Modeling the Momentum Spillover Effect for Stock Prediction via Attribute-Driven Graph Attention Networks

Rui Cheng, Qing LiAAAI 2021 · DOI

Introduction

In modern financial markets, the movement of a stock is rarely an isolated event. Instead, firms are interconnected through complex networks of supply chains, competition, and shared industries. A critical phenomenon in this context is "momentum spillover," where the past performance of one firm serves as a predictor for the future returns of its linked neighbors. While Graph Neural Networks (GNNs) have been applied to model these dependencies, they often struggle with two primary limitations: the use of static, noisy predefined relations and an "attribute-blind" approach to information aggregation.

This paper, originally presented at AAAI-21, introduces the Attribute-Driven Graph Attention Network (AD-GAT). This framework addresses these gaps by dynamically learning latent firm relationships and employing a state-sensitive aggregation mechanism that considers the specific attributes of connected firms to determine the strength and relevance of momentum spillover.

The AD-GAT Framework

The AD-GAT architecture is designed to capture high-level interactions between diverse market signals and model the conditional nature of financial influence. The process is divided into three main stages:

  1. Tensor Fusion (TF) Module: Rather than simple concatenation, the TF module utilizes a bilinear tensor product to fuse numerical market data with textual news sentiment. This allows the model to capture synergistic effects, such as how specific news sentiment might amplify the impact of a sudden price spike.
  2. Sequential Embedding: The fused historical data is passed through Recurrent Neural Networks (RNNs) to generate sequential embeddings ($v_i^t$) for each firm, preserving temporal dependencies.
  3. Attribute-Driven Graph Attention (AD-GAT): This is the core innovation, consisting of two sub-modules:
  4. Attribute-Mattered (AM) Aggregator: Unlike standard linear aggregators, the AM Aggregator uses a gate mechanism. It performs element-wise multiplication between a firm's features and a nonlinear transformation of its neighbor's attributes (e.g., P/E ratio, turnover). This ensures that spillover is only recorded if the "state" of the connected firm warrants it.
  5. Relation Building (RB) Module: To move beyond static industry labels, the RB module uses an unmasked attention mechanism to infer latent, dynamic relations from market data in real-time.
The overall architecture of the AD-GAT framework, illustrating the flow from historical data to final stock movement predictions.

Experimental Setup

The model was rigorously evaluated using a custom dataset of S&P 500 firms from February 2011 to November 2013. The data included:

  • Technical Indicators: Open, high, low, close prices, and trading volume.
  • Textual Media: Financial news from Reuters and Bloomberg, processed for sentiment using the L&M financial dictionary.
  • Predefined Relations: Five categories (industry, supply chain, competition, customer, and alliance) were used for baseline comparisons.

Experimental Results

AD-GAT demonstrated superior performance over a wide range of baselines, including traditional time-series models (LSTM, GRU) and state-of-the-art graph models like Temporal Graph Convolution (TGC).

Performance comparison of AD-GAT against various baseline models in terms of Directional Accuracy (DA) and Area Under Curve (AUC).

As shown in the results, AD-GAT achieved a Directional Accuracy (DA) of 56.47% and an AUC of 0.5894. This represents a significant relative improvement of approximately 6.4% in DA and 10.7% in AUC over the best-performing baselines.

Key findings from the analysis include:

  • Latent Relations vs. Human Labels: The model performed better when allowed to infer relationships dynamically (RB module) than when restricted to predefined categories like "Industry." This suggests that the market's underlying influence network is more fluid than static classifications suggest.
  • Conditional Spillover: Ablation studies confirmed that the Attribute-Mattered Aggregator is vital. By recognizing that a price drop in a peer firm is only a relevant signal under specific conditions (e.g., high trading volume), the model avoids the noise inherent in standard GNN aggregation.

Conclusion

AD-GAT provides a robust framework for stock prediction by treating momentum spillover as a conditional, attribute-driven process. By combining tensor-based feature fusion with dynamic relation building and state-sensitive aggregation, the model effectively navigates the noise of financial markets. Future work may involve integrating more advanced LLM-based sentiment analysis and testing the framework on more volatile asset classes like bulk futures or options.

Summary updated

BibTeX
@inproceedings{cheng2021modeling,
  author = {Rui Cheng and Qing Li},
  title = {{Modeling the Momentum Spillover Effect for Stock Prediction via Attribute-Driven Graph Attention Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i1.16077},
}
Figure