Comments on : “ Model Complexity Control for Regression Using VC

In [1], various model selection approaches were experimentally inter-compared; one of the considered model selection criteria was the Schwarz Information Criterion (SIC); however, SIC was incorrectly implemented. The same mistake has been repeated in other more recent papers. Here, we show why the SIC formula originally employed was wrong. We report instead the correct approach, which is well-known in statistics literature. We then show that the SIC performance is far better than the one described in [1], by repeating several experiments of the original paper. Nevertheless, we confirm that VC-based model selection is more powerful than SIC, especially for small samples.