Mixed-Frequency Deep Learning for Cross-Sectional Stock Prediction
A mixed-frequency model combining CNN-LSTM price-volume features with MLP-encoded fundamentals via self-attention, trained with a pairwise ranking objective on CSI 300 stocks.
- +25% peak validation RankIC (0.0268 → 0.0335) from adding fundamentals to a price-only baseline
- Portfolio backtests: 69.3% annualized (Sharpe 3.28) at N = 10, 144.7% (Sharpe 2.98) at the extreme N = 1; +43.1% excess vs CSI 300 (simulation)
面向横截面股票预测的混频深度学习模型
将 CNN-LSTM 量价特征与 MLP 编码的财务基本面通过自注意力融合， 并以成对排序损失训练，用于沪深 300 股票的横截面排序预测。
- 引入财务信息使验证集峰值 RankIC 提升 25%（0.0268 → 0.0335）
- 组合回测：N = 10 年化 69.3%（夏普 3.28），N = 1 年化 144.7%（夏普 2.98）；较沪深 300 超额 +43.1%（模拟）
