Anomaly Replication Lab
  • The zoo
  • Screener
  • Multiple testing
  • Correlations
  • Composite
  • Methodology
  • Guides
  • Resources
Sign in
Anomaly Replication LabSign in
  • The zoo
  • Screener
  • Multiple testing
  • Correlations
  • Composite
  • Methodology
  • Guides
  • Resources

Educational replication, not advice. Long-short returns are decile spreads or published factor series from the Kenneth R. French Data Library and AQR, value-weighted, US equities, and gross of transaction and shorting costs — which for high-turnover signals like short-term reversal is the whole story. Reproducing a premium in-sample is not evidence it will persist. Past performance does not predict future results.

© 2026 Austin Levenberg. All rights reserved. This site’s code, design, written analysis, and computed results are not licensed for reuse or redistribution — terms of use · privacy.

Data: Kenneth R. French Data Library, AQR Capital Management public data sets.

Part of austinlev.dev — four research tools, and a note on where this is headed.

← the zoo
ReversalGone

Long-term reversal (2–5 years)

Stocks that lost over the prior 2–5 years beat prior long-term winners.

De Bondt & Thaler · Journal of Finance · published 1985 · original sample 1926–1982 (reported ≈ +8.0%/yr, t ≈ 2.5)

How our version differs from the paper. De Bondt & Thaler equal-weighted 3-year winner/loser portfolios; the value-weighted decile version is smaller and mostly a value bet. The decay figure compares our replication in-sample to our replication after publication, so it isn’t affected by this — but the level gap between “reported” and “our replication” partly is.
Our replication
+3.2%
1931–2026, per year
Full-sample t
2.0
Sharpe 0.24
Post-publication
+0.5%
since 1986, t = 0.1
Decay
95%
of the premium is gone

The whole history

Growth of $1 in the long-short strategy, gross of trading and shorting costs. The dashed line marks the year the paper was published — everything to its right is out-of-sample.

$1 in the long-short strategy, gross of costs. Grey = the sample the paper had; blue = everything since.

Before and after publication

The comparison the site is built around.

WindowReturn/yrVolSharpet-statMax DD
Paper's sample (1926–1982)+7.6%24.4%0.412.6-74%
After publication (1986–2026)-1.3%19.5%0.030.1-82%
Full sample (1931–2026)+3.2%22.2%0.242.0-82%

Against the factor models

Is there a risk-adjusted alpha left once you control for the standard factors? Alpha is annualized; t-stats use Newey–West standard errors.

ModelAlpha/yrAlpha tR²Biggest loadings
CAPM (market only)+3.1%1.10.00Mkt-RF +0.03
Fama–French 3-factor-1.8%−0.90.45HML +0.92, SMB +0.78, Mkt-RF +0.01
Fama–French 5-factor-1.5%−0.80.52SMB +0.69, HML +0.66, CMA +0.64
FF5 + momentum-0.6%−0.30.52SMB +0.69, CMA +0.67, HML +0.61

Rolling performance

Trailing 3-year Sharpe ratio. Long stretches below zero are common.

By decade

1930s+25.9%/yrSharpe 0.58
1940s+14.4%/yrSharpe 0.72
1950s-6.3%/yrSharpe -0.55
1960s+5.9%/yrSharpe 0.44
1970s+11.1%/yrSharpe 0.52
1980s-2.0%/yrSharpe -0.14
1990s+2.4%/yrSharpe 0.14
2000s+10.1%/yrSharpe 0.54
2010s-7.5%/yrSharpe -0.42
2020s+0.2%/yrSharpe 0.01

The risk explanation

Long-term losers are distressed, so this overlaps the value premium — much of it disappears once you control for book-to-market.

The behavioral explanation

The original 'overreaction' paper: investors push prices too far from fundamentals over multi-year horizons and they mean-revert.

Why it might have decayed

Largely subsumed by the value factor once that was formalised in 1992; as a standalone signal it has been weak for decades.

Long-short = top-decile minus bottom-decile of stocks univariately sorted on the characteristic. Value-weighted, US equities, monthly, gross of transaction and shorting costs. The original-paper figures are approximate and annualized for comparison.