Detecting Multivariate Market Regimes Via Clustering Algorithms

Abstract

In this paper we study the joint dynamics of multivariate time series using an unsupervised learning technique and demonstrate its use in pairs trading and portfolio design. We present a novel non-parametric market regime detection method for multidimensional data. The detection procedure is based on a k-means clustering algorithm which makes use of distances between distributions to discriminate between different market regimes, which are categorised by their mean, variance and correlation. In particular, we empirically investigate the performance of the algorithm endowed with either Wasserstein distances or Maximum Mean Discrepancies. We suggest a two-step approach to clustering multivariate data using this new method, which we show to be effective on both synthetic and real world data. We demonstrate how our new approach can be used to obtain approximations to the mean, variance and correlation between two assets at a given point in time. We further show that these values can be used in the context of Modern Portfolio Theory to form profitable trading strategies when using two assets.

Keywords: k-means clustering, p-Wasserstein, MMD, Mean-Variance optimisation, Stochastic processes, numerical methods, stochastic volatility models, regime classification, unsupervised learning

JEL Classification: C02, C14, C15, C45, C82, G6, G11