A major criticism on the binning algorithm as well as on the WoE transformation is that the use of binned predictors will decrease the model predictive power due to the loss of data granularity after the WoE transformation. While talk is cheap, I would use the example below to show that using the monotonic binning algorithm to pre-process predictors in a GRNN is actually able to alleviate the over-fitting and to improve the prediction accuracy for the hold-out sample.

First of all, the whole dataset was split into half, e.g. one as the training sample and another as the hold-out sample. The smoothing parameter, e.g. sigma, was chosen through the random search and happened to be 2.198381 for both GRNNs.

1. For the first GRNN with untransformed raw predictors, the AUC for the training sample is 0.69 and the AUC for the hold-out sample is 0.66.
2. For the second GRNN with WoE-transformed predictors, the AUC for the training sample is 0.72 and the AUC for the hold-out sample is 0.69.

In this particular example, it is clearly shown that there is roughly a 4% – 5% improvement in the AUC statistic for both training and hold-out samples through the use of monotonic binning and WoE transformations.