4.7 Case studies

4.7.1Post-publication decay of return predictors (McLean and Pontiff, 2016)#

Context By 2010 academics had published scores of variables that predict the cross-section of stock returns. If the effects were data-mined, they should vanish out of sample. If they were real mispricing, they should shrink after publication as traders exploit them.

Evidence McLean and Pontiff (Journal of Finance 71(1), February 2016, pp. 5–32) study 97 published predictors. Their abstract reports that portfolio returns are 26% lower out-of-sample (after the original sample period ends but before publication) and 58% lower post-publication. They read the out-of-sample decline as an upper bound on data-mining effects, and estimate a 32% (58% − 26%) lower return from publication-informed trading. Post-publication declines are larger for predictors with higher in-sample returns, and predictor portfolios become more correlated with other published-predictor portfolios after publication.

Mechanism, with this module's tools The 26% decline is what 4.6 predicts: the published in-sample estimate is the best of many trials and is biased upward by selection. The further decline after publication is the arbitrage channel: once known, the predictor attracts capital, which moves prices until the remaining return roughly covers costs and risk. Larger in-sample returns decay more, which is consistent with both channels (bigger selection bias and more attractive to arbitrageurs).

Arithmetic for a trader Suppose a published predictor had an in-sample $t$-statistic of 3.0. If post-publication mean returns are 42% of in-sample (58% lower) and volatility is unchanged, a post-publication sample of the same length would have an expected $t$ of about $0.42\times3.0=1.26$. You would need roughly $(2/1.26)^2\approx2.5$ times as much data to confirm it at $t=2$.

4.7.2The 2009 momentum crash#

Context After the 2008 crisis, the losers over the trailing year were banks, autos and levered industrials; the winners were defensive stocks.

Timeline (Daniel and Moskowitz, NBER w20439, US data)

• Beginning of March 2009: firms in the loser decile were down on average 84% from their peaks, including Citigroup, Bank of America, Ford, GM and International Paper. The past-winner portfolio held defensive or counter-cyclical firms such as AutoZone.
• March 2009: WML −42.28% (market +8.97%). The two-year market return before formation was −44.90%.
• April 2009: WML −45.52% (market +10.20%).
• March–May 2009: market +26%, loser decile +163%, winner decile +8%.
• August 2009: WML −30.54%.
• The paper's figure of the subsequent period shows that from March 2009 to March 2013 the losers produced more than twice the profits of the winners.

Mechanism Proposition 4.7: after a two-year market fall of about 45%, the sort on past returns loaded the short side with high-beta stocks, so WML was net short the market just as it rallied. The magnitude, far beyond a linear beta, comes from the option-like payoff of distressed equity (M08). Daniel and Moskowitz show that such crashes are partly forecastable: they occur in "panic" states, after market declines and when volatility is high.

What a swing trader should take away

1. A breakout or Trend-Template system run in early 2009 would have been mostly in cash: few stocks met the template, and the market filter (Definition 4.8) was negative. That is a feature. The danger is the short side for those who trade it, and the late re-entry for long-only systems: the first leaders of a new bull market are often the previous losers, which do not pass a 52-week-high filter.
2. Scale exposure by forecast volatility, and reduce gross momentum exposure after large market declines with high volatility.
3. The equivalent Indian question can be studied with the IIMA WML series around March 2009 and March–April 2020. Its calendar 2025 figure (−5.53% to the December release, against a long-run 10.65% a year) is a reminder that momentum has losing years outside crashes too. Problem 4.15 asks you to examine this.

4.7.3Chart rules and data snooping: from Brock–Lakonishok–LeBaron (1992) to Sullivan–Timmermann–White (1999)#

Context Brock, Lakonishok and LeBaron (Journal of Finance 47(5), 1992, pp. 1731–1764) tested 26 variants of two rules, moving-average crossovers and trading-range breaks, on the Dow Jones index from 1897 to 1986. They found strong support: returns after buy signals were higher and less volatile than after sell signals, returns after sell signals were negative, and the pattern was inconsistent with random-walk, AR(1), GARCH-M and EGARCH nulls in bootstrap tests.

The question Those 26 rules were not drawn at random. They were the rules that a century of chartists had converged on, and the Dow was the most studied series in finance. The relevant family is every rule that could have been chosen. Sullivan, Timmermann and White (Journal of Finance 54(5), 1999, pp. 1647–1691) applied White's Reality Check (4.6.5) to an expanded universe of rules built around BLL's, using about 100 years of daily Dow data, to measure the data-snooping bias directly.

Mechanism This is Proposition 4.13 at scale. With thousands of rules, the best rule's Sharpe ratio under the null is several cross-trial standard deviations above zero, and the Reality Check's bootstrap maximum captures exactly that, with the correlation among rules accounted for. The paper also examines a sub-sample after BLL's data ended, which is the cleanest out-of-sample test available for a rule chosen in 1992. Read Sections III–IV of the paper for the in-sample and post-1986 results before forming a view; Problem 4.16 asks you to set out, in advance, what result would support each interpretation.