Applying Convolutional Neural Networks to Chinese Stock Selection

Summary

The document summarizes a study applying convolutional neural networks (CNNs) to multi-factor stock selection in the Chinese A-share universe. It proposes arranging factor observations as two-dimensional inputs, treating convolution as factor combination, using a single convolution layer, and omitting pooling to preserve factor meaning. It also warns that the order of factors in the input representation can affect what the model learns.

The study compares CNNs with a fully connected neural network and linear regression using annual training and test splits over a historical sample. It reports that CNN-derived factors performed better in single-factor and quintile tests, while portfolio results depended on the benchmark: neural models lagged linear regression against the CSI 300, whereas the CNN led the alternatives against the CSI 500. These are historical backtest results, not evidence of future performance. The summary gives limited detail on costs, implementation, robustness, or whether the results persist beyond the tested sample, and it identifies further CNN design choices as open research directions.

Key ideas

• Factor observations can be arranged as two-dimensional inputs for a CNN stock-ranking model.
• The study treats convolution as factor synthesis and omits pooling to retain factor interpretability.
• The ordering of factors in the input representation can influence model results.
• Historical comparisons with neural and linear models vary depending on the benchmark used.
• The reported backtests do not establish that the strategy will perform similarly in the future.

Tags

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