Using 1D CNN and ResNet to Rank A-Shares by Short-Term Returns
Summary
This research note explains how one-dimensional convolutional neural networks can extract local patterns from financial time series, and how ResNet skip connections help address performance degradation as networks deepen. Its proposed stock-ranking setup feeds engineered market features through fully connected layers and reshaped residual blocks, then predicts forward returns. The study uses seven basic price and trading variables to construct 98 factors and predicts five-day returns for China A-share constituents of the CSI 500. It describes missing-value filling, standardization, clipping, a training and prediction split, and rolling retraining. Every five days, the model ranks stocks and selects the top 20 for weighted positions.
The note also cites image-classification results to illustrate ResNet, but those are not evidence of trading performance. The supplied text ends before detailed financial backtest results, so it does not establish the strategy's returns or robustness. The described experiment omits stop-loss, take-profit, and trend filters, and the article flags that its platform instructions refer to an older version.
Key ideas
- One-dimensional convolutions can extract local patterns from sequential market features.
- ResNet skip connections are presented as a way to reduce degradation in deeper networks.
- The stock-ranking experiment uses 98 engineered factors to predict five-day returns for CSI 500 constituents.
- The described process retrains on rolling windows and selects the highest-ranked stocks every five days.
- The available text does not provide detailed trading results, so profitability and robustness remain unverified.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.