Episode 316 - Andrew Chen: "Is everything I was taught about cross-sectional asset pricing wrong?!"

Episode Summary

Executive Summary: Andrew Chen argues that most cross-sectional asset pricing predictors are real out-of-sample patterns, but their apparent alpha shrinks sharply after publication and is largely eliminated by transaction costs and modern market efficiency. He emphasizes careful terminology, replication quality, and skepticism toward strong narratives or theoretical stories that outpace the data.

Main Topics: Predictors vs. factors and what counts as a real anomaly (Priority: 5/5): Chen distinguishes predictors from factors and explains the common but imprecise language in asset pricing. He defines predictability broadly as evidence that a variable forecasts returns, while noting that factor, anomaly, and predictor are often conflated. Replication and the open-source asset pricing project (Priority: 5/5): The discussion centers on Chen’s open-source replication effort covering hundreds of predictors from major meta-studies. The project aimed to verify published results, improve transparency, and counter exaggerated claims of a replication crisis. Out-of-sample decay, false discovery, and publication bias (Priority: 5/5): Chen argues that many predictors decay by about half after publication, but they do not vanish immediately, which suggests they are not mostly statistical figments. He sharply criticizes prior work for overstating failure rates by equating insignificant results with false discoveries. Transaction costs and investability (Priority: 5/5): A major theme is that gross predictive returns are materially reduced once trading frictions are included. Chen says transaction costs can wipe out much of the apparent alpha, especially in the modern era, making many anomalies effectively uninvestable in simple form. Technology, market efficiency, and the post-2005 kink (Priority: 4/5): Chen highlights a notable decline in anomaly performance after the mid-2000s and ties it to improved information technology and market efficiency. He treats this as evidence that historical premiums are time-varying and unlikely to persist unchanged. Theory, machine learning, and the limits of storytelling (Priority: 4/5): He finds little evidence that strong theoretical foundations improve out-of-sample performance. Chen argues that both fancy equilibrium models and machine-learning-driven data mining can uncover temporary patterns, but investors should trust numbers more than narratives. Implications for investors and the future of asset pricing research (Priority: 4/5): The conversation ends with a pragmatic takeaway: investors should be skeptical, act quickly, and not expect old factor papers or elegant stories to deliver durable excess returns. Research is evolving toward more rigorous methods, but claims still need verification.

Key Arguments: Most literature-backed predictors can be replicated; the main issue is not inability to reproduce code but decay in performance and misinterpretation of results. The open-source replication project found only 3 failures out of roughly 200 predictors that actually claimed predictability in the original papers. A lot of prior criticism overstated the replication crisis by calling insignificant results failures even when the direction of the effect was correct. Out-of-sample returns typically decline by about 50% after publication, but the decay is gradual rather than immediate, which argues against pure false discovery. False discovery rates in Chen’s analysis are below 10%, far smaller than some prior influential claims suggested. Transaction costs are large enough to eliminate much of the gross alpha; using basic but mitigated cost assumptions leaves many anomaly strategies effectively at zero net return in recent decades. The mid-2000s appear to mark a structural break in predictor performance, likely tied to internet/IT-driven efficiency gains and faster information diffusion. Strong theoretical justification does not reliably translate into better out-of-sample performance; in some cases, more sophisticated models do worse. Machine learning and traditional academic publishing are similar in that both can be forms of data mining; the edge is temporary unless acted on quickly. Investors should trust the numbers more than the text, and should be highly skeptical of strong claims unsupported by the empirical evidence.

Data Points: Estimated documented cross-sectional predictors: ~200 - Chen’s estimate of predictors clearly documented in the academic literature. Anomalies discussed in famous meta-study: 450 - A cited paper inflated 150 predictors into 450 anomalies by varying definitions. Underlying predictors in that paper: 150 - Chen argues the headline anomaly count was misleading. Variables in open-source replication set: 300 - Comprehensive list from prior meta-studies examined for predictability. Variables that actually showed predictability in original papers: ~200 - Of the 300 variables, only about 200 had original evidence of predictability. Replication failures in open-source project: 3 - Chen says only three results failed to replicate satisfactorily out of about 200 replications. Approximate replication failure rate: 1-2% - Derived from 3 failures out of about 200 predictors. Post-publication/out-of-sample decay: ~50% - Returns from 1990 onward were about half of pre-publication/in-sample levels. Early out-of-sample decay: ~25% - In the first few years after publication, decay was milder than the long-run decline. Estimated false discovery rate: <10% - Chen’s estimate for false discoveries in cross-sectional asset pricing research. Transaction cost impact on gross returns: ~25-30% - Estimated share of original sample trading strategy returns eaten by trading frictions. Mid-2000s structural break: Around 2005 - Chen observes a kink in factor performance around the time information technology accelerated. Time horizon for data-mining advantage: Up to decades earlier - He notes many later-published anomalies could have been found long before publication via systematic search.

Pivotal Quotes: "If you see some people claim some big thing, you probably want to verify that the numbers actually support that." — Andrew Chen: Advice to investors and researchers on preferring empirical validation over narrative claims. "The returns are very close to zero." — Andrew Chen: Summary of net expected returns after accounting for transaction costs and modern decay. "You have to trust the numbers more than the text." — Andrew Chen: Core takeaway on how to read academic asset-pricing papers and factor stories.

Implications: For investors, simple factor replication is unlikely to deliver durable alpha after costs. For the field, the emphasis should shift from stories to verified numbers, rapid testing, and skepticism toward claims of permanent premiums.