EP-GAT: Energy-based Parallel Graph Attention Neural Network for Stock Trend Classification

This paper introduces EP-GAT, an Energy-based Parallel Graph Attention Neural Network that utilizes dynamic stock graphs derived from energy differences and a parallel attention mechanism to effectively model evolving inter-dependencies and hierarchical intra-stock dynamics, achieving superior performance in stock trend classification across multiple global markets compared to existing baselines.

The Gist

Imagine you are trying to predict the weather, but instead of looking at clouds, you are looking at the stock market. The problem is that the stock market is chaotic. Stocks don't just move on their own; they are like a giant, noisy crowd where everyone is whispering to everyone else. If one person (a stock) sneezes, the person next to them might catch a cold, and the whole row might start coughing.

For a long time, computer programs trying to predict these movements made two big mistakes:

1. They treated stocks like strangers: They assumed Stock A doesn't care about Stock B, even though they are in the same industry or influenced by the same news.
2. They used a static map: They drew a map of who talks to whom based on old rules (like "companies in the same sector are friends") and never updated it, even though the market changes every second.

Enter EP-GAT (Energy-based Parallel Graph Attention Neural Network). Think of this as a super-smart, real-time weather forecaster for the stock market. Here is how it works, broken down into simple concepts:

1. The "Energy" Map (Dynamic Graph Generation)

Most old models used a fixed map. EP-GAT draws a new map every single day.

• The Analogy: Imagine a room full of people (stocks). Some are jumping around with high energy (volatile stocks), and some are sitting calmly (stable stocks).
• How it works: The model calculates the "energy" of each stock based on its recent history. It then uses a physics concept called the Boltzmann Distribution (think of it as a "temperature" gauge) to figure out who is likely to influence whom.
• The Result: If Stock A has high energy and Stock B is calm, the model realizes, "Hey, Stock A is probably going to shake up Stock B." It draws a line between them. If the energy levels change the next day, the lines on the map change too. This captures the evolving relationships between stocks.

2. The "Parallel" Listening Party (Parallel Graph Attention)

Once the map is drawn, the model needs to listen to the conversations. Old models listened to the crowd in a single line, which often got confusing. If you listen to a conversation too many times, you might forget the original point or get the details mixed up.

• The Analogy: Imagine a group of friends trying to solve a puzzle.
  • Old Way: They pass the puzzle piece down a line. By the time it reaches the last person, the piece might be bent or lost.
  • EP-GAT Way: They all sit in a circle and look at the puzzle piece at the same time from different angles.
• How it works: The "Parallel Graph Attention" mechanism lets the model look at the stock's history from multiple "layers" or perspectives simultaneously. It preserves the hierarchical features—meaning it remembers the big picture (the trend) and the small details (the daily fluctuations) without letting one distort the other. It's like having a team of detectives where one looks at the crime scene, another looks at the timeline, and a third looks at the suspects, and they all share their findings instantly without losing their original notes.

3. The Prediction

After drawing the dynamic map and listening to the multi-layered conversations, the model makes a guess: Will this stock go up or down tomorrow?

Why is this better?

The authors tested EP-GAT on five different stock markets (like the US NASDAQ and UK FTSE) with hundreds of stocks.

• The Competition: They compared it to five other smart models.
• The Winner: EP-GAT won consistently. It was better at spotting the hidden connections between stocks and didn't get confused by the noise.

The Bottom Line

Think of EP-GAT as a smart, adaptive radar system.

• Instead of using a static map of the world, it updates the map every second based on how "energetic" the stocks are.
• Instead of listening to one voice at a time, it listens to the whole choir in harmony, keeping the distinct voices clear.

This allows it to predict stock trends much more accurately than the old methods, helping investors and policymakers see the future with a clearer lens.

One small catch: The current version assumes the influence goes both ways (if A affects B, B affects A). In reality, a giant company might influence a small one, but the small one might not influence the giant. The authors plan to fix this in future versions to make the "map" even more realistic.

Technical Summary

1. Problem Statement

Stock market prediction is a challenging task due to the stochastic nature and high volatility of financial markets. While deep learning models (like RNNs and Transformers) have improved performance, existing Graph Neural Network (GNN) approaches for stock forecasting suffer from two critical limitations:

• Static Graph Construction: Most existing methods rely on static, pre-defined factors (e.g., industry sectors, supply chains) to model inter-stock dependencies. These fail to capture the dynamic, evolving inter-dependencies between stocks that change over time.
• Loss of Hierarchical Features: Standard GNN propagation mechanisms often distort or lose hierarchical intra-stock features (temporal patterns at different scales) as information propagates through layers, leading to suboptimal representation learning.

The paper aims to address these gaps by developing a framework that dynamically models changing stock relationships while preserving the hierarchical structure of individual stock data.

2. Methodology: EP-GAT

The proposed Energy-based Parallel Graph Attention Neural Network (EP-GAT) consists of two core modules:

A. Energy-based Stock Graph Generation

Instead of using static industry relations, EP-GAT constructs a dynamic stock graph at each time step $t$ based on the energy differences between stocks and the Boltzmann distribution.

• Energy Definition: The "energy" ($E_i$) of a stock is calculated as the sum of squares of its historical feature vector (lag window) at time $t$.
• Boltzmann Distribution: The probability of a connection (edge weight) between stock $i$ and stock $j$ is modeled using the Boltzmann distribution:
  $$A_{i,j}^t = \frac{e^{-\frac{|E_i - E_j|}{k\tau}}}{\sum_{o=1}^{N} e^{-\frac{|E_i - E_o|}{k\tau}}}$$
  Where:
  • $|E_i - E_j|$ is the energy difference.
  • $\tau$ (lag window size) acts as the system temperature ($T$), controlling the sensitivity to energy differences.
  • $k$ is a scaling factor.
• Sparsification: A threshold $s$ is applied to the adjacency matrix to remove weak connections and reduce noise, ensuring the graph remains sparse and meaningful.

B. Parallel Graph Attention Mechanism

To address the distortion of hierarchical features during GNN propagation, EP-GAT introduces a Parallel Graph Attention mechanism.

• Decoupling: Unlike standard GNNs where representation learning and propagation are tightly coupled, this mechanism decouples them.
• Preservation: It preserves the latent representations from each propagation layer ($H_{l-1}$) and transforms them using a Multi-Head Attention mechanism.
• Concatenation: The final representation for a layer is formed by concatenating the preserved features with the transformed features:
  $$H'_l = \gamma(H'_{l-1} || (\eta(H_{l-1}) + H_{l-1}W_l))$$
  This ensures that hierarchical temporal features are not lost or distorted as the model deepens.

C. Objective Function

The model performs a binary classification task (Upward vs. Downward trend) for the next trading day ($\phi=1$). It is trained using a standard cross-entropy loss function over the predicted labels and ground truth.

3. Key Contributions

1. Dynamic Graph Construction: A novel method for generating stock graphs based on energy differences and the Boltzmann distribution, allowing the model to adapt to stochastic market changes rather than relying on static industry labels.
2. Parallel Graph Attention: A new architecture that preserves hierarchical intra-stock features by decoupling representation learning from propagation, preventing feature distortion in deep GNNs.
3. Comprehensive Empirical Validation: Extensive experiments on five real-world datasets (LSE, FTSE, NYSE, NASDAQ, S&P) covering 503 stocks over five years, demonstrating consistent superiority over baselines.

4. Experimental Results

The authors evaluated EP-GAT against five competitive baselines:

• GNN-based: GraphWaveNet, HyperStockGAT, STGCN.
• Transformer-based: Informer, ST-TIS.

Performance Metrics: Accuracy (ACC), Matthews Correlation Coefficient (MCC), and F1-Score.

Key Findings:

• Superiority: EP-GAT consistently outperformed all baselines across all five datasets.
  • Average Improvement: +7.61% in ACC, +2.63 $\times 10^{-2}$ in MCC, and +0.06 in F1-Score compared to the best baseline.
• Dynamic vs. Static: Models using dynamic graphs (EP-GAT, GraphWaveNet) significantly outperformed those using static graphs (HyperStockGAT, STGCN), validating the importance of modeling evolving inter-stock relations.
• Hierarchical Importance: Removing the parallel attention mechanism caused a performance drop of ~8%, proving the necessity of preserving hierarchical features.
• Ablation Study: Replacing the energy-based graph with a pre-defined industry graph resulted in a ~2.25% drop in accuracy, confirming the value of the energy-based approach.

Hyperparameter Sensitivity:

• Lag Window ($\tau$): Optimal performance was found in the range of 14–23 days.
• Scaling Factor ($k$): Varies by dataset; lower values (e.g., 0.08) worked best for LSE/NASDAQ, suggesting these markets have more intricate interdependencies.
• Threshold ($s$): A balance is required; too low introduces noise, while too high removes critical connections.

5. Significance and Limitations

Significance:
This work bridges the gap between statistical physics (Boltzmann distribution) and financial time-series forecasting. It provides a robust framework for capturing the non-stationary, stochastic nature of stock markets while ensuring that the internal temporal dynamics of individual stocks are not lost during graph propagation. It sets a new state-of-the-art for multi-stock trend classification.

Limitations:

1. Undirected Graphs: The current model assumes undirected edges, ignoring the fact that influential stocks may affect others without being affected in return (directed influence).
2. Scalability: The parallel attention mechanism involves feature-dependent operations (concatenation) and multi-head attention, which may become computationally expensive and difficult to train when scaling to very large numbers of stocks.

Future Work: The authors plan to incorporate directed inter-stock relations and optimize the parallel attention mechanism using retention mechanisms to handle larger markets.