Robust Detection of Multivariate Outliers in Asset Returns and Risk Factors Data

Abstract

It is well-known that outliers exist in the type of multivariate data used by financial practitioners for portfolio construction and risk management. Typically, outliers are addressed prior to model fitting by applying some combination of trimming and/or Winsorization to each individual variable. This approach often fails to detect and/or mitigate multivariate outliers. Existing literature documents the use of the robust Mahalanobis squared distance (RSD) based on the minimum covariance determinant (MCD) estimator to detect and to shrink multivariate outliers in financial data. We use MCD-based RSDs, along with a modified version of the Iterated Reweighted MCD methodology of Cerioli, to illustrate the presence of outliers in the asset returns and firm fundamental data that equity portfolio managers would use to build and monitor portfolios. We demonstrate how RSDs based on the MCD estimate are superior to Mahalanobis distances based on the classical mean and covariance estimates for detecting multivariate outliers. In the process, we show that univariate trimming and Winsorization are insufficient to deal with multivariate outliers in financial data.