Allocation-Focused Regimes and Applications to Dynamic Factor Investing
提出一种混合识别-预测框架，通过统计跳跃模型识别配置中心体制，用XGBoost预测，并基于组合表现优化超参数，应用于美国因子组合数据，显著提升夏普比率。
The authors define allocation-focused regimes through the relative performances of investment strategies rather than broad economic conditions. To identify and forecast allocation-focused regimes, the authors propose a hybrid identification-forecast framework that integrates regime identification using statistical jump models, regime forecasting using XGBoost classifiers, and performance-driven hyperparameter optimization. To address interpretability and integration challenges when combining disjoint models, the framework decouples the identification and forecasting components to minimize interference and to improve robustness. The two steps are then unified through end-to-end hyperparameter optimization based on portfolio performance. This hybrid approach allows investors to identify persistent patterns of strategy outperformance/underperformance in comparison, and learn ex post the market conditions that may drive these patterns. Using empirical data on US equity factor portfolios from 1960 to 2024, the authors show that incorporating regime-aware forecasts into factor allocation strategies significantly improves performance compared to passive factor investing. Specifically, the authors demonstrate two applications: active allocation relative to the equal-weighted benchmark and dynamic allocation between complementary pairs of factors: value/growth, momentum/reversal, and size. Both strategies consistently outperform their benchmarks in terms of Sharpe ratios and achieve positive information ratios.