Software

by BrainMine's President Clemens van Brunschot.

BrainMine has released a new component of its Scorecard Builder, a package that offers low-priced data preprocessing and model building processes with a GUI shell, as add-ons for the widely used SAS System.

The Reject Inference module is meant for inferring a 0/1 target variable for cases where this variable is missing because the cases were not accepted or not selected in the past. Such a procedure is necessary if the aim is to build a predictive model on the basis of previous accepts AND REJECTS. This is common practice in e.g. credit scoring. But the idea can be extended to marketing and pharmaceutical modeling where many cases were not exposed to the stimulus before.

The module has two modes: an automatic and a manual one. In automatic mode the user picks the predictor variables per measurement level, and declares the values that should be considered as 'missing'. The program will automatically develop a (binned) logistic regression model based on the accepted cases, and apply it to the rejects, making necessary adjustments.

One of the controls is a user-defined parameter that indicates the user's trust in the previous selection process. There are other parameters that can be tuned, like the size of the buckets used for evaluation. And the process can be applied to selected cases in the dataset.

In manual mode the user makes use of a model that has already been developed. In both modes the reject inference is evaluated in a table and a graph that show the score distribution of both known and total goods and bads. Screenshots of the interface and output can be found at www.brainmine.nl, where a free but fully functional trial copy can be requested.

Together with recent additions to the Model Evaluation module this addition has now turned the Brainmine Scorecard Builder into a complete but inexpensive package for developing and testing scoring models. It facilitates optimal binning of predictor variables of all measurement levels, allows interactions between predictor variables, and produces SAS datastep code for calculating the score on the basis of the original variables, respecting missing values.