Information Aggregation and Data Snooping

This paper studies the interaction between information aggregation and data snooping in the context of predicting stock returns. Using simulations, we demonstrate that the aggregation of predictors by standard techniques amplifies the distortions of test sizes produced by data snooping. The proposed alternative aggregation technique, which is a modification of 3PRF/PLS, penalizes likely spurious predictors and thereby mitigates data snooping concerns. We illustrate our approach by applying various aggregation methods to three sets of return predictors suggested in the literature. We find that the forecasting ability of combined predictors in two cases cannot be fully explained by data snooping.