Cross-Sectional vs. Time-Series Scaling for Stock Selection

Summary

This study compares time-series and cross-sectional standardization of price and volume factors in an A-share stock-selection setting. It uses a dataset spanning 2011–2022 with 98 daily factors and a five-day return label, and compares LightGBM and deep neural network models. The evaluation uses time-ordered folds, with three years for training and the next year for prediction, to better reflect sequential use of market data.

The authors report that cross-sectional scaling performs better than time-series scaling for this factor set. They also simulate a portfolio based on the DNN predictions from 2017 through 2021, selecting the 50 highest-ranked stocks each day, and report annualized return, cumulative return, and Sharpe ratio results. The text gives limited detail on the experiment’s full results, execution assumptions, and robustness checks, so the findings should be treated as specific to this dataset and backtest. It also argues that validation error alone is insufficient for judging overfitting and recommends checking prediction correlation as well.

Key ideas

• The study compares unscaled, time-series-scaled, and cross-sectionally scaled factors with LightGBM and DNN models.
• Its validation uses rolling time-ordered folds with three years of training followed by one year of prediction.
• The authors report better results from cross-sectional scaling for the DeepAlpha factors examined.
• They recommend considering validation-set correlation alongside error metrics when assessing model fit.
• The reported portfolio simulation selects the 50 stocks with the highest predicted returns each day.

Tags

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