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

TL;DR

This work presents the Energy-based Parallel Graph Attention Neural Network, a novel approach for predicting future movements for multiple stocks that generates a dynamic stock graph with the energy difference between stocks and Boltzmann distribution, capturing evolving inter-dependencies between stocks.

Abstract

Graph neural networks have shown remarkable performance in forecasting stock movements, which arises from learning complex inter-dependencies between stocks and intra-dynamics of stocks. Existing approaches based on graph neural networks typically rely on static or manually defined factors to model changing inter-dependencies between stocks. Furthermore, these works often struggle to preserve hierarchical features within stocks. To bridge these gaps, this work presents the Energy-based Parallel Graph Attention Neural Network, a novel approach for predicting future movements for multiple stocks. First, it generates a dynamic stock graph with the energy difference between stocks and Boltzmann distribution, capturing evolving inter-dependencies between stocks. Then, a parallel graph attention mechanism is proposed to preserve the hierarchical intra-stock dynamics. Extensive experiments on five real-world datasets are conducted to validate the proposed approach, spanning from the US stock markets (NASDAQ, NYSE, SP) and UK stock markets (FTSE, LSE). The experimental results demonstrate that EP-GAT consistently outperforms competitive five baselines on test periods across various metrics. The ablation studies and hyperparameter sensitivity analysis further validate the effectiveness of each module in the proposed method. The raw dataset and code are available at https://github.com/theflash987/EP-GAT.

Figures (4)

• Figure 1: The overall architecture of EP-GAT is structured as follows: 1) utilizing multiple historical stock indicators time series to construct the stock graph $\mathcal{G}_t(\mathbf{X}_t, \mathbf{A}_t)$, where the adjacency matrix $\mathbf{A}_t$ is generated by energy differences between stocks and the Boltzmann Distribution. It is designed to capture the evolving inter-dependencies between stocks; 2) applying the parallel graph attention blocks on the generated stock graph $\mathcal{G}_t(\mathbf{X}_t, \mathbf{A}_t)$, which learns the latent representation $\mathbf{H}'_l$ and preserves hierarchical temporal features within stocks; after $L$ blocks learning process the final representation $\mathbf{H}'_L$ is obtained. Then, a linear reshaping layer is applied to generate the predicted label matrix $\mathbf{C}_t$, which represents the stock movements on future $\phi$ steps.
• Figure 2: Detailed illustration of the parallel graph attention mechanism. Here, $\mathbf{H}_l$ denotes the latent representation from propagation, and $\mathbf{H}'_{l}$ denotes the latent representation matrix from the parallel graph attention. The concatenation is performed along the columns, and the GATv2Conv represents the layer-wise graph attention operation from brody2021attentive.
• Figure 3: The EP-GAT comprises two components: energy-based stock graph generation and parallel graph attention mechanism. The ablation studies are conducted by replacing energy-based stock graph generation with the pre-defined stock graph or removing the parallel graph attention mechanism on the backbone GATv2Conv. The grey dot lines denote the best-performing baseline models.
• Figure 4: Hyperparameter sensitivity results of the lag window size $\tau$, scaling factor $k$ and threshold $s$ on five datasets over test periods.