Abstract
Most factor-based equity portfolios maintain static exposures across the cycle: equal or fixed weights to value, momentum, quality, low-volatility, and so on, rebalanced periodically. Empirically, however, individual factors exhibit pronounced regime dependence. Value and momentum can spend years in compensating phases. Quality and low-volatility tend to lead in deteriorating risk environments and lag in expansions. A static blend averages those phases together without trying to exploit them.
This working paper proposes a framework for dynamic factor allocation based on machine-learning models trained on regime-indicating features. The framework takes a basket of factor sleeves and adjusts their weights through time, conditioning on macroeconomic, volatility, and cross-sectional dispersion inputs. The architecture, the feature set, and the cross-validation protocol are described in full, with particular attention to leakage controls and to the separation between in-sample fitting and out-of-sample evaluation.
The empirical section validates the framework out-of-sample across multiple equity universes and across regimes the model did not see in training. The two questions central to any active factor-timing claim are addressed directly: whether the timing signal survives realistic transaction costs, and whether the behavior holds up under cross-period testing rather than concentrating in a single regime.
The paper also addresses the trade-off that defines most ML-driven systematic work: more model complexity tends to buy more in-sample fit but typically less out-of-sample robustness. The architecture choices documented in the paper are conscious moves toward the robust end of that spectrum.