MALP-Meta: An Empirical Evaluation of Maximum Agreement Linear Predictors for Directional Equity Market Forecasting

Typical machine learning forecasting approaches, optimized via mean squared error, struggle in decision settings requiring directional consistency. The Maximum Agreement Linear Predictor (MALP) framework addresses this by maximizing Lin’s Concordance Correlation Coefficient (CCC), but its meta-learner value under real, weak-signal conditions is untested. We evaluate a two-stage MALP-Meta architecture (four base models, a CCC-optimized meta-learner) on real hourly S&P 500 (SPY) data (n = 5032; 2023–2026), with every result cross-validated across two computing environments. Only Ridge achieves significant directional accuracy (52.7%, p = 0.037); no MALP-Meta variant improves on this, and we diagnose a previously undocumented pathology in which CCC optimization collapses toward a constant under weak signal. Of three temporal robustness strategies, only stable feature selection matches Ridge’s significance (52.5%, p = 0.048). Under an externally-defined stress regime (VIX>25), one meta-learner configuration improves simultaneously on concordance and direction, an unconfirmed signal (p = 0.291). Sensitivity analysis shows the primary result is fragile to the evaluation split. These findings affirm rather than overturn the conventional agreement–error trade-off: agreement-based meta-learning’s real-world value is modest, conditional on genuine market stress, and detectable only through rigorous cross-validation and sensitivity analysis.