Lag-Resilient Meta-Ensembles for Next-Day Equity Returns: An ML Framework for Algorithmic Trading

Abstract

This study develops a lag-resilient meta-ensemble for next-day equity return forecasting in algorithmic trading. The framework is motivated by a practical constraint in live trading systems: forecasts must be accurate, but they must also react quickly enough to changing market conditions to remain actionable. To address this problem, the proposed architecture combines linear econometric models, tree-based machine-learning models, and a recurrent neural meta-learner within a two-layer forecasting design. The first layer included ARIMAX, gradient-boosted trees, random forests, and an LSTM base learner. The second layer, termed MetaNet, is a GRU-based meta-learner that receives base forecasts and rolling error features to adapt the model weights over time. The empirical setting uses a daily panel of U.S. large-cap equities from January 2015 through December 2024, together with macroeconomic variables, implied volatility indicators, and news-derived sentiment signals. The evaluation follows a rolling one-step-ahead protocol designed to reflect live trading conditions. Across the test period, the proposed system achieved the strongest overall performance, with a lower forecast error, improved directional accuracy, and a materially lower signal lag than the benchmark models. This paper also discusses interpretability, latency, execution, and risk management considerations relevant to institutional deployment. The resulting framework is intended as a practical forecasting architecture for systematic trading environments in which temporal responsiveness is as important as statistical fit.