Establish Prospective, Capacity, and Live-Operation Validity of Nonlinear Equity ML Models

Establish whether nonlinear machine-learning models for equity return prediction and portfolio construction retain their reported value under prospective evaluation, realistic capital capacity, and live operation.

Background

The paper reviews evidence that trees, boosting methods, neural networks, and other nonlinear machine-learning models can improve historical out-of-sample prediction and portfolio performance in broad equity panels. However, the cited evidence is primarily retrospective and evaluated within specific historical universes, rebalancing rules, and institutional assumptions. The authors explicitly identify prospective validation, capacity, and live-operation performance as unresolved dimensions requiring further investigation before such models can support stronger claims about durable investment profitability.

References

Their historical universes, rebalancing rules and institutional assumptions define a particular opportunity set. Cross-sectional predictability may compensate investors for risk, reflect mispricing, or combine both, and a factor-model intercept depends on which risks the model recognizes. Moreover, portfolio sorts amplify small statistical differences and can concentrate exposure in illiquid stocks. In the profile of Table~\ref{tab:evidence_profile}, their support is strongest on time-valid historical prediction and disclosed portfolio simulations; prospective, capacity and live-operation dimensions remain open.

— Artificial Intelligence in Equity and Crypto Markets: Progress, Profitability Evidence, and the Limits of Automated Investing  (2609.04917 - Zhu et al., 4 Sep 2026) in Section 4.1, “From linear characteristics to nonlinear cross-sectional models”