Predictability has a location: what Mosaics of Predictability actually tests

Return predictability can vary sharply across assets and market conditions. The April 2026 working paper Mosaics of Predictability makes that heterogeneity the research object. Cong, Feng, He and Wang ask where a common modeling approach finds stronger signal, rather than only whether a larger global model improves average accuracy. For us, the useful implication is that forecast evaluation should show where a gain comes from.

The paper studies monthly U.S. equity returns from 1973–2022, using 51 firm characteristics and eight market predictors. Its Panel Tree partitions observations into groups and fits group-specific models. For cross-sectional evaluation, the first 30 years form the training sample and the most recent 20 years form the out-of-sample period. The time-series regime analysis uses the entire 50-year sample. Those designs give different evidentiary status to the resulting maps.

Predictability is stronger in groups associated with large earnings surprises, high earnings-price ratios and low trading volume. The paper also reports weak or negative predictive R² in some groups, including combinations involving low earnings surprises, high spreads and strong momentum. At the regime level, high dividend yields and constrained liquidity coincide with stronger predictability. Its portfolio tests include transaction-cost robustness checks.

Read a map without turning it into a trading rule

A learned group describes a relationship among characteristics. Membership can change as a firm's characteristics change. “This stock belongs to a predictable group” therefore needs an observation date and the fitted partition that assigned it. Treating the label as a permanent property of a ticker would change the paper's object.

Likewise, a negative predictive R² means the tested model lost a forecast-error comparison against its benchmark on that sample. It does not establish that every possible model or information source fails on those assets. The result narrows what the tested pipeline successfully extracted. That narrower conclusion remains valuable: it identifies where allocating more trust to that pipeline would need additional evidence.

The regime finding needs similar precision. A full-sample partition can describe an economically interpretable historical pattern. To turn that description into a prospective agent input, a new experiment must construct the partition using only earlier observations and assign future observations without looking ahead. Otherwise, the apparent ability to identify favorable states benefits from the future it is supposed to forecast.

This distinction changes the deployment question. An agent cannot wait until an economic episode is complete to decide whether it was a high-predictability episode. It needs currently available features, a frozen assignment rule and an explicit response when observations sit near a boundary.

Our proposed test combines signal location with feasible access

We propose a two-dimensional report for future work. The first dimension records forecast improvement by prespecified asset-state group. The second records the feasible trade capacity and measured execution conditions of the same group. Plotting both reveals whether the strongest signal sits in a market region the intended agent can actually use.

For an illustrative comparison, one group might show a reliable forecasting gain while offering little executable depth. Another might have a smaller gain across a much broader liquid universe. Ranking the groups by forecast error alone answers the research question about signal strength. Ranking them by cost-aware opportunity answers a different question about deployment suitability.

Freeze grouping rules before the test period. Compare group-specific models with a pooled model using the same predictors and the same development budget. Report how many observations enter each group, its forecast-error difference and uncertainty that respects repeated observations across firms and months. Also record transitions between groups. A result driven by a few unusual dates deserves a different interpretation from a persistent comparison.

Then rerun the evaluation under the intended liquidity and position constraints. If excluding thinly traded assets removes the forecasting gain, the evidence identifies a restricted opportunity rather than a broadly deployable one. If a gain persists, it justifies the next execution test.

The working paper's contribution is a serious alternative to treating market predictability as one constant. Our proposed extension would make the location of signal and the accessibility of that signal visible together, with prospective group assignment supplying the essential next receipt.