Using XGBoost to Forecast Factor Returns and Dynamically Reweight a Stock Portfolio

Summary

This report describes a Chinese equity strategy that predicts the future effectiveness of seven style factors and adjusts their portfolio weights over time. It uses historical information coefficients, macroeconomic data, and market variables as model features, with XGBoost forecasting each factor’s next-period information coefficient. Factors with stronger predicted effectiveness receive greater weight in a composite score, and the portfolio equal-weights the highest-scoring group of stocks.

The reported backtest says rolling model training tracked changing market styles more effectively than equal-weighted factors or a fixed model. It also reports that the dynamic strategy’s higher turnover can erode returns through transaction costs, and that limiting turnover improved results. The figures are historical backtest claims, with one reported win-rate value missing from the source text. The report warns that changing market structure, trading behavior, or wider adoption of similar strategies may weaken performance; the results do not establish future effectiveness.

Key ideas

• The model uses factor history and macroeconomic and market variables to predict future factor information coefficients.
• Predicted factor effectiveness determines weights in a composite stock-ranking score.
• The portfolio equal-weights stocks in the top-scoring group and updates its factor weights dynamically.
• Rolling model training is reported to outperform fixed and equal-weighted alternatives in the historical backtest.
• Higher turnover creates transaction-cost drag, while limiting turnover is reported to improve results.
• Backtest performance may not persist as markets and participant behavior change.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.