CMStatistics 2020: Start Registration
View Submission - CFE
Title: Market regime detection via realized covariances: A comparison between unsupervised learning and nonlinear models Authors:  Vito Ciciretti - Independent (Germany) [presenting]
Andrea Bucci - universita degli studi g d annunzio di chieti pescara (Italy)
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 stressed regimes, leading to the failure of common diversification methods. We aim to identify market regimes from correlation matrix features and detect transitions towards stressed regimes, hence improving tail-risk hedging. Starting from the time series of fractionally differentiated sentiment-like future values (such as gold, VIX, dollar index, etc.), we first build the realized correlation matrix for each period. Hence, we calculate several correlation features such as the distribution of the eigenvalues, the cophenetic index, the condition number. On these features adjusted to deal with collinearity, we apply an unsupervised learning methodology, the agglomerative hierarchical clustering, for labelling two latent market regimes, and compare its regime identification accuracy with a smooth transition autoregressive model applied on the covariance series. Finally, we fit a random forest classifier and evaluate its SHAP values to understand the most important correlation features when filtering market regimes. Our results show that the stressed regime is easier to be classified and that the cophenetic index and the percentage of variance explained by the eigenvalues above the Marchenko-Pastur upper bound are the most relevant explaining variables.