Search

We use empirical Bayes (EB) to mine data on 140,000 long-short strategies constructed from accounting ratios, past returns, and ticker symbols. This "high-throughput asset pricing" produces out-of-sample performance comparable to strategies in top finance journals. But unlike the published strategies, the data-mined strategies are free of look-ahead bias. EB predicts that high returns are concentrated in accounting strategies, small stocks, and pre-2004 samples, consistent with limited attention theories. The intuition is seen in the cross-sectional distribution of t-stats, which is far from the null for equal-weighted accounting strategies. High-throughput methods provide a rigorous, unbiased method for documenting asset pricing facts.

Mining 29,000 accounting ratios for t-statistics over 2.0 leads to cross-sectional return predictability similar to the peer review process. For both methods, about 50% of predictability remains after the original sample periods. Predictors supported by peer-reviewed risk explanations or equilibrium models underperform other predictors post-sample, suggesting peer review systematically mislabels mispricing as risk, though only 20% of predictors are labelled as risk. Data mining generates other features of peer review including the rise in returns as original sample periods end and the speed of post-sample decay. It also uncovers themes like investment, issuance, and accruals -- decades before they are published.

Researchers are more likely to share notable findings. As a result, published findings tend to overstate the magnitude of real-world phenomena. This bias is a natural concern for asset pricing research, which has found hundreds of return predictors and little consensus on their origins. Empirical evidence on publication bias comes from large scale meta-studies. Meta-studies of cross-sectional return predictability have settled on four stylized facts that demonstrate publication bias is not a dominant factor: (1) almost all findings can be replicated, (2) predictability persists out-of-sample, (3) empirical $t$-statistics are much larger than 2.0, and (4) predictors are weakly correlated. Each of these facts has been demonstrated in at least three meta-studies. Empirical Bayes statistics turn these facts into publication bias corrections. Estimates from three meta-studies find that the average correction (shrinkage) accounts for only 10 to 15 percent of in-sample mean returns and that the risk of inference going in the wrong direction (the false discovery rate) is less than 10%. Meta-studies also find that $t$-statistic hurdles exceed 3.0 in multiple testing algorithms and that returns are 30 to 50 percent weaker in alternative portfolio tests. These facts are easily misinterpreted as evidence of publication bias effects. We clarify these misinterpretations and others, including the conflating of ``mostly false findings'' with ``many insignificant findings,'' ``data snooping'' with ``liquidity effects,'' and ``failed replications'' with ``insignificant ad-hoc trading strategies.'' Meta-studies outside of the cross-sectional literature are rare. The four facts from cross-sectional meta-studies provide a framework for future research. We illustrate with a preliminary re-examination of equity premium predictability.

We characterize the structure and origins of missingness for 159 cross-sectional return predictors and study missing value handling for portfolios constructed using machine learning. Simply imputing with cross-sectional means performs well compared to rigorous expectation-maximization methods. This stems from three facts about predictor data: (1) missingness occurs in large blocks organized by time, (2) cross-sectional correlations are small, and (3) missingness tends to occur in blocks organized by the underlying data source. As a result, observed data provide little information about missing data. Sophisticated imputations introduce estimation noise that can lead to underperformance if machine learning is not carefully applied.

I develop simple and intuitive bounds for the false discovery rate (FDR) in cross-sectional return predictability publications. The bounds can be calculated by plugging in summary statistics from previous papers and reliably bound the FDR in simulations that closely mimic cross-predictor correlations. Most bounds find that at least 75% of findings are true. The tightest bound finds that at least 91% of findings are true. Surprisingly, the estimates in Harvey, Liu, and Zhu (2016) imply a similar FDR. I explain how Harvey et al.'s conclusion that most findings are false stems from misinterpreting ``insignificant factor'' as ``false discovery.''

Many scholars have called for raising statistical hurdles to guard against false discoveries in academic publications. I show these calls may be difficult to justify empirically. Published data exhibit bias: results that fail to meet existing hurdles are often unobserved. These unobserved results must be extrapolated, which can lead to weak identification of revised hurdles. In contrast, statistics that can target only published findings (e.g. empirical Bayes shrinkage and the FDR) can be strongly identified, as data on published findings is plentiful. I demonstrate these results theoretically and in an empirical analysis of the cross-sectional return predictability literature.

We develop an estimator for publication bias and apply it to 156 hedge portfolios based on published cross-sectional return predictors. Publication bias adjusted returns are only 12% smaller than in-sample returns. The small bias comes from the dispersion of returns across predictors, which is too large to be accounted for by data-mined noise. Among predictors that can survive journal review, a low t-stat hurdle of 1.8 controls for multiple testing using statistics recommended by Harvey, Liu, and Zhu (2015). The estimated bias is too small to account for the deterioration in returns after publication, suggesting an important role for mispricing. [ABSTRACT FROM AUTHOR]

We estimate asset pricing models with multiple risks: long-run growth, long-run volatility, habit, and a residual. The Bayesian estimation accounts for the entire likelihood of consumption, dividends, and the pricedividend ratio. We find that the residual represents at least 80% of the variance of the price-dividend ratio. Moreover, the residual tracks most recognizable features of stock market history such as the 1990's boom and bust. Long run risks and habit contribute primarily in crises. The dominance of the residual comes from the low correlation between asset prices and consumption growth moments. We discuss theories which are consistent with our results. [ABSTRACT FROM AUTHOR]