Extending a machine learning pairs trading strategy

Abstract

One of the most popular sources of alpha (trading strategy) used by asset managers is called pairs trading. It exploits the idea that two co-moving assets exposed to the same risk factor should keep moving together, relative to that common factor. Any divergence from this relationship can be exploited by buying the undervalued security and selling the overvalued one. In the literature there have been many works contributing different methods attempting at increasing it’s profitability but overall it’s fragmented. This fragmentation is due to the lack of gold standards, which leads to discrepancies in methodologies that make comparisons impossible. Building on recent attempts to unify all the different methods into one modular assembly line type framework, with machine learning methods applied at each station, we seek to answer a key question, which methods maximize financial performance at each station of the pairs trading framework? Thus a summary of alternative techniques deployed in the asset representation and pairs search stage are discussed. These methods are be implemented and evaluated in this work. We manage to identify K-Means as the clustering technique that provides the best financial performance. In the explanatory analysis we also find that most clustering algorithms produce results that correlates with the current state of the market. That is, if the overall market is doing well then the performance from the clustering method will also do well and vice versa. We also found interesting edge cases. The Agglomerative clustering using Complete Linkage performs better under certain factors, in this case in a stressed Minimum Spanning Tree (MST), via the feature describing the normalized tree length. We also manage to identify improved financial performance when Autoencoder based feature representations are used over PCA based ones. Furthermore we observe a relationship between the entropy of the frequency distribution of the generated representations with the final financial performance. These findings will need to be further investigated, but they are steps forward in our understanding of these methods applied to this pairs trading challenge.