Methods for outlier detection in prediction

Abstract If a prediction sample is different from the calibration samples, it can be considered as an outlier in prediction. In this work, two techniques, the use of the uncertainty estimation and convex hull method are studied to detect such prediction outliers. Classical techniques (Mahalanobis distance and X-residuals), potential functions and robust techniques are used for comparison. It is concluded that the combination of the convex hull and the uncertainty estimation offers a practical way for detecting outliers in prediction. By adding the potential function method, inliers can also be detected.