Style Rotation with Machine Learning

Research evidence excerpt

Style Rotation with Machine Learning

Luo's QES Research

Quantitative Research

ESG

Portfolio Strategy

Multi-Dimensional Alpha June 2, 2026

STYLE ROTATION WITH MACHINE LEARNING

Time Series Modeling, Deep Learning, Foundation Models, and Recommender Systems

• Style Rotation and Factor Timing (Revisited). This paper tackles one of active investing's oldest implementation

problems—style rotation and factor timing—by reframing it as a capital-ranking exercise over 228 investable factor

portfolios (built on US and Global large-cap equity universes, respectively) using the latest tabular machine learning

techniques. Forecasts are produced at the economically interpretable factor-portfolio level and then (optionally)

projected back into stocks.

• Broad Feature Suite: Macro and Time Series Analytics. The conditioning set is deliberately broad: cross-asset

market variables (regions, sectors, industries, currencies), point-in-time macro surprises and consensus revisions

built from 25 economic releases (payrolls, CPI, PMIs, GDP, housing, jobless claims, etc.), and a wide menu of

econometric time-series features — ARIMA / ARIMAX / SARIMAX, state-space (Kalman filter) latent expected

returns, full-network / shrinkage / low-rank VAR, regime-switching VAR with spell-duration conditioning, and

GAM-based nonlinear controls.

• ML Model Bench: From Boosting, Deep Learning, to Recommender Systems. We benchmark the modern ML

stack end-to-end: regularized logistic regression and XGBoost as supervised baselines; deep sequence learners on

daily factor paths (DLinear, iTransformer, LSTM/GRU, TCN, PatchTST, TFT, TimeMixer); pretrained foundation time-

series models (Chronos, TimesFM); and graph / recommender architectures (ridge and shrunk-correlation lagged