Comparing Pointwise GBDT Regression with Listwise Stock Ranking

Summary

This forum post describes a Chinese equity backtest comparison between standard regression models and a platform stock ranking model. The author trained XGBoost, LightGBM, and CatBoost regressors on five-day returns using rolling historical periods, then selected the four highest predicted stocks for five-day holdings. The same general template factors and parameters were used, and special treatment stocks were excluded. The author reports that the stock ranker’s rolling results were weaker than those from the three regressors.

The central question is why a listwise learning-to-rank approach did not outperform pointwise regression followed by sorting. The post provides a useful experimental setup and highlights that a ranking objective does not guarantee better portfolio returns. However, the charts or detailed result figures are absent from the text, and it does not analyze differences in labels, training objectives, model tuning, portfolio construction, or statistical significance. It raises the comparison without resolving which method is superior or why.

Key ideas

• The experiment compares three GBDT regression models with a platform listwise stock ranker.
• The strategy buys the four highest predicted stocks and holds them for five days.
• The author reports weaker rolling results for the ranker in this particular setup.
• A ranking objective alone does not guarantee higher returns from a top-ranked portfolio.
• The post poses the cause as an open question and supplies no detailed statistical analysis.

Tags

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