Abstract (EN)

A simple and efficient exogenous outlier detection & estimation algorithm introduced in a regularized version of the Kalman filter is presented. Exogenous outliers that may occur in the observations are considered as an additional stochastic impulse process in the observation equation of the Kalman filter that requires a regularization of the innovation in the recursive equations of the Kalman filter. Regularizing with a l1 or l2-norm needs to determine the value of the regularization parameter. Since the innovation error of the KF is assumed to be Gaussian we propose to first detect the possible occurrence of a non-Gaussian spike and then to estimate its amplitude using an adapted value of the regularization parameter. The algorithm is first validated on synthetic data and then applied to a concrete financial case that deals with the analysis of hedge fund returns. We show that the proposed algorithm can detect anomalies frequently observed in hedge returns such as illiquidity issues.