... and the Cross-Section of Expected Returns
Read the paperopens doi.org in a new tab
What they found
The authors catalogued 316 factors that had been published as predictors of stock returns and asked a basic question: given that hundreds of researchers have tested thousands of candidates, how strong should the evidence be before we believe a new factor? Using multiple-testing corrections, they argue that the usual t-statistic threshold of 2.0 is far too lenient and that a new factor should clear roughly 3.0. By that standard most published factors are likely false discoveries, though the major ones (market, value, momentum) survive.
What you can use
- Most published anomalies are probably not real; the factor zoo is largely the product of data mining.
- Demand a much higher bar of evidence for any 'new' edge, including your own backtests: a t-stat of 2 is not enough.
- The classic factors clear the higher bar; obscure ones typically do not.
Caveats
The correction methods make assumptions about the number of unpublished tests, which is unknowable. Some argue the 3.0 threshold is too strict for economically motivated factors. A free version is on SSRN.
Tags: anomalies, multiple-testing, replication, statistics
Summaries are our own reading of the paper, not the authors' words. Educational only, not advice. Discuss it in Book Club.