The MSA data set was divided into 10,000 different learning and test sets of equal size. The value of α which produces the maximal median percent of AUC improvement in accuracy of contact prediction was obtained for each learning set and then used to assess the median percent of AUC improvement for the corresponding test set. The distributions of the median percent of improvement obtained for the test sets are shown for OMES, MI, DCA and PSICOV. The mean values of these test sets distributions obtained using the different methods are similar to those of their learning sets, thus, showing that the improvement is not due to over-fitting. In the case of OMES, MI and DCA, the mean difference between the values of α which maximizes the median of the percent of AUC improvement for the learning and test sets equals zero, thus, reflecting the stability of the values of α. In the case of PSICOV, the variance of that difference is high due to the asymptotic nature of the median percent of AUC improvement as a function of α.