Deep Neural Networks Versus LightGBM for Ranking A-Shares

Summary

This practice report compares a DeepAlpha deep neural network with LightGBM in a daily A-share stock-selection strategy. Both models use the same 32 short-term price and volume factors and five-year training sample. The portfolio buys the ten highest-ranked stocks each day, caps each holding at 20%, and holds positions for five trading days. The neural network uses three hidden layers with ReLU activation.

The author reports substantially higher returns and lower maximum drawdown for the DNN in the stated backtest, while noting that win rate did not improve. The report also argues that DNN rankings may support diversification because top-ranked candidates were described as relatively similar, but this is the author’s interpretation. Results are historical and tied to the stated sample and test period; the document provides no independent replication or detailed discussion of costs, validation design, or robustness. It notes that concentrated positions could raise drawdown and that model complexity and computational demands make DNN harder to use.

Key ideas

• The comparison holds the stated factors, training sample, and portfolio settings constant while changing the model.
• The strategy ranks A-shares daily and holds selected stocks for a fixed period with a per-stock cap.
• The author reports better returns and lower drawdown for DNN, except for win rate.
• The report suggests DNN rankings may help balance diversification and returns, based on the author’s observations.
• Backtest results do not establish robustness beyond the reported setup, and DNN requires more computation and expertise.

Tags

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