Dynamic Forecasting and Temporal Feature Evolution of Stock Repurchases in Listed Companies Using Attention-Based Deep Temporal Networks
Dynamic Forecasting and Temporal Feature Evolution of Stock Repurchases in Listed Companies Using Attention-Based Deep Temporal Networks
This paper presents a dynamic early warning system for stock repurchases using a hybrid deep temporal network that integrates Temporal Convolutional Networks (TCN), Long Short-Term Memory (LSTM) networks, and a temporal attention mechanism. The model is designed to overcome the limitations of traditional static models (e.g., logistic regression, XGBoost) that ignore temporal dependencies in panel data. By reconstructing financial indicators into 3D sliding window tensors, the network captures both short-term fluctuations and long-term evolutionary patterns in corporate financial conditions. The use of Focal Loss addresses the severe class imbalance inherent in repurchase prediction. Experimental results on Chinese A-share data from 2014 to 2024 demonstrate that the proposed model consistently outperforms baseline models in terms of AUC, PR-AUC, and F1-score across multiple test years. Furthermore, the integration of Explainable AI (XAI) via SHAP and temporal attention weights provides economic interpretability: the attention mechanism reveals that financial improvements closer to the decision year are more influential, while SHAP attribution confirms the 'Undervaluation Hypothesis' (low Tobin's Q as a long-term driver) and the 'Free Cash Flow Hypothesis' (cash flow abundance as a short-term trigger). This work offers a powerful tool for quantitative investment and financial supervision, bridging deep learning with classical corporate finance theory.
Highlights
- 1Proposes a hybrid deep temporal network (TCN-Att-LSTM) for dynamic early warning of stock repurchases using low-frequency panel data.
- 2Reconstructs static financial indicators into 3D temporal tensors via sliding windows to capture long-term dependencies and lag effects.
- 3Integrates Focal Loss to handle severe class imbalance in repurchase prediction.
- 4Employs SHAP and temporal attention weights to provide explainable AI insights, revealing undervaluation as a long-term motive and cash flow as a short-term trigger.
- 5Outperforms traditional logistic regression and static XGBoost in AUC, PR-AUC, and F1-score across multiple lead times on Chinese A-share data from 2014 to 2024.
Methods
- MTemporal Convolutional Network (TCN) with dilated causal convolutions for local financial fluctuation extraction.
- MLong Short-Term Memory (LSTM) network for modeling long-term evolution of capital structure.
- MTemporal Attention mechanism (additive alignment) for identifying critical decision-triggering time steps.
- MSHAP (SHapley Additive exPlanations) for feature attribution and economic interpretation.
Results
- RThe proposed TCN-Att-LSTM model achieves superior AUC, PR-AUC, and F1-score compared to logistic regression, XGBoost, standalone TCN, and standalone LSTM across test years 2021-2024.
- RTemporal attention weights show that financial conditions closer to the decision year have higher impact, with a sharp improvement in cash flow acting as a trigger.
- RSHAP analysis confirms that prolonged undervaluation (low Tobin's Q) is the primary long-term motive, while cash flow abundance is the key short-term trigger.
- RThe model maintains robust predictive performance even with lead times of 2-3 years, demonstrating effective forward-looking early warning capability.
Analyze Paper
Generate insights from "Dynamic Forecasting and Temporal Feature Evolution of Stock ...".