Inductis dedicated an offshore Risk Analytics team for the development of the Risk Scorecards in multiple phases.
The team produced sophisticated hybrids of logistic regression models (SAS) with various non-parametric tools [CART and MARS] to address two business needs: an existing customer model using internal data and a new customer model using personal guarantor consumer data.
Secondarily, we blended the model scores with commercial scores to assess possible enhancement of the precision of the Acceptance and Rejection regions for approval decision making.
To adjust for potential bias due to the limitation of complete data to approved accounts only, we used Reject Inferencing techniques to ensure that model parameters would represent the behavior of all applicants and that the acceptance/rejection cut-offs would be appropriate across all applications.
The advanced models enabled us to capture both additive effects and interaction effects effectively and therefore deliver robust and interpretable solutions to the client.
We tested the predictive power and robustness of the models through multiple tests including out of time and out of sample validation, variable sensitivity analysis (see example below), coefficient blasting (see below) and bootstrapping, in addition to evaluating model performance on statistical metrics such as the KS Statistic, Lift, Concordance, the Confusion Matrix etc.
Illustrations of Techniques used for Assessing Robustness 
We further assessed the optimal operating equilibrium which maximized the net benefits for the client through financial impact analysis of alternate cut-offs for the acceptance region as shown below.
The client received detailed documentation to guide the implementation of these models in their production environment.
|
