Outliers in Asset Pricing

This paper examines the effects of multivariate outliers on factor models of the stochastic discountfactor (SDF) through expected returns estimates. To demonstrate these effects clearly within a machine learning context we use the estimation method of Kozak et al. (2020) and an outlier robust method based on the Minimum Covariance Determinant (MCD) estimator. Using a closed-form solution of the SDF coefficients we provide intuition on the effects of outliers on estimation. Furthermore, we show the economic interpretation of the MCD estimator as an optimal decision rule under ambiguity on the proportion of outliers in the data. We prove the method shrinks the covariance matrix and increases the upper bound on the prior Sharpe Ratio. On factor returns data, the robust method significantly outperforms the classical method in terms of out-of-sample R^2. Furthermore, the robust method is able to identify three regimes which correspond to a low volatility, high volatility and market crash state.The effects of outliers on expected return and SDF coefficient estimates are in line with the intuition.The effects are especially strong for factors with strongly market state-dependent returns like momentumfactors