the abnormal returns of all stocks in that size decile), see Figure 9.1. More than 50% of the January premium happens in the first week, especially on the first trading day of January. People rationalize this by portfolio rebalancing (higher demand in January because of investing year-end bonuses) and tax-loss selling (a drop in demand in December to realize less capital gains or more capital losses for tax purposes so the selling pressure is smaller in January). September effect: in the U.S., September is by far the worst month in terms of monthly average return. Weekend effect: Lakonishok & Smidt (1988) use 90 years of daily data on the Dow Jones Industrial Average to find that stocks have good returns on Friday and bad returns on Monday. Holiday effect: Lakonishok & Smidt (1988) also find stocks have good returns before holidays. Turn-of-the-month effect: stocks tend to have good returns around the turn of the month.

季节性效应：Heston 和 Sadka (2008) / seasonality effect Heston & Sadka (2008) 用 1965 年 1 月至 2002 年 12 月 NYSE 与 AMEX 上市股票数据（脚注：Heston & Sadka 2010 表明同样模式在全球成立）运行回归 \(r_{i,t}=\alpha_{k,t}+\gamma_{k,t}\,r_{i,t-k}+u_{i,t}\) (9.1)，其中 \(r_{i,t}\) 为第 \(t\) 月的月收益、\(r_{i,t-k}\) 为滞后 \(k\) 月的月收益。作者先对每个月做横截面回归，再报告各 \(k\) 下 \(\gamma_{k,t}\) 的时间序列平均（图 9.2）。可见每隔 12 个月就有一个尖峰——12 个月前的月收益总能很好地预测当前月收益，该模式甚至追溯到 20 年滞后仍成立。一个可能原因是公司有按年频率发生的某些经常性事件、显著驱动收益。这类效应是行为金融的早期焦点，因为它与弱式 EMH 不可调和。Heston & Sadka (2008) use data on NYSE and AMEX listed stocks from January 1965 to December 2002 (footnote: Heston & Sadka 2010 show the same patterns hold globally) to run the regression \(r_{i,t}=\alpha_{k,t}+\gamma_{k,t}\,r_{i,t-k}+u_{i,t}\) (9.1), where \(r_{i,t}\) is the monthly return in month \(t\) and \(r_{i,t-k}\) is the \(k\)-month lagged monthly return. The authors first run the regression in the cross-section for each month, then report the time-series average of \(\gamma_{k,t}\) for each \(k\) (Figure 9.2). We can see a spike every 12 months — the monthly return 12 months ago always well-predicts the monthly return now, and the pattern is true even going back to a 20-year lag. One possible reason is that firms have certain recurrent events that happen at a yearly frequency and drive the returns significantly. This type of effect was an early focus in behavioral finance because it's irreconcilable with the weak-form EMH.

图 9.1、9.2（已转述 / Figures 9.1, 9.2, paraphrased） 图 9.1（按规模与月份的异常收益）：横轴为市值十分位（最小到最大），纵轴百分比异常收益；"一月"那条线在最小市值十分位处异常收益极高（约 0.5），随市值增大迅速下降；"二月至十二月"诸线则平贴在 0 附近——一月溢价集中于小盘股。图 9.2（(9.1) 中 \(\gamma_{k,t}\) 的时间序列平均）：横轴为滞后阶数 \(k\)（0 到约 240 个月），纵轴回归系数估计；曲线在 \(k=12,24,36,\dots\) 等 12 的倍数处出现规律性正向尖峰——年度周期性极强。Figure 9.1 (abnormal returns by size and month): the horizontal axis is the market-value decile (smallest to largest), the vertical axis percentage abnormal return; the "January" line is extremely high at the smallest-cap decile (about 0.5) and falls sharply with size, while the "February through December" lines lie flat near 0 — the January premium concentrates in small-cap stocks. Figure 9.2 (time-series average of \(\gamma_{k,t}\) in (9.1)): the horizontal axis is the lag \(k\) (0 to about 240 months), the vertical axis the regression coefficient estimate; the curve shows regular positive spikes at multiples of 12 (\(k=12,24,36,\dots\)) — a very strong annual periodicity.

Jegadeesh 和 Titman (1993)：动量 / momentum 动量指个股异常收益在 2 个月到 12 个月的水平上很可能正自相关。Jegadeesh & Titman (1993) 用如下策略构造组合：计算每只股票过去 \(J\) 个月的收益，按收益升序排名；把最高十分位（最小收益十分位）称为"输家"、最低十分位（最大收益十分位）称为"赢家"。每个月买入赢家、卖空输家（等权），持有该组合 \(K\) 个月不做改变；于是每月，过去 \(K-1\) 个月构造的组合保持不变，而 \(K\) 个月前构造的组合被本月新构造的组合替换。动量策略几乎总能赚正收益；一周滞后的安排（图 9.3 面板 B）在某些情形下能进一步改善收益（图 9.3）。Momentum means the abnormal returns of an individual stock are very likely positively autocorrelated at a 2-months to 12-months horizon. Jegadeesh & Titman (1993) construct portfolios by the following strategy: calculate the return of every stock over the past \(J\) months, and rank all stocks in ascending order based on those returns; call the top decile (smallest-return decile) the "losers" and the bottom decile (largest-return decile) the "winners". Every month, buy the winners and short the losers with equal weight to form a portfolio, and hold that portfolio for \(K\) months without making any changes; so every month, the portfolios formed in the previous \(K-1\) months remain unchanged, but the portfolio constructed \(K\) months ago is replaced with a new portfolio constructed in the current month. The momentum strategy almost always earns positive returns; a one-week lag arrangement (Panel B of Figure 9.3) improves the returns in some cases even further (Figure 9.3).

图 9.3（\(J\)-\(K\) 动量策略的收益，已转述 / Figure 9.3, paraphrased） 表格按形成期 \(J\)（行）× 持有期 \(K\)（列）列出各动量组合的月均收益，面板 A 为"算完滞后收益后立即形成组合"、面板 B 为"算完滞后收益一周后再形成组合"。每格给出 Sell（空腿，仅卖空输家）、Buy（多腿，仅买入赢家）、Buy-sell（多空策略，买赢家卖输家）的收益。绝大多数 \(J\)-\(K\) 组合的 Buy-sell 收益为正且显著，面板 B（滞后一周）的收益往往更高。A table listing the average monthly return of each momentum portfolio by formation period \(J\) (rows) × holding period \(K\) (columns), with Panel A "forming the portfolio immediately after calculating the lagged returns" and Panel B "forming the portfolio one week after calculating the lagged returns". Each cell gives the return of Sell (short leg, shorting losers only), Buy (long leg, buying winners only), and Buy-sell (long-short strategy, buying winners and shorting losers). The vast majority of the \(J\)-\(K\) portfolios have positive and significant Buy-sell returns, and Panel B (one-week lag) often has higher returns.

De Bondt 和 Thaler (1985)：长期反转 / long-horizon reversals 长期反转指个股异常收益在 3–8 年水平上很可能负自相关。De Bondt & Thaler (1985) 比较极端赢家与输家（按过去 3 年衡量）的未来收益表现。用 NYSE 普通股 1926 年 1 月至 1982 年 12 月的 CRSP 月度收益数据；对每只股票 \(j\) 计算其累积超额收益 \(CU_j\)，定义为前 36 个月残差（脚注：残差按 CAPM 计算）的算术和；该过程在 1930 年 1 月至 1977 年 12 月间的 16 个非重叠三年窗口中各重复一次。在每个三年窗口中，把所有公司按 \(CU_j\) 从低到高排名，最高 35 只入赢家组合 \(W\)、最低 35 只入输家组合 \(L\)。然后对 16 个三年窗口各算 \(W\) 与 \(L\) 的累积平均残差收益 \(CAR\)（组合残差收益的算术和），平均 \(CAR\) 定义为 16 个窗口 \(CAR\) 的平均。作者发现：赢家的平均 \(CAR\) 为负、输家的平均 \(CAR\) 为正，这是市场过度反应的强证据（图 9.4）。Long-horizon reversals mean the abnormal returns of an individual stock are very likely negatively autocorrelated at a 3–8 years horizon. De Bondt & Thaler (1985) compare the future return performances of extreme winners and losers measured over the past 3 years. They use CRSP monthly return data on common stocks traded on NYSE from January 1926 to December 1982; for every stock \(j\) they compute the cumulative excess return \(CU_j\) as the arithmetic sum of the prior 36 months' residuals (footnote: residuals calculated as residuals in CAPM); this process is repeated 16 times in 16 non-overlapping three-year windows between January 1930 and December 1977. In each three-year window, all firms are ranked from low to high by \(CU_j\), with the top 35 stocks assigned to the winner portfolio \(W\) and the bottom 35 to the loser portfolio \(L\). Then for each of the 16 three-year windows, the cumulative average residual return \(CAR\) is computed for both \(W\) and \(L\) (the arithmetic sum of the portfolio's residual returns), and the average \(CAR\) is defined as the average of the 16 windows' \(CAR\). The authors find: the average \(CAR\) of winners is negative, while the average \(CAR\) of losers is positive, which is strong evidence of market overreaction (Figure 9.4).

注 9.1（短期反应不足 + 长期过度反应 / Remark 9.1） 结合动量与长期反转，大致可得结论：市场在短期反应不足、在长期过度反应。于是动量得以延续（价格未充分针对信息调整），但不幸的是动量使价格走得太远，故长期反转随后发生以向回调整。Combining momentum and long-horizon reversals, we can probably reach the conclusion that the market under-reacts in the short run and overreacts in the long run. So momentum keeps going on as the market price is not sufficiently adjusted for information, but unfortunately the momentum makes the price go too far, so the long-horizon reversal then takes place to adjust backwards.

图 9.4（16 个三年窗口的平均 \(CAR\)：长期反转，已转述 / Figure 9.4, paraphrased） 横轴为"组合形成后的月份数"（0 到约 35），纵轴平均 \(CAR\)；输家组合线持续上行至正值（约 +0.10 以上），赢家组合线下行至负值（约 −0.05）——形成期的输家在此后跑赢、赢家跑输，呈现反转。The horizontal axis is "months after portfolio formation" (0 to about 35), the vertical axis the average \(CAR\); the loser portfolio line rises continuously to positive values (above about +0.10), while the winner portfolio line declines to negative values (about −0.05) — past losers subsequently outperform and past winners underperform, displaying reversal.

规模效应与 B/M 效应 / size effect and B/M effect 特征效应指收益可由某些公司特征预测。规模效应 (size effect)：低市值公司有更高收益（脚注：规模由市值定义，即价格 × 流通股数）（脚注：由 Fama & French 1992 记录）。规模效应与一月效应密切相关：一月之外没有强规模效应，低市值股票之外没有强一月效应。B/M 效应（脚注：B/M 为账面市值比，是市值账面比的倒数）：高账面市值比公司有更高收益（脚注：也由 Fama & French 1992 记录）。高 B/M 股票称价值股、低 B/M 股票称成长股。价值溢价可能与长期反转有关——因为成为高 B/M 股的一种方式就是过去有一连串低收益；故价值溢价可能部分由长期反转解释。Characteristic effects mean returns are predictable by some firm characteristics. Size effect: low-market-capitalization firms have higher returns (footnote: size is defined by market capitalization, price times shares outstanding) (footnote: documented by Fama & French 1992). The size effect is closely related to the January effect: outside January there is no strong size effect, and outside low-market-capitalization stocks there is no strong January effect. B/M effect (footnote: B/M is the book-to-market ratio, the inverse of the market-to-book ratio): firms with a high book-to-market ratio have higher returns (footnote: also documented by Fama & French 1992). High-B/M stocks are called value stocks and low-B/M stocks are called growth stocks. The value premium is potentially related to long-horizon reversals — because one way to become a high-B/M stock is by having a sequence of low returns in the past; so the value premium could potentially be partly explained by long-horizon reversals.

Cutler et al. (1988) 与 Huberman & Regev (2001)：新闻 vs 非新闻 / news vs non-news Cutler et al. (1988) 考察 1941–1987 年的重大历史事件与标普指数变动，记录到股票对真正重要的事件（新闻，图 9.5）反应不大，却对不清不楚的原因（非新闻，图 9.6）反应。Huberman & Regev (2001) 通过对 EntreMed 股价的事件研究表明，市场反应受投资者注意力约束。EntreMed 是一家小型生物科技公司，拥有将一种潜在抗癌过程商业化的权利（图 9.7 为其 1997 年 10 月 1 日至 1998 年 12 月 30 日股价）。1997 年 11 月 28 日，因前一天《Nature》发表了介绍该治疗过程的文章（可视为基本面新闻），价格与成交量上升。1998 年 5 月 4 日（周一），因《纽约时报》5 月 3 日（周日）的头版文章，价格从上周五的 USD 12.063 暴涨到近 USD 52。1998 年 11 月 11 日，因当天上午《华尔街日报》报道实验室未能复现原结果，价格下跌。显然，价格对真新闻（1997/11/28 与 1998/11/11）的反应远弱于对非新闻（1998/5/4）的反应。此外，1998 年 5 月 4 日前后市场对非新闻的反应在整个生科行业都很剧烈，体现为显著更高的行业收益、更高成交量、以及当日日收益超过 5% 的公司比例更高（图 9.8）。有时人们会对信息感到困惑、在错误的股票上交易（见 §7.5.1 Rashes 2001 的讨论）。Cutler et al. (1988) investigate the major historical events and S&P Index changes between 1941 and 1987, and document that stocks do not react much to the really important events (the news, Figure 9.5) but react to unclear causes (the non-news, Figure 9.6). Huberman & Regev (2001) show market reactions are bounded by investors' attention through an event study on EntreMed's stock price. EntreMed is a small biotech company with the rights to commercialize a potentially cancer-curing process (Figure 9.7 is its stock price from October 1, 1997 to December 30, 1998). On November 28, 1997, the price and volume increased because of the previous day's publication of a Nature article that introduced the curing process (can be regarded as fundamental news). On May 4, 1998 (Monday), the price rocketed from USD 12.063 on the previous Friday to almost USD 52 because of a front-page article in the New York Times on May 3, 1998 (Sunday). On November 11, 1998, the price dropped following that morning's Wall Street Journal article saying laboratories had failed to reproduce the original results. Clearly, the price reactions to real news (Nov 28, 1997 and Nov 11, 1998) are much weaker than the reaction to non-news (May 4, 1998). Moreover, the market reactions to non-news around May 4, 1998 are severe in the entire biotech industry, reflected by significantly higher industry returns, higher volume, and a higher fraction of firms with daily returns exceeding 5% on May 4, 1998 (Figure 9.8). Sometimes people are confused about information and trade on the wrong stocks (see the discussion in §7.5.1 on Rashes 2001).

图 9.5–9.8（已转述 / Figures 9.5–9.8, paraphrased） 图 9.5（重大事件与标普指数对应变化）：列出珍珠港、肯尼迪遇刺等重大事件当日的标普百分比变化，多数变动较小——重大新闻并未引起大幅价格反应。图 9.6（标普最大变动日与《纽约时报》解释）：列出指数变动最大的若干日及报纸给出的解释，多数解释含糊（非新闻）。图 9.7（EntreMed 收盘价与成交量，1997/10–1998/12）：价格在 1997/11/28（Nature）、1998/5/4（NYT，暴涨至约 52）、1998/11/11（WSJ，下跌）三处出现跳动，5/4 的非新闻反应最猛烈。图 9.8（1998/5/4 前后生科指数表现）：行业收益与成交量在 5/4 附近显著抬升——非新闻引发全行业反应。Figure 9.5 (major events and corresponding percentage change in the S&P Index): lists the S&P percentage change on the day of major events such as Pearl Harbor and the Kennedy assassination, most changes being small — major news did not cause large price reactions. Figure 9.6 (largest changes in the S&P Index and New York Times explanations): lists the days of largest index changes and the newspaper's explanations, most of which are vague (non-news). Figure 9.7 (EntreMed closing price and volume, 1997/10–1998/12): the price jumps at 1997/11/28 (Nature), 1998/5/4 (NYT, rocketing to about 52), and 1998/11/11 (WSJ, dropping), with the 5/4 non-news reaction the most violent. Figure 9.8 (biotech index performance around 1998/5/4): industry returns and volume rise significantly around 5/4 — the non-news triggers an industry-wide reaction.

Frazzini & Lamont (2007) 与 Barber et al. (2013) / earnings announcement premium Frazzini & Lamont (2007) 用 1973–2004 年全部美股数据表明：在第 \(t\) 月做盈余公告的股票有更高的异常收益（脚注：异常收益定义为公告者平均月收益减去非公告者平均月收益），称盈余公告溢价；并在此后每隔三个月有更高的异常成交量（脚注：异常成交量先把公司本月成交量标准化为本月成交量与过去 12 个月平均月成交量之比，再算公告者平均标准化成交量减非公告者平均标准化成交量），因为盈余通常按季度公告（图 9.9）。Barber et al. (2013) 用 20 个国家 1991 年 1 月至 2010 年 12 月的数据，在日度层面证实 Frazzini & Lamont (2007) 的月度发现。特别地发现：(1) 公告前的异常收益高于公告后（图 9.10a）；(2) 公告日起的三日窗口内平均日成交量出现尖峰（图 9.10b）；(3) 公告日及次日的平均日特质波动率出现尖峰（图 9.10c）。这些结果提示若干解释：注意力吸引效应（成交量越高收益越高，支持此说），或理性风险解释（特质波动率越高收益越高）。Frazzini & Lamont (2007) use data on all U.S. stocks from 1973 to 2004 to show that stocks that make an earnings announcement in month \(t\) have a higher abnormal return (footnote: abnormal return defined as the average monthly return of announcers minus the average monthly return of non-announcers), called the earnings announcement premium, and a higher abnormal volume (footnote: to calculate abnormal volume, first scale the firm's monthly volume as the ratio of the firm's volume this month over the average monthly volume in the previous 12 months, then calculate the average scaled volume of announcers minus the average scaled volume of non-announcers) every three months later, because earnings are typically announced on a quarterly basis (Figure 9.9). Barber et al. (2013) use data from 20 countries from January 1991 to December 2010 to confirm, on a daily basis, the monthly findings of Frazzini & Lamont (2007). In particular they find: (1) abnormal returns pre-earnings are higher than post-earnings (Figure 9.10a); (2) average daily volume in a three-day window starting at the announcement day spikes (Figure 9.10b); (3) average daily idiosyncratic volatility during the announcement day and the day after spikes (Figure 9.10c). These results suggest several types of explanations: the attention-grabbing effect (returns higher when volume higher, which supports this) or a rational risk-based story (returns higher when idiosyncratic volatility higher).

图 9.9、9.10（已转述 / Figures 9.9, 9.10, paraphrased） 图 9.9（公告月之后的异常收益与成交量）：横轴为"相对公告月的月份数 \(t+k\)"，左轴月异常收益（柱）、右轴月异常成交量（线），二者均在 \(k=0,3,6,\dots\) 等每隔三个月处出现尖峰——季度公告节律。图 9.10（盈余公告前后各日）：三幅图分别为 (a) 异常收益（公告前为正、公告后回落）、(b) 相对非公告日的日成交量（公告日附近三日尖峰）、(c) 相对非公告日的日特质波动率（公告日及次日尖峰）。Figure 9.9 (abnormal return and volume subsequent to announcement month): the horizontal axis is "month relative to announcement \(t+k\)", the left axis monthly abnormal return (bars), the right axis monthly abnormal volume (line), both showing spikes every three months at \(k=0,3,6,\dots\) — the quarterly announcement rhythm. Figure 9.10 (days around earnings announcements): three panels for (a) abnormal returns (positive before the announcement, falling back after), (b) daily volume relative to non-announcement days (a three-day spike around the announcement), (c) daily idiosyncratic volatility relative to non-announcement days (a spike on the announcement day and the day after).

Bernard 和 Thomas (1989)：盈余公告后漂移 / PEAD Bernard & Thomas (1989) 用 1974–1981 年数据为每只股票计算累积异常收益 \(CAR\)，定义为按规模调整（脚注：异常收益定义为公司股票收益减去同一规模十分位内各公司的平均收益）的日收益的算术和。\(CAR\) 从盈余公告前 60 多天起算、到公告后 60 多天结束。公司按标准化未预期盈余（脚注：SUE 表示一阶自回归盈余预期模型的预测误差、按其估计期标准差标准化）(SUE) 分为十个十分位组，第 10 组为 SUE 排名最高者。盈余公告后的价格运动表现为：高盈余惊喜股票向上漂移、低盈余惊喜股票向下漂移——从图 9.11 可见，从最高 \(CAR\) 到最低 \(CAR\)，组合排名恰好从 10 到 1。注意图 9.11a 与 9.11b 完全是同一回事，唯一区别在于：9.11a 报告的 \(CAR\) 从公告前 60 多天起算、中间不重启累积，而 9.11b 在公告日重启 \(CAR\) 的累积。还要注意，公告前的价格运动并不能给出任何交易策略，因为按未来事件给组合排序无法告诉我们过去任何可行的策略。Bernard & Thomas (1989) use data from 1974 to 1981 to calculate the cumulative abnormal return \(CAR\) for each stock, defined as the arithmetic sum of size-adjusted (footnote: abnormal return defined as the firm's stock return minus the average return of firms in the same size decile) daily returns. The \(CAR\) starts at over 60 days before the earnings announcement and ends at over 60 days post the earnings announcement. Firms are assigned to ten decile groups based on standardized unexpected earnings (footnote: SUE represents the forecast error from a first-order autoregressive earnings-expectations model scaled by its estimation-period standard deviation) (SUE), with group 10 having the firms with the highest SUE rankings. The post-earnings price movement leads to an upward drift of high-earnings-surprise stocks and a downward drift of low-earnings-surprise stocks — as can be seen from Figure 9.11, the ranking of portfolios is exactly from 10 to 1 when going from the highest \(CAR\) to the lowest \(CAR\). Note that Figure 9.11a and 9.11b are exactly the same thing; the only difference is that 9.11a reports \(CAR\) starting from over 60 days prior to the announcement and doesn't restart the accumulation in between, whereas 9.11b restarts the accumulation of \(CAR\) on the announcement day. Also note that the price movement prior to the announcement doesn't give us any trading strategy at all, because sorting portfolios by future events cannot inform us of any feasible strategies in the past.

PEAD 与特征效应的关系 / relation between PEAD and characteristic effects PEAD 可部分由动量解释，但在控制动量后仍有一些 PEAD。PEAD 也可部分由 B/M 效应解释——特别地，高 B/M 股票的异常收益集中在盈余公告附近。PEAD can be partially explained by momentum, but after controlling for momentum there is still some PEAD. PEAD can also be partially explained by the B/M effect — in particular, the abnormal returns on high-B/M stocks are concentrated around the earnings announcement.

图 9.11（盈余公告后漂移，已转述 / Figure 9.11, paraphrased） 横轴为相对公告日的交易日数，纵轴 \(CAR\)；按 SUE 分的十组曲线自上而下从第 10 组排到第 1 组——高盈余惊喜组在公告后持续上漂、低惊喜组持续下漂。(a) 前后相连（CAR 从公告前 60 多天连续累积）；(b) 前后分离（公告日重启累积）；两图本质相同。The horizontal axis is trading days relative to the announcement, the vertical axis \(CAR\); the ten SUE-sorted curves run from group 10 at the top down to group 1 at the bottom — high-surprise groups keep drifting up after the announcement and low-surprise groups keep drifting down. (a) prior and post connected (CAR accumulated continuously from over 60 days before); (b) prior and post separated (accumulation restarts on the announcement day); the two are essentially the same.

四类顾虑及其对策 / Four concerns and their solutions 联合假设问题：对 EMH 的拒绝可能实为拒绝"EMH 与资产定价模型"的联合假设，并不必然削弱 EMH 本身的可信度。对策：呈现多个基准资产定价模型、表明证据对模型不敏感；或论证模型选择与所识别的实证模式无关。经济显著性：EMH 可能只在经济上不重要的方面被违背，使人们不顾不利证据仍信赖 EMH。对策：聚焦在经济上显著的异象。统计显著性：异象可能由数据挖掘得到；有人会批评其统计显著性源于过小的标准差、未正确考虑寻找异象的数据挖掘过程。对策：尝试用新数据看用旧数据识别的模式能否在新时段或其他国家很好复现。异象是否仍在？：异象一旦在学术期刊发表，可能就在数据中消失。对策：频繁用新数据复现异象。Joint hypothesis problem: the rejection of EMH might actually be rejecting the joint hypothesis of EMH and the asset pricing model, which does not necessarily undermine the credibility of EMH itself. Solutions: present many benchmark asset pricing models and show insensitivity of the evidence to models; or argue that model choices are irrelevant to the empirical patterns identified. Economic significance: it's possible that EMH is violated in an economically unimportant way such that people still have faith in EMH regardless of the unfavorable evidence. Solutions: focus on economically significant anomalies. Statistical significance: it's possible that the anomalies are found by data mining; people may critique that the statistical significance comes from a too-small standard deviation that doesn't correctly account for the data-mining process of finding the anomalies. Solutions: try to look at new data to see whether the patterns identified with past data replicate well in new periods or in other countries. Are effects still there?: after anomalies get published in academic journals, they might disappear in the data. Solutions: use new data to replicate the anomalies frequently.

注 9.2 / Remark 9.2 基于关于 EMH 违背的丰富证据，"行为因素存在、EMH 有时被违背"几乎无可争辩。但关键问题在于：行为因素与 EMH 违背是否对市场与经济至关重要。因此好的行为研究不应只满足于指出一个异象，还应聚焦于为什么这个异象重要。Based on the rich evidence on violations of EMH, it's almost indisputable that behavioral factors exist and EMH is sometimes violated. But the key question is whether the behavioral factors and EMH violations are central to the market and the economy. So good behavioral research should not be satisfied just with pointing out an anomaly, but should also focus on why this anomaly is crucial.