The devil in the details: a multiverse view of pockets of predictability

The growing complexity of forecasting models increases the number of decision nodes in the research process, raising the risk of overfitting to specific design choices. We illustrate this issue using the recent concept of “pockets of predictability,” which posits that return predictability is time-varying and that short windows of high predictability can be identified ex ante. In this study, we reassess the robustness and practical applicability of this approach. By analyzing 19,440 variations of the original methodology, we find that its effectiveness depends critically on various seemingly minor methodological decisions. Furthermore, return predictability has declined significantly in recent decades, and the potential economic gains are highly sensitive to trading costs. Overall, strategies based on pockets of predictability should be approached with caution.