Building and Rolling a Deep Neural Network for Chinese Stock Selection

Summary

The report describes a daily Chinese equity selection model built with a fully connected deep neural network. It derives 98 features from seven basic price and volume inputs using rolling averages, extrema, dispersion, ranks, weighted averages, correlations, and lagged observations. The model predicts five-day returns, standardizes and clips inputs, then ranks stocks for a portfolio. A rolling design trains on three years and tests on the following year; the baseline also compares changes to learning rate, optimizer, loss function, batch size, and data processing.

The report presents results from both a fixed train/test split and rolling backtests, including returns, drawdowns, volatility, and Sharpe ratios, and argues that results persist across several parameter variants. These are historical, platform-reported backtests rather than evidence of live performance. The document’s own caveats include substantial drawdowns, limited short-term sequence memory, reliance on basic price-volume features, and the risk that past results may not generalize. Its claims of robustness should therefore be assessed with attention to implementation details and backtest design.

Key ideas

• The model predicts five-day stock returns from engineered features based on price and volume data.
• A fully connected neural network combines 98 standardized inputs through three hidden layers.
• The study compares parameter variants and uses rolling three-year training and one-year testing.
• Reported historical returns come with sizable drawdowns and do not establish future or live performance.
• The authors identify longer-term sequence information and richer features as areas for further research.

Tags

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