蜡烛图隐含了价格趋势信息吗？—— 基于中国股票市场的证据
基于注意力机制 CNN 的 A 股蜡烛图二分类预测，多空组合夏普比率达 3.30（周频）和 1.50（月频）。
Research on whether candlestick charts contain implicit price trend information in the Chinese stock market.
Problem
Can image-based deep learning extract predictive signals from candlestick charts that traditional quantitative features miss?
Method
Built an attention-based CNN model that visualizes A-share price, volume, and moving averages as candlestick images for binary classification prediction. Compared CNN prediction signals with classic factors and conducted empirical analysis using logistic regression.
Result
The Attention-CNN performed best at the 20-day window for both weekly and monthly frequencies, with long-short Sharpe ratios of 3.30 and 1.50 respectively, improving annualized return by 2% over baseline CNN. CNN prediction signals showed independent pricing factor characteristics, with stronger predictive power in small-cap and high-turnover stocks, validating the effectiveness of image deep learning for A-share trend prediction.
核心成果
- Long-short Sharpe 3.30 (weekly), 1.50 (monthly)
- Attention-CNN outperforms baseline by 2%
- Independent pricing factor characteristics
- Stronger in small-cap & high-turnover stocks