Using Statistical Models to Counter Consumer Correlations Conundrum. (Consumer Lending)
Mays, Elizabeth, The RMA Journal
Statistical models aren't just for credit scoring anymore. Institutions are applying the same analytical technique to credit exposures from certain loan products or portfolio segments, realizing it's the necessary antecedent to appropriate lending policies and risk-based pricing adjusters.
Traditional credit risk analysis relies on tracking delinquency or charge-off rates by portfolio segment to evaluate how individual loan characteristics affect delinquency and default. Although a helpful first step in understanding portfolio risk, such tracking is by no means sufficient and, in certain cases, can be very misleading. It can even create lending policies that result in money-losing loans or lost opportunities for profitable loans.
Correlations Cause Confusion
Because the risk characteristics of consumer loans are often correlated with one another, it makes sense to measure credit risk with statistical models. Frequently, borrowers who exhibit a "risky" level of one characteristic, such as debt-to-income ratio, may also have other high-risk characteristics, such as poor credit histories. Some correlations among characteristics occur naturally, but many are created by past credit policies.
For example, mortgage lenders historically have set loan-to-value (LTV) ratio limits for loans on non-owner-occupied (investor) homes that are lower than those for owner-occupied homes. Thus, in an attempt to manage risk, they have built in a correlation between the risk characteristic "investor property" and LTV. If we observe that the delinquency history of investor properties is lower than that for primary homes, we do not know if such homes do, indeed, carry lower risk or if we are observing the effect of their low LTVs. Even if the delinquency rates of investor homes are higher than for owner-occupied homes, we cannot know how much additional risk is imparted by its being an investor property. That is because we are observing the joint effect of investor property and low LTV ratio, as well as any other characteristics correlated with occupancy type.
Sorting It Out
To disentangle the effects of numerous risk characteristics on loan performance and learn the incremental effect of each, we return to scorecards. Multivariate regression models let us assess the effect of multiple loan characteristics, or variables (hence, the term "multivariate"), on loan performance and identify the separate influence of each. Once the incremental risk of each characteristic is known, appropriate policies can be designed to limit the risk, or appropriate risk-based pricing adjusters can be charged to compensate the lender for the extra risk.
How Regression Models Work
Two cases presented below demonstrate the weakness of delinquency rate analysis, whereas statistical models reveal the characteristics' true incremental risk. Both examples use actual data from a large consumer lender.
Home loan. Let's examine the delinquency history of mortgages made for cash-out refinances where the borrower is refinancing his loan and receives a portion of his equity as cash back from the lender. Borrowers with a lot of high-interest debt outstanding and who wish to consolidate their debt into a single lower-interest loan will frequently favor this type of loan. The table below compares the percent of cash-out refinances that have gone 90 days past due (DPD) in a two-year period to the percent of for-purchase mortgages that have reached the 90 DPD mark.
The table shows that cash-out refinance loans have been seriously delinquent at a rate that is only 58% (.78/1.35) that of purchase-money mortgages. What this simple analysis ignores, however, is that loan type is correlated with other borrower and loan characteristics. In fact, as for investor homes, the lender has limited the LTV ratios on cash-out refinances to lower levels than those for purchase-money mortgages.
Multiple regression analysis quantifies the effect of a set of variables on another variable we are interested in explaining or predicting. …