Model Entropy

The model entropy gives you an idea of how useful each variable could be in a predictive model for the probability of default.

The best possible predictor is one that, for each generated bin, contains cases with the same value as the guide variable; thus, the guide variable can be perfectly predicted. Such a predictor has an undefined model entropy. This generally does not occur in real-world situations and may indicate problems with the quality of your data.
The worst possible predictor is one that does no better than guessing; the value of its model entropy is dependent upon the data. In this dataset, 1256 (or 0.2512) of the 5000 total customers defaulted and 3744 (or 0.7488) did not; thus, the worst possible predictor would have a model entropy of −0.2512 × log₂(0.2512) − 0.7488 × log₂(0.7488) = 0.8132.

It is difficult to make a statement more conclusive than that variables with lower model entropy values should make better predictors, since what constitutes a good model entropy value is application and data-dependent. In this case, it appears that variables with a larger number of generated bins, relative to the number of distinct categories, have lower model entropy values. Further evaluation of these binning input variables as predictors should be performed using predictive modeling procedures, which have more extensive tools for variable selection.