Graph Learning for Foreign Exchange Rate Prediction and Statistical Arbitrage

TL;DR

This paper introduces a novel graph learning approach for foreign exchange rate prediction and statistical arbitrage, leveraging multi-currency relationships and time lag considerations, resulting in significant performance improvements.

Contribution

It develops a two-step graph learning framework for FX prediction and arbitrage, addressing prior gaps by modeling complex relationships and time delays explicitly.

Findings

Significant reduction in mean squared error for FX prediction.

61.89% higher information ratio in FX arbitrage strategy.

45.51% higher Sortino ratio compared to benchmark.

Abstract

We propose a two-step graph learning approach for foreign exchange statistical arbitrages (FXSAs), addressing two key gaps in prior studies: the absence of graph-learning methods for foreign exchange rate prediction (FXRP) that leverage multi-currency and currency-interest rate relationships, and the disregard of the time lag between price observation and trade execution. In the first step, to capture complex multi-currency and currency-interest rate relationships, we formulate FXRP as an edge-level regression problem on a discrete-time spatiotemporal graph. This graph consists of currencies as nodes and exchanges as edges, with interest rates and foreign exchange rates serving as node and edge features, respectively. We then introduce a graph-learning method that leverages the spatiotemporal graph to address the FXRP problem. In the second step, we present a stochastic optimization problem to exploit FXSAs while accounting for the observation-execution time lag. To address this problem, we propose a graph-learning method that enforces constraints through projection and ReLU, maximizes risk-adjusted return by leveraging a graph with exchanges as nodes and influence relationships as edges, and utilizes the predictions from the FXRP method for the constraint parameters and node features. Moreover, we prove that our FXSA method satisfies empirical arbitrage constraints. The experimental results demonstrate that our FXRP method yields statistically significant improvements in mean squared error, and that the FXSA method achieves a 61.89% higher information ratio and a 45.51% higher Sortino ratio than a benchmark. Our approach provides a novel perspective on FXRP and FXSA within the context of graph learning.