Pairs Trading via Unsupervised Learning on the JSE

Abstract

Pairs trading, a strategy that capitalises on temporary price discrepancies between two correlated assets, has garnered attention for its potential to generate profits in financial markets. This research explores the viability of employing unsupervised learning techniques for pairs trading on the Johannesburg Stock Exchange (JSE). Using clustering algorithms to identify pairs and considering both price data and firm characteristics, the study examines the performance of pairs trading portfolios constructed via different clustering methods. Empirical results reveal that while agglomerative clustering shows promise with the highest monthly mean return for long-short portfolios, none of the strategies consistently outperform benchmark indices. Furthermore, considering only momentum features in the clustering process leads to deteriorated portfolio performance, emphasizing the importance of incorporating firm characteristics. Despite the potential benefits offered by unsupervised learning, challenges such as the limited number of listed stocks and algorithm selection hinder the strategy's effectiveness on the JSE. The findings suggest that further research is needed to refine methodologies and address practical implementation challenges for pairs trading strategies in emerging markets like the JSE