Combined Effects of Fat‐Tail and Spread Forecasting on Pairs Trading: A Hybrid Model Based on Integrating VAR With GRU Models

ABSTRACT Pairs trading, a popular algorithmic trading strategy, exploits the short‐term price difference (spread) between two comoving assets. Empirically, the spread distribution of most assets in pairs trading has a fat‐tail characteristic that does not follow a normal distribution. However, existing trading strategies did not consider these characteristics and attempted to improve performance by applying complex machine learning techniques. This paper targets cryptocurrency assets that show the fat‐tail characteristics of the spread distribution well due to their high volatility and presents a new hybrid model that can solve these characteristics in machine learning, namely, VAR‐GRU‐QT (Vector AutoRegression–Gated Recurrent Unit–Quantile Transform). The VAR‐GRU model is proposed for spread prediction, whereas the QT‐based pair trading strategy seeks to capture trading signals. The empirical analysis results show that the VAR‐GRU‐QT model improves prediction performance over the comparative model and significantly improves pairs trading performance. In addition, it was confirmed that as the threshold of the trading signal increases, the fat‐tail effect appears more prominently in the hybrid model. In conclusion, by considering the non‐Gaussian distribution of the spread, the prediction accuracy of the hybrid model is improved, and by additionally applying the quantile‐based trading methodology, the trading performance of the pair trading strategy is significantly improved over the comparative model.