An important problem in higher education is to find the most suitable admission procedure that can distinguish between students with high academic potential and future dropouts. Admissions usually rely on pre-enrolment achievement measures; therefore, it is crucial that these selection criteria have high predictive validity on academic achievement. In this study, we use sophisticated statistical learning methods, such as receiver operating characteristic curve analysis, logistic and Tobit regression to analyse the predictive validity of the Hungarian university entrance score on final university performance, in particular on degree completion and qualification. We place particular emphasis on drawing statistically grounded conclusions. The analysis is built on data of 21,547 undergraduate students from the Budapest University of Technology and Economics. We find that the current Hungarian centralised entrance score is a valid predictor, however, its predictive validity varies significantly across disciplines. We find that high school grades have strong predictive validity, and general knowledge is more important than program-specific knowledge. We also find that the academic performance of females is underpredicted and that of the males is overpredicted by the university entrance score.