We address the problem of outlier detection for more reliable credit scoring. Scoring models are used to estimate the probability of loan default based on the customer’s application. To get an unbiased estimation of the model parameters one must select a set of informative objects (customers). We propose an object selection algorithm based on analysis of the covariance matrix for the estimated parameters of the model. To detect outliers we introduce a new quality function called specificity measure. For common practical case of ill-conditioned covariance matrix we suggest an empirical approximation of specificity. We illustrate the algorithm with eight benchmark datasets from the UCI machine learning repository and several artificial datasets. Computational experiments show statistical significance of the classification quality improvement for all considered datasets. The method is compared with four other widely used methods of outlier detection: deviance, Pearson and Bayesian residuals and gamma plots. Suggested method performs generally better for both clustered and non-clustered outliers. The method shows acceptable outlier discrimination for datasets that contain up to 30–40% of outliers.