Consumer lending is dominated by automation and analytics. Lenders approve billions of dollars in consumer loans a year, relying partly--and, in many cases, fully--on segmentation, generic credit scores, custom credit scores, and other auto-decisioning tools. For example, the majority of credit card applications at major U.S. financial services firms are decided without a manual underwriter.
This article answers the following questions:
* How can credit review add value in examining the automated underwriting process in credit card, home equity, and auto lending?
* How can credit review use data analytics to both supplement and target manual transaction testing?
* How is the testing of origination scorecards by credit review different from the model validation work of a model risk management group?
How Credit-Scoring Models Are Developed and Used in the Automated Environment
A credit review professional needs to understand the objective of the custom score. Origination custom credit scores are designed to rank-order credit applicants based on a predefined metric, such as the probability the loan will become 60 or more days past due within its first 24 months.
Once the score has been developed, the scorecard will progress through the validation and monitoring process, which includes a statistical validation prior to implementation. This validation will include, for example, how statistically powerful the custom score is by using the Kolmogorov-Smirnov test for goodness-of-fit.
Understanding the objective of the scorecard and the monitoring of its performance allows you to better determine if it is being used as intended.
At most major financial institutions, consumer loan underwriting involves an automated process in which applications flow through a detailed decision engine. A rule-based underwriting tool, the decisioning engine usually contains these multiple steps:
1. Segmentation / dual matrix.
3. Credit policy.
4. Credit limit assignment.
7. Final decision.
Segmentation / Dual Matrix
Segmentation by economic profitability (EP) is traditionally the first step of any decisioning engine. Its objective is to split loans into similarly profitable segments. Potential customers are split by certain characteristics, such as the length or depth of their bank relationship and by the amount of revolving debt that would be correlated with higher or lower EP
After flowing through a decision tree, the potential customer is assigned to a predefined bucket containing similarly profitable applicants. Each bucket (or node, as it is called in a modeling environment) has a scorecard assigned to it. These scorecards are at the heart of the decisioning engine. Usually, the scorecards are a dual matrix with a FICO score on one axis and a custom origination credit score on another axis. In cases where different nodes have very similar profitability values, nodes often are aggregated into segments and the scorecard is examined at the segment level.
In the simplified example below, the decisioning engine segments applicants based only on the size of their deposit relationship with the bank. In the example, deposit relationships with the bank were found to have been an attribute most correlated to economic profitability. Consequently, when the decisioning engine was designed, this deposit size attribute was selected for use in segmenting customers.
In the example, if the applicant has a deposit relationship equal to or greater than $10,000, he or she would flow to a more "lenient" dual matrix since the bank would earn a higher economic profit for the loan. Similarly, if the applicant has no deposit relationship or a deposit relationship less than $10,000, the applicant would flow to a more "strict" dual matrix. …