Skip to content

Open Source Cross-Sectional Asset Pricing: summary and key takeaways

The authors wrote open code for 319 firm-level stock return predictors and compared reproduced t-stats with hand-collected figures from the original papers. Nearly all predictors that were clearly significant in the originals reproduce. The paper argues that claims of widespread replication failure come from loose definitions of what counts as an anomaly.

Chen, A.Y.; Zimmermann, T. · 2021 · Finance and Economics Discussion SeriesSummary updated 20 November 2026
Published2021Journal article
Asset focusEquitiesFactor library
TypeResearchQuant Crises & Methods
Reading time5 minthis summary

Open Source Cross-Sectional Asset Pricing at a glance

The question
Can the published cross-sectional stock return predictors be reproduced when the original methods are followed, and do earlier reports of mass replication failure hold up?
Dataset & period
U.S. firm-level stock data (CRSP, Compustat, IBES and others), monthly characteristics, 319 characteristics drawn from 153 papers. Each predictor is tested over the sample period of its original paper. A single overall sample range is Not stated in the paper's prose.
Method
The authors read each original paper and record the key table, the sign of predictability, the t-stat and the portfolio design. They then code the characteristic and form long-short portfolios, mostly with the original weighting, sort and rebalancing choices. Each characteristic is placed in one of four evidence groups: clear predictor, likely predictor, not-predictor or indirect signal. Success is judged against the original evidence, with a raw long-short t-stat above 1.96 as the test for clear predictors.

Key findings of Open Source Cross-Sectional Asset Pricing

  • The dataset has 319 characteristics from 153 papers. 161 are clear predictors and 44 are likely predictors.
  • Among the clear predictors, 98% of the reproduced long-short portfolios have t-stats above 1.96. The count is 158 of 161, and one of the three misses reaches 1.93.
  • Regressing reproduced t-stats on the original long-short t-stats gives a slope of 0.88 and an R2 of 82%.
  • Across 205 clear and likely predictors, only three fail to match the original evidence of significance: R&D ability and the two shareholder activism measures.
  • Decile mean returns rise with the predictor. The paper puts the probability that decile k beats the decile two steps below at roughly 96%.
  • The Hou, Xue and Zhang failure rate is traced to classification: 117 out of these 118 clear predictors achieve t-stats above 1.96, and the last one reaches 1.93.

Robustness, as reported in the paper

  • Mean returns are monotonic across deciles at the median and at the quartiles of the predictor distribution.
  • Using only clear predictors, the paper reproduces three McLean and Pontiff post-publication facts: decay grows with in-sample return and t-stat, and returns stay positive.
  • Value-weighting or screening out small stocks cuts mean returns by about 30% (about 20 bps per month). Typical returns fall from around 60 bps to about 40 bps per month.
  • Less frequent rebalancing lowers returns steadily as the interval moves from one month to twelve.

Caveats and limits

  • The paper does not test whether the predictors give implementable profits. It points to other work finding that bid-ask spreads remove most post-publication returns.
  • Reproductions are attempted reproductions. Standard choices such as a six-month accounting data lag, raw returns and monthly rebalancing differ from some original papers.
  • Some results depend on the authors' judgment when sorting characteristics into clear, likely, not-predictor and indirect-signal groups. The authors note that reasonable people can disagree on some of these calls.
  • Coding errors or small deviations from the originals likely remain among hundreds of characteristics (inferred from the authors' own statement of this risk for the three failures).

Items marked (inferred) are our reading of the design, not statements by the authors.

Why it matters

The paper gives practitioners a public, auditable set of predictor signals and long-short portfolios to test against. It also shows that the size of reported replication failure depends on how an anomaly is defined. Anyone using these signals still has to account for trading costs and liquidity.

Questions about this paper

How many predictors does the paper reproduce successfully?

Among the clear predictors, 158 of 161 have long-short t-stats above 1.96. Only three of the 205 clear and likely predictors fail to match the original evidence.

Why does this differ from Hou, Xue and Zhang's finding of mass replication failure?

The paper says much of that failure comes from counting studies that never showed long-short significance. Of their clear predictors, 117 out of these 118 pass the 1.96 cutoff in the reproductions.

Do liquidity screens change the returns?

Yes. Value-weighting or market-equity screens lower typical mean returns from around 60 bps to about 40 bps per month.

Where can I read the original paper?

The original is at its publisher: https://doi.org/10.17016/feds.2021.037. We link to the paper and do not host the PDF.

Follow new quant research

Download MacroFinance market data as CSV or Excel, or pull it by API, and test the ideas in these papers yourself. New summaries are added as we review them.

Create a free account
Summaries are written by MacroFinance from the published papers, in our own words, for information only. They are not investment advice and may omit detail. Papers are not hosted here: each page links to the original at its publisher.