Pair Trading: Multivariate Pairs Formation with Machine Learning Forecasting-Based Strategy

Published on Jun 7, 2024. Pair Trading, a popular statistical arbitrage strategy, involves shorting overvalued assets and buying undervalued ones to profit from their return to equilibrium. This research explores and compares aspects of pair trading across pair formation and trading stages. Multivariate trading pairs are built by employing the OPTICS algorithm, empirical study reveals that diversification characteristics of multivariate pairs enhance risk-adjusted returns, but they may not necessarily outperform univariate pairs in dynamic market conditions. Furthermore, a forecasting-based trading framework is proposed to capture short-term trend reversals and improve risk-adjusted returns. Among the forecasting methods employing the ARIMA, XGBoost, and LSTM models, the ARIMA models demonstrate the best forecasting and trading performance, achieving a notable 23.038% annualized return, a 5.612 Sharpe ratio, and an over 80% win rate throughout the testing period.