A LightGBM Rolling Model for Five-Day Stock Ranking

Summary

This post outlines an A-share stock-selection framework that uses eight cross-sectionally ranked features covering valuation, momentum and reversal, and liquidity. A LightGBM regression model predicts relative returns over a five-trading-day horizon; the strategy holds the ten highest-scored stocks at equal weights and rebalances every five trading days. It filters out risk-flagged, suspended, and recently listed stocks, and retrains the model every 90 trading days. The design also saves trained models for later loading and tracks training and trading activity.

The author reports that LightGBM ran faster than the prior XGBoost version in the described environment, with similar prediction quality, but supplies no validation details or strategy performance results. The stated fit metric is training-set R-squared only, so it cannot establish out-of-sample predictive value and may conceal overfitting. The post also identifies missing validation and hyperparameter search, liquidity safeguards, and exit rules; consequently the framework is a starting point for research rather than evidence of a robust deployable strategy.

Key ideas

• The model ranks stocks using valuation, momentum and reversal, and liquidity features.
• It predicts five-day relative returns and equally weights the ten highest-scoring stocks.
• The strategy rebalances every five trading days and retrains every 90 trading days.
• Training-set R-squared alone does not measure out-of-sample performance.
• The write-up flags missing liquidity filters, validation, and explicit exit rules.

Tags

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