Machine Learning Models for Short-Term Chinese Equity Trend Forecasting

Summary

The document reviews tactical and dynamic asset allocation, then focuses on predicting short-term movements in Chinese equity indices. It compares logistic regression, artificial neural networks, and support vector machines across different inputs and training windows. For the CSI 300, the reported strongest monthly direction result came from logistic regression trained on a 36-month window, with accuracy reaching 65%. The analysis also groups returns into four or six classes to distinguish large and small rises or falls; the four-class setup reports a 64% win rate for its best model and higher excess returns than the binary setup.

The authors apply the CSI 300 trend signal to the CSI 500 and CSI 1000, reporting excess returns for both and improved Sharpe and Calmar ratios in an equally weighted combination of the three indices. The six-class model performs poorly, which the document attributes to too few observations in each class. These are reported historical findings; the excerpt provides no detailed sample dates, transaction-cost treatment, or robustness analysis, so the figures do not establish out-of-sample performance or live-trading reliability.

Key ideas

• The study compares logistic regression, neural networks, and support vector machines for monthly equity trend prediction.
• A 36-month logistic regression training window is reported as the strongest setup for the CSI 300.
• Classifying returns into four directional and magnitude groups reportedly improves excess returns over binary classification.
• The six-class approach performs poorly, with limited observations per class cited as a reason.
• Applying trend signals to three Chinese indices reportedly improves combined Sharpe and Calmar ratios.

Tags

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