Directional Predictability of the CSI 300 Index: A Walk-forward Evaluation of a BILSTM-attention Model

Abstract

This study looks at short-term predictability in China’s stock market, using the CSI 300 index as the main example. A Bidirectional Long Short-Term Memory (BiLSTM) model with an Attention mechanism is built that uses recent price- and volatility-based technical indicators and is tested with an expanding-window, walk-forward method for both direction classification and return-size regression. The model choice follows earlier work showing that deep sequence models and volatility-aware hybrids can capture nonlinear patterns and changing volatility, even though the economic value of daily predictions is usually small. In the results, the BiLSTM-Attention model gives a stable but modest edge in one-day-ahead direction prediction—AUC, PR-AUC, F1, and hit ratio are all above random—while predicted return sizes are still very noisy, which is consistent with weak-form market efficiency. The paper also discusses a tradability-oriented framework that links calibrated probabilities to position sizing under transaction costs and regime shifts, and outlines possible extensions using volatility-aware hybrids and multi-scale or global attention to improve robustness across different regimes and assets. Overall, the findings show that short-horizon predictability is fragile, and that careful implementation, probability calibration, and risk control are crucial if weak statistical signals are to be turned into investable strategies.