Daily Returns Predictability: A Replication, Reconciliation and Extension of the Pockets Debate.

Abstract

This thesis shows that procedure matters as much as signal. On the expanded 1926–2024 sample, the original two-sided implementation yields relatively few, long, and strong pockets with sizable in-pocket improvements – consistent with the original results from Farmer et al. (2023) – but these gains are inflated by future-information leakage in the identification stage. Switching to the one-sided pocket classification materially reduces in-pocket explanatory power, shortens pockets, and increases their count – exactly what we expect when classifications are made in real time. Economic performance likewise compresses: Clark and West statistics and portfolio alphas and Sharpe ratios move closer to the prevailing-mean benchmark once leakage is removed. Introducing fixed Bayesian shrinkage – by smoothing forecasts before pockets identification – partially restores performance. Shrinkage reduces forecast variance and improves the bias-variance trade-off, leading to cleaner (though typically weaker than two-sided) pockets and modestly higher statistical and economic metrics relative with respect to the raw one-sided approach. The overall picture is internally consistent: (i) two-sided pockets overstate predictability because they have the benefits of future information on their side; (ii) one-sided pockets give a conservative, real-time baseline; and (iii) disciplined shrinkage can recover some signal without re-introducing look-ahead by improving the bias-variance trade-off. Substantively, the following lessons emerge. First, daily excess returns exhibit time-varying predictability that is state-dependent and fragile to methodology – supporting cautious, regime-aware modelling. Second, strict, out-of-sample one-sided evaluation is non-negotiable for credible inference in high-noise financial data. Third, simple regularization (here, fixed Bayesian shrinkage) can deliver practical gains while respecting real-time constraints. Fourth, as noted also by Cheng et al. (2025), time series of predictive variables tend to be globally nonstationary. Fifth, as discussed in section 1.2, when predicting equity premia, one should not expect very high values of 𝑅2, to the 71 point that the results obtained in Table 1 for in-pocket periods are already quite high. For practitioners, the implication is to treat pockets as signals to monitor, not guarantees; for researchers, the agenda is to refine one-sided, low-leakage procedures and to study how pocket characteristics evolve across macro and market regimes, especially in the post-2016 environment. My analysis only focused on the U.S. market as to be coherent with the original material, but future studies could also investigate different markets to see whether the same conclusions hold and whether similar pockets of predictability can be discovered as periods where predictability seems to be increased. Taken together, the replication, correction, and shrinkage experiments provide a reconciled account of pockets of predictability and a transparent template for real-time evaluation with data through 2024 – a setting where methodological discipline is the difference between spurious alphas and robust, economically meaningful evidence.