Classification of the optimal rebalancing frequency for pairs trading using machine learning techniques

Abstract

Selection of the optimal rebalancing frequency (ORF) is crucial for the pair trading algorithm (PTA) that periodically rebalances the allocation of two assets. This study proposes a machine learning (ML) approach to predict ORF ranges. To improve ML accuracy, pairs were categorized into three subgroups based on their correlation levels: positively, weakly, and negatively correlated. The statistical distribution of the ORF values is also presented. Accuracy scores show that random forest, logistic regression, and support vector classifiers perform competitively for the ORF range classification in both short- and long-term applications. The negatively correlated pairs showed the best classification performance, whereas the positively correlated pairs showed the lowest accuracy rate. Furthermore, the robustness of the proposed ML procedure is verified using a validation dataset, demonstrating the applicability of ORF range classification in practical exchange markets.