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

This paper proposes a hybrid Temporal Convolutional Network and Attention-based LSTM model using Chinese A-share data to outperform traditional static methods in predicting stock repurchases, while leveraging Explainable AI to reveal that prolonged undervaluation serves as a long-term motive and sharp cash flow increases act as short-term triggers for these decisions.

The Gist

Imagine you are trying to predict when a company will decide to buy back its own shares from the stock market. This is a big deal: it's like a company saying, "We think our stock is on sale, and we have enough cash to buy it back to make our shareholders happy."

For a long time, economists tried to predict this by looking at a single "snapshot" of a company's finances, like taking a photo of a runner at one specific second and trying to guess if they will win the race. The problem is, companies don't make decisions based on just one moment; they make them based on a movie of their history.

This paper introduces a new, super-smart way to watch that "movie" using Deep Learning (a type of advanced AI) and Explainable AI (AI that can tell you why it made a guess).

Here is the breakdown of their approach using simple analogies:

1. The Problem: The "Photo" vs. The "Movie"

Traditional methods look at financial data like a photo album. They see that a company had low stock prices last year and high cash this year, but they treat these as separate, unrelated events. They miss the story connecting them.

• The Paper's Solution: They treat the data like a movie reel. They look at the last 3 years of a company's life as a continuous sequence. They want to see not just what happened, but how things evolved over time.

2. The Engine: A Three-Part Detective Team

The authors built a digital detective team (a neural network) with three specialized members to analyze this "movie":

• The "Noise Filter" (TCN): Imagine a noisy room where people are shouting. This part of the AI is like a noise-canceling headphone. It filters out the daily market chaos and short-term panic to find the smooth, underlying trends in the company's finances.
• The "Memory Keeper" (LSTM): This is the AI's long-term memory. It remembers that a company might have been struggling with debt for three years straight, even if they had one good month recently. It understands that long-term patterns matter more than a single lucky break.
• The "Spotlight" (Attention Mechanism): This is the most clever part. Imagine a detective watching a crime scene video. They don't stare at every second equally; they zoom in on the specific moment the suspect pulled the trigger.
  • The AI uses this "Spotlight" to realize: "Hey, the stock price has been low for years (the motive), but the decision to buy back shares happened specifically because cash flow suddenly spiked last month (the trigger)."

3. The "Black Box" Problem: Making the AI Talk

Usually, deep learning is a "Black Box." You put data in, and a number comes out, but you don't know why. In finance, you can't just trust a number; you need to know the logic.

• The Solution (XAI): The authors added a "translator" (using a technique called SHAP). It's like asking the AI, "Why did you pick this company?"
• The Answer: The AI can now point to the screen and say: "I predicted a buyback because the company's stock was undervalued for a long time (the motive), AND they suddenly had a massive pile of cash (the trigger). However, if they had too much debt, I would have said 'No' because debt acts like a brake."

4. What They Discovered (The "Aha!" Moments)

By watching these "movies" of thousands of Chinese companies, the AI found two main rules that drive stock buybacks:

1. The "Underwater" Motive: If a company's stock has been "underwater" (undervalued) for a long time, the management wants to buy it back. This is the long-term reason.
2. The "Cash Spigot" Trigger: Wanting to buy back shares isn't enough; you need the money. The decision usually happens the moment the "cash spigot" opens up (a sudden, sharp increase in cash flow). This is the short-term trigger.

The Twist: The AI also found that Debt is a "Veto Power." Even if a company is undervalued and has cash, if they have too much debt, they cannot buy back shares. The debt acts like a heavy anchor holding them down.

5. Why This Matters

• For Investors: It's like having a crystal ball that doesn't just guess, but explains its reasoning. It helps find companies that are about to buy back their stock (which usually makes the stock price go up) much earlier than traditional methods.
• For Regulators: It helps spot risky companies or market anomalies before they become a crisis.
• For Science: It proves that looking at the history of a company (the movie) is much better than looking at a single snapshot (the photo).

In a nutshell: The authors built an AI that watches the financial history of companies like a movie, filters out the noise, remembers the long-term plot, zooms in on the critical moments, and then explains exactly why it thinks a company is about to buy back its stock. It turns a "black box" guess into a clear, logical story.

Technical Summary

1. Problem Statement

The paper addresses the limitations of traditional methods in predicting corporate stock repurchases, a critical financial strategy for signaling undervaluation and optimizing capital structure.

• The Gap: Traditional econometric models (e.g., Logistic Regression, Probit) and static machine learning algorithms (e.g., XGBoost, SVM) treat financial data as independent "cross-sectional snapshots." They fail to capture non-linear temporal dependencies and lagging effects inherent in corporate decision-making.
• The Challenge: Stock repurchase decisions are often the result of multi-year strategic planning (e.g., long-term undervaluation) combined with short-term triggers (e.g., sudden cash flow surges). Static models lose this "temporal memory" and the dynamic evolution of financial features, leading to suboptimal prediction accuracy and a lack of interpretability regarding when and why a decision is made.
• Data Imbalance: Repurchase events are a minority class compared to non-repurchase events, causing standard models to suffer from the "Accuracy Paradox" (predicting the majority class to achieve high accuracy).

2. Methodology

The authors propose a Dynamic Early Warning System that integrates deep temporal networks with Explainable AI (XAI).

A. Data Reconstruction (Static to Dynamic)

• Dataset: Large-scale panel data of Chinese A-share listed companies (2014–2024).
• Feature Engineering: Instead of flattening data, the authors reconstruct static financial indicators (Tobin's Q, Net Operating Cash Flow, Debt-to-Asset Ratio, etc.) into 3D temporal tensors ($S \times L \times F$) using a Sliding Window Mechanism.
  • $S$: Number of samples.
  • $L$: Time window length (e.g., 3 years).
  • $F$: Number of financial features.
• Preprocessing: Z-Score standardization based on training sets and temporal forward-filling for missing data to prevent look-ahead bias.

B. Model Architecture: TCN-Att-LSTM

The core prediction engine is a hybrid deep learning model designed to capture both local shocks and long-term trends:

1. Temporal Convolutional Network (TCN):
   • Uses 1D Dilated Causal Convolutions to extract local financial fluctuations and short-term shocks.
   • Ensures strict temporal causality (no future information leakage) and expands the receptive field to cover macroeconomic cycles.
   • Includes Residual Blocks to mitigate gradient vanishing.
2. Long Short-Term Memory (LSTM):
   • Processes the TCN output sequence to model long-term temporal dependencies (e.g., multi-year debt accumulation or valuation depression).
   • Uses gating mechanisms (Forget, Input, Output gates) to retain or forget historical financial states.
3. Temporal Attention Mechanism:
   • Applies an Additive Alignment Model to compute attention weights ($\alpha_t$) for each time step in the window.
   • Identifies which specific years (e.g., $T-1$ vs. $T-3$) contributed most to the final decision, effectively highlighting "trigger points."
4. Output Layer:
   • A Fully Connected layer with Dropout and Sigmoid activation outputs the probability of a repurchase.
   • Loss Function: Uses Focal Loss to address class imbalance, down-weighting easy negative samples and focusing optimization on hard-to-classify minority samples.

C. Explainability (XAI)

To overcome the "black box" nature of deep learning, the study integrates:

• Temporal Attention Weights: Visualizes the evolution of importance over time.
• SHAP (SHapley Additive exPlanations): Based on Cooperative Game Theory, it decomposes the prediction to quantify the marginal contribution of each feature at each time step, validating economic theories.

3. Key Contributions

1. Novel Architecture: First application of a hybrid TCN-Attention-LSTM framework specifically for low-frequency, multi-year corporate financial decision prediction (stock repurchases).
2. Temporal Feature Reconstruction: A rigorous methodology to transform static panel data into 3D tensors, preserving the autocorrelation and evolutionary logic of financial conditions.
3. Economic Interpretability: Successfully bridges deep learning and classical finance theory. The model does not just predict if a repurchase will happen, but reveals the temporal causal chain: distinguishing between long-term "underlying motives" and short-term "action triggers."
4. Robustness Validation: Demonstrates superior performance in highly imbalanced, real-world market scenarios using rolling-window backtesting.

4. Experimental Results

• Performance Metrics: The proposed TCN-Att-LSTM model significantly outperformed baselines (Logistic Regression, XGBoost, standalone TCN, and LSTM) across all test years (2021–2024).
  • AUC: Reached 0.860 (2023) and 0.847 (2024), compared to ~0.68 for XGBoost.
  • PR-AUC: Stabilized above 0.60, indicating strong performance in identifying the minority class (repurchases).
  • F1-Score: Consistently higher than static models, proving better balance between precision and recall.
• Lead-Time Robustness: Even with a 3-year lead time, the model maintained an AUC of 0.782, surpassing the 1-year performance of traditional models. This confirms that strategic repurchase signals are embedded in long-term financial trajectories.
• Ablation Study: Confirmed that combining TCN (local features) and LSTM (long-term memory) with Attention (trigger identification) yields a synergistic "1+1 > 2" effect.

5. Significance and Economic Insights

The study provides micro-algorithmic evidence that validates and extends classical corporate finance theories:

• Dual-Core Drivers: Tobin's Q (valuation) and Net Operating Cash Flow are the dominant drivers.
  • Undervaluation Hypothesis: Long-term depression in Tobin's Q acts as the underlying motive (stable high importance over 3 years).
  • Free Cash Flow Hypothesis: A sharp surge in cash flow acts as the decisive trigger (spike in importance at $T-1$).
• Market Timing: The attention mechanism reveals that management often exploits short-term mispricing windows, with decision weights heavily skewed toward the immediate pre-decision period.
• Financial Constraints: SHAP analysis on high-leverage firms showed that Debt-to-Asset Ratio becomes the dominant feature, effectively "vetoing" repurchase decisions regardless of cash flow or valuation, validating the Financial Constraints theory.
• Synergistic Effects: The model identified a non-linear threshold where abundant cash flow only triggers repurchases when combined with extreme undervaluation.

Conclusion

This paper establishes a new paradigm for intelligent financial supervision and quantitative investment. By moving from static cross-sectional analysis to dynamic temporal deep learning, it offers a tool that not only predicts stock repurchases with high accuracy but also provides logical traceability of the decision-making process, effectively bridging the gap between advanced AI and traditional economic theory.