Using Gradient-Boosted Ranking Models to Mine A-Share Factor Returns
Summary
This report presents an A-share stock-selection framework that applies StockRanker, a gradient-boosted learning-to-rank model, to fundamental, financial, and trading factors. It tests 282 individual factors across short, medium, and long holding horizons, ranking stocks by predicted future returns and forming portfolios from the leading names. The report also explains factor IC as a measure of predictive consistency and describes portfolio exposure and transaction-cost analyses as ways to assess model behavior.
The reported experiments use historical training and test periods, with examples where market-cap ranking and price-to-book ranking performed well in different periods. The authors report stronger factor-return capture than conventional selection methods in their comparisons, while noting that results varied with market conditions and transaction costs. The study is limited: it tests factors individually, does not neutralize industry or risk exposures, and leaves factor correlation, multi-factor performance, and strategy capacity for future work. Its historical findings therefore do not establish out-of-sample or live profitability.
Key ideas
- The framework uses gradient-boosted learning to rank stocks from factor data and future-return labels.
- The study tests 282 factors individually over multiple holding horizons.
- Information coefficients are presented as a way to assess factor predictive consistency.
- Historical comparisons report period-dependent results for market-cap and price-to-book factors.
- The authors identify missing factor-correlation, neutralization, multi-factor, and capacity analyses as limitations.
- Short-horizon strategies are described as more sensitive to trading costs because they rebalance more often.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.