Corrigendum: Man versus Machine Learning

Abstract

We revisit the empirical results in Binsbergen, Han, and Lopez-Lira (2023) and correct for a data leakage problem in firms’ past earnings. While the results at the one-quarter and one-year forecasting horizon are not affected by this data leakage concern, the results for the two-quarter, three-quarter, and two-year horizons are. After correcting for this data leakage problem, we document that the paper’s main conclusions are qualitatively unaffected by these data leakage concerns. Some of the quantitative results are affected, and we document each of them in this corrigendum. In summary, we find:

1. Machine learning models maintain their superior accuracy over analysts’ forecasts across all horizons.
2. The conditional bias, defined as the difference between analysts’ earnings forecasts and machine-learning forecasts, is, on average, positive (upward-biased) and increases with the forecast horizon, following a similar pattern relative to the original results.
3. The short legs of well-known anomalies contain firms with excessively optimistic earnings forecasts with a similar pattern to the original results.
4. Managers of companies with the largest upward-biased earnings forecasts are more likely to issue stocks, as in the original results.
5. While conditional biases are associated with negative cross-sectional return predictability, the magnitude of these results is smaller than in the original results. The value-weighted long-short portfolio sorted on conditional bias earns significant alphas of -0.63% (t-stat = -2.39) per month, -0.77% (t-stat = -3.44), and -0.48% (t-stat = -1.85) relative to the CAPM, the Fama-French three-factor, and the Fama-French five-factor models, respectively. The conditional biases do not significantly predict cross-sectional stock returns in Fama-MacBeth cross-sectional regressions.