Market Regime Detection via Realized Covariances: A Comparison between Unsupervised Learning and Nonlinear Models

Abstract

There is broad empirical evidence of regime switching in financial markets. The transition between different market regimes is mirrored in correlation matrices, whose time-varying coefficients usually jump higher in highly volatile regimes, leading to the failure of common diversification methods. In this article, we aim to identify market regimes from covariance matrices and detect transitions towards highly volatile regimes, hence improving tail-risk hedging. Starting from the time series of fractionally differentiated sentiment-like future values, two models are applied on monthly realized covariance matrices to detect market regimes. Specifically, the regime detection is implemented via vector logistic smooth transition autoregressive model (VLSTAR) and through an unsupervised learning methodology, the agglomerative hierarchical clustering. Since market regime switches are unobservable processes that describe the latent change of market behaviour, the ability of correctly detecting market regimes is validated in two ways: firstly, randomly generated data are used to assess a correct classification when regimes are known; secondly, a na\"{i}ve trading strategy filtered with the detected regime switches is used to understand whether an improvement is showed when accounting for regime switches. The results point to the VLSTAR as the best performing model for labelling market regimes.