Tuning a CNN for Chinese Equity Return Prediction

Summary

This study evaluates a convolutional neural network that predicts five-day returns across the Chinese stock market. Its baseline uses 2012–2017 data for training and tests on 2018 through October 2021. The authors vary batch size, optimizer, loss function, feature-map width, kernel size, dropout, and batch normalization, then compare results across repeated runs and rolling yearly tests. Reported measures include annualized return, drawdown, and Sharpe ratio.

The reported tests suggest that adding batch normalization and using a deeper network improved results in several periods, while smaller model settings sometimes performed better. The article also notes substantial variation between training runs and sensitivity to epoch count. Its conclusions are limited by the historical sample, the reported backtest setup, and the absence of detail here on transaction costs or other implementation effects. The authors describe the findings as evidence of possible stability across market regimes, but they do not establish that the model will generalize to future data. The article also says its platform instructions and resources refer to an older version.

Key ideas

• The model predicts five-day stock returns using a convolutional neural network trained on Chinese equity data.
• The tests compare architectural and training choices, including network depth, batch size, optimizer, loss function, dropout, and batch normalization.
• Reported results vary across repeated runs, and the article identifies epoch selection as an important influence.
• Rolling tests show mixed yearly outcomes, so aggregate performance does not imply consistent results in every period.
• The article cautions that lower model complexity may help reduce overfitting, but its backtests alone cannot establish future performance.

Tags

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