Online Ensemble Learning for Sector Rotation: A Gradient-Free Framework

The introduction of a gradient-free online ensemble learning algorithm marks a notable advancement in the field of machine learning, particularly for sector rotation strategies. This algorithm integrates forecasts from a diverse set of 16 models, including both linear benchmarks like OLS, PCR, and LASSO, as well as nonlinear learners such as Random Forests and Gradient-Boosted Trees. By adapting weights based on recent predictive performance, the ensemble consistently outperforms individual models and traditional offline ensembles. The empirical findings suggest that sector returns are more predictable and stable than those of individual assets, making this approach particularly suitable for cross-sectional forecasting. The theoretical framework provides a guarantee on forecast accuracy, bounding the online forecast regret in terms of realized out-of-sample R-squared. This development not only enhances predictive accuracy but also promises stronger risk-adjusted returns across various …