Dynamic Forecasting and Temporal Feature Evolution of Stock Repurchases in Listed Companies Using Attention-Based Deep Temporal Networks

Battle Count: 163

Highly relevant for quantitative trading as it provides a forward-looking early warning system for stock repurchase events, which are known to generate significant excess returns (Alpha). The model achieves AUC of 0.84-0.86 with 1-year lead time and maintains 0.78 AUC even at 3-year lead time. The XAI component identifies specific financial triggers (cash flow surges, undervaluation) that can be translated into actionable trading signals. The micro-case analysis shows probability lifts of 11-12x over baseline, indicating strong signal-to-noise ratio for alpha generation.

The architecture involves multiple components: TCN with dilated causal convolutions and residual blocks, LSTM with gating mechanisms, additive attention mechanism, Focal Loss optimization, and SHAP attribution. The data preprocessing requires reconstructing panel data into 3D tensors with sliding windows. Training uses Adam optimizer with BPTT. The XAI post-hoc analysis adds computational overhead. However, all components are well-established in deep learning literature, and standard frameworks (PyTorch/TensorFlow) support these architectures.

The paper provides detailed mathematical formulations, hyperparameter descriptions (sliding window size, dilation factors, network dimensions), and experimental settings (rolling-window backtesting, data splitting). However, no code repository is mentioned, and specific hyperparameter values (learning rate, batch size, number of layers, window size L) are not fully enumerated. The dataset is based on publicly available Chinese A-share financial data, but exact preprocessing scripts are not provided.