Prediction of the Profitability of Pairs Trading Strategy Using Machine Learning

Pairs trading strategy is one of the well-known quantitative trading strategy developed in 1980s by the team of scientists. There are many researchers trying to study and create the mathematical model to improve the pairs trading strategy on various assets such as cointegration method, OLS, Kalmann filter, Machine learning, etc. The purpose of the models is to generate the precise signals from pairs of assets to maximize the return based on statistical arbitrage of pairs trading strategy. In this paper, Stress Indicator pairs trading strategy is studied further. Stress Indicator pairs trading strategy is easy, straightforward and profitable. However, There are many factors which influence the profitability of the strategy, causing the loss trades. We purpose a novel approach by using the machine learning algorithm to learn the historical trades of Stress Indicator pairs trading strategy in foreign exchage rates and to predict the profitability in the future trades. The pairs of the exchange rate are filtered by choosing only the pairs which generate the positive average return per trade from Stress Indicator pairs trading strategy in the past. The capability of the ML models is to classify whether the signals from Stress Indicator pairs trading strategy is profitable or not before opening the positions. The powerful ML models, Artificial Neural network and XGBoost, are implemented in this study. Several factors which could influence the profitability such as correlation, volatility OLS beta are collected and used to train the model following the common step of ML training procedures such as features selection, Hyperparameter tuning and k-Fold cross validation to generate the capable models. Next, the performance of ANN and XGBoost is compared that which one performs better by the score matrix. The result shows that the performance of predicting the profitability is not significantly different. Both models mostly achieve 60% accuracy in In-sample data, but the accuracy in out-of-sample data is quite fluctuated. In other words, ML models are capable to classify the profitable signal from price behavior but may lack of consistency.