MTRGL:Effective Temporal Correlation Discerning through Multi-modal Temporal Relational Graph Learning

TL;DR

The paper tackles identifying temporal correlations for pair trading by constructing a dynamic multi-modal graph from time-series prices and discrete features, reframing correlation discovery as a temporal link-prediction problem. It introduces MTRGL, a memory-based temporal graph neural network with encoder (message, memory, embedding) and decoder components, trained via binary cross-entropy with contrastive learning and a lag-one batching scheme. Empirical results on multiple markets show that MTRGL outperforms static-feature, LSTM, CNN, and cointegration baselines, with ablations confirming the benefit of both feature information and graph structure. The approach has practical potential to refine automated pair trading and could extend to other finance problems requiring integrated multi-modal temporal graph learning.

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.

Figures (2)

• Figure 1: Illustration of the Pair Trading Strategy. Entities A and B are two correlated entities whose price tends to move together. The circled section indicates a temporal divergence of the price of the two entities due to market fluctuation or inefficiency. During the divergent period, the trader can simultaneously take a long position on entity B and a short position on entity A, and benefit from the divergence period when the movement of A and B converge again.
• Figure 2: Figure \ref{['fig:graph_construct']} displays how the time series data spanning from $[0,T]$ is partitioned into a sequence of intervals $\{ \Delta_1,...,\Delta_n \}$. The latter portion of this figure presents how the information from these time series segments is amalgamated with feature data to create the sequence of temporal investment graphs $\{G_1,...,G_n\}$. On the other hand, Figure \ref{['fig:train_pipe']} provides a visualization of MTRGL's training process. The constructed graphs are processed as event batches and employed for model updates. The loss of event batches $B_i$, denoted as $\mathcal{L}(B_i, \theta_{k}^{(j)})$, is calculated using the parameter $\theta_{k}^{(j)}$ obtained from the $k$-th epoch and $j$-th iteration. The parameters derived from the final iteration of each epoch serve as initial values for the subsequent epoch, denoted as $\theta_{k}^{(K-1)} = \theta_{k+1}^{(0)}$.