Exploring Macro Factor-Driven Multi-Asset Tactical Allocation Strategies with Machine Learning

Abstract

This study constructs a machine learning-based tactical asset allocation (TAA) framework to address the limitations of traditional strategic asset allocation (SAA) in a volatile macro environment. The study obtains macroeconomic indicators from FRED and combines Exchange Traded Fund (ETF) historical data with lagged and rolling standardization methods to avoid forward-looking bias. In this paper, ridge regression, random forest, and XGBoost prediction models are compared, and a model selection mechanism based on prediction accuracy is proposed. The results show that ridge regression is the most robust in terms of forecasting accuracy and directional consistency. Based on the forecast signal, two types of portfolio strategies are designed: the Top-N equal-weight strategy and the mean-variance optimization strategy. The empirical results show that the mean-variance optimization strategy outperforms the 60/40 benchmark portfolio in terms of risk-return performance: its annualized return rate is 15.86%, slightly higher than 15.73% of the 60/40 benchmark portfolio; The Sharpe ratio reached 1.564, up about 16.9% from 1.338 of 60/40; Volatility also decreased from 11.75% to 10.14%, significantly improving risk control. These findings verify the potential of machine learning-driven asset rotation, provide practical reference for investors and asset managers, and lay a methodological foundation for future research introducing more factors, more complex models, and cross-market validation.