MTRGL: EFFECTIVE TEMPORAL CORRELATION DISCERNING THROUGH MULTI-MODAL TEMPORAL RELATIONAL GRAPH LEARNING

Su Junwei, Wu Shan, Li Jinhui

Abstract

Identifying stable correlations between financial assets is a cornerstone of market-neutral strategies like pair trading. However, traditional statistical methods often fail to capture the dynamic and multi-modal nature of market relationships. This paper presents MTRGL (Multi-modal Temporal Relation Graph Learning), a novel framework that reframes pair trading as a dynamic graph link prediction task. By integrating time-series price data with discrete metadata through a memory-based Temporal Graph Neural Network (TGNN), MTRGL effectively discerns evolving asset correlations. Experimental results on the KOSPI, S&P 500, and HSI indices demonstrate that MTRGL significantly outperforms traditional and deep learning baselines in both accuracy and error reduction.

1. Introduction

Pair trading is a quantitative strategy designed to exploit temporary market inefficiencies. It involves identifying two assets whose prices historically move in tandem. When the price ratio of these assets diverges from its historical mean, a trader takes a long position in the undervalued asset and a short position in the overvalued one, profiting when the prices eventually converge.

The primary challenge in this strategy is temporal correlation discerning: the ability to identify which asset pairs will maintain their relationship in a noisy, non-stationary market. Traditional statistical approaches, such as cointegration tests, are often too static to adapt to rapid market shifts. Conversely, standard machine learning models often overlook the structural relationships between assets or fail to incorporate fundamental metadata that could stabilize correlation predictions. MTRGL addresses these gaps by modeling the market as an evolving network of interconnected nodes.

2. The MTRGL Framework

MTRGL treats the stock market as a dynamic graph where stocks are represented as nodes and their correlations are represented as edges. The framework operates through three distinct stages:

2.1 Temporal Graph Construction

The timeline is partitioned into discrete intervals. For each interval, a graph is constructed based on the Normalized Historical Difference (NHD) between asset prices. An edge is established between two nodes if their price correlation exceeds a predefined threshold. This transforms the problem of finding pairs into a link prediction task: predicting which edges will exist in the subsequent time step.

2.2 The Encoder: Memory-based TGNN

The core of MTRGL is its encoder, which utilizes a memory-based Temporal Graph Neural Network to capture both long-term trends and immediate neighborhood effects:

  • Message & Memory: The model maintains a node-specific memory updated via a Gated Recurrent Unit (GRU). As new correlation "events" occur, the GRU updates the stock's state, allowing the model to retain a historical context of the asset's behavior.
  • Multi-modal Embedding: MTRGL combines internal memory with discrete features, such as sector classifications and market capitalization. This multi-modal approach ensures that the model considers both price action and fundamental commonalities.
  • Temporal Graph Attention: A graph attention mechanism aggregates information from a stock’s neighbors. This allows the model to learn from the collective behavior of related assets, providing a more robust signal than analyzing a single pair in isolation.

2.3 The Decoder

The final stage is a Multi-Layer Perceptron (MLP) decoder. It takes the learned embeddings of two candidate stocks and outputs the probability of a correlation existing between them in the next time interval.

3. Experimental Evaluation

3.1 Datasets

The researchers evaluated MTRGL using daily closing price data from 2015 to 2019 across three major global indices:

  • KOSPI (South Korea)
  • S&P 500 (United States)
  • HSI (Hong Kong)

The study focused on the pre-pandemic period to establish a baseline for standard market dynamics, avoiding the "black swan" distortions of 2020. Discrete features included sector classifications and market capitalization.

3.2 Performance Metrics

MTRGL was compared against four baselines: Cointegration (statistical), MLP (feature-based), and CNN/LSTM (time-series based). The model demonstrated superior performance across all indices:

  • Accuracy: On the S&P 500, MTRGL achieved an Average Precision (AP) of 74.2%, significantly higher than the 65.6% achieved by the next best performer (LSTM).
  • Error Reduction: The Mean Absolute Percentage Error (MAPE), measuring the variance in predicted correlation values, was notably lower for MTRGL. In the HSI index, MTRGL reached a MAPE of 16.8%, compared to 23.3% for the LSTM baseline.

4. Discussion and Key Insights

The success of MTRGL highlights two critical factors in financial modeling:

  1. The Power of Memory: Financial relationships are not instantaneous; they are built on history. The GRU-based memory module allows MTRGL to filter out transient noise and focus on persistent relationships.
  2. Multi-modal Synergy: Price data alone is often insufficient for predicting long-term correlations. By integrating "static" metadata like sector information, the model gains a necessary prior that helps distinguish between coincidental price movements and fundamentally driven correlations.

Ablation studies confirmed these insights, showing a sharp decline in performance when either the structural graph information or the multi-modal features were removed.

5. Conclusion and Future Work

MTRGL provides a robust framework for discerning temporal correlations in complex financial markets. By reframing pair trading as a dynamic link prediction task, it leverages the strengths of Temporal Graph Neural Networks to outperform traditional methods.

Future research will focus on testing the model's resilience during extreme market volatility (such as the COVID-19 pandemic) and optimizing the computational efficiency of the TGNN to handle larger datasets with thousands of nodes. Additionally, the authors suggest that this multi-modal graph approach could be adapted for broader financial applications, including portfolio optimization and systemic risk assessment.

Summary updated

Abstract

In this study, we explore the synergy of deep learning and financial market applications, focusing on pair trading. This market-neutral strategy is integral to quantitative finance and is apt for advanced deep-learning techniques. A pivotal challenge in pair trading is discerning temporal correlations among entities, necessitating the integration of diverse data modalities. Addressing this, we introduce a novel framework, Multi-modal Temporal Relation Graph Learning (MTRGL). MTRGL combines time series data and discrete features into a temporal graph and employs a memory-based temporal graph neural network. This approach reframes temporal correlation identification as a temporal graph link prediction task, which has shown empirical success. Our experiments on real-world datasets confirm the superior performance of MTRGL, emphasizing its promise in refining automated pair trading strategies.
BibTeX
@inproceedings{junwei2024mtrgl,
  author = {Su Junwei and Wu Shan and Li Jinhui},
  title = {{MTRGL: EFFECTIVE TEMPORAL CORRELATION DISCERNING THROUGH MULTI-MODAL TEMPORAL RELATIONAL GRAPH LEARNING}},
  booktitle = {Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing},
  year = {2024},
  url = {https://openreview.net/forum?id=0h5Yc1vzPK},
}
Figure