The paper proposes an algorithm for estimating the performance of classification models under covariate shift and shows that it outperforms existing methods on multiple real-world datasets.