5 minutes
Credit unions have rich data that would be uniquely valuable in evaluating credit risk: their historic lending results, such as which members paid in full, which loans were charged off, and what their balances were when they were charged off.
Credit union borrower history, distinct from individual members’ credit reputation with reporting agencies, could be the most accurate predictor of future repayment. But CUs may not have considered how their data could be leveraged as a decisioning tool and, if they have, they may not have had the means or expertise to access and activate their history in this way.
Unique Markets Need Unique Solutions
We talk a lot about the differences between credit unions and banks, but what about the differences between credit unions and other credit unions? CUs share a philosophy and organizational structure that differentiate them from banks, but when you consider mission, field of membership and location, CUs’ truly individual natures show. It’s revealed in different product and service offerings, and it shows up in differences in member behavior, including repayment behavior.
Consider Communicating Arts Credit Union, a $30 million community development CU in Detroit (with whom I am a volunteer director). Communicating Arts CU is working to offer credible financial services to its members, many of whom have been underserved and/or unbanked, and some of whom have routinely relied on check cashers and payday lenders.
CUES member Hank Hubbard, our CEO, says “We’re always trying to get to yes. We don’t want to make bad loans, but it is our strategic direction to walk close enough to the line with our approvals that we are putting loans in the hands of any member we think will perform. We realize this strategy means that we have to expect significantly higher-than-peer charge-offs, and we plan our business accordingly.”
In all, Communicating Arts CU knows it is approving applications with metrics (income, credit scores, debt-to-income, etc.) that even other mission-driven CUs wouldn’t touch.
“Not all 500 credit scores are the same,” Hubbard adds. “Once you get below a normal subprime score, it’s an automatic ‘no’ for many banks and some credit unions. We want to say ‘yes’ much deeper into the scores, so the trick is to use other attributes to make a credible decision—to use our experience to say ‘yes’ to the right loans. This can be a very labor-intensive process, and it’s our hope that predictive modeling will allow us to get a quick and consistent decision on those gray-area applications.”
Individualized Predictive Modeling
Most credit unions have been using predictive modeling for decades, in the form of credit scores, as a consideration in their lending decisions. These models are not specific to any particular institution but, rather, a one-size-fits-all-borrowers algorithm intended to yield a measure of creditworthiness. Because credit scores are such broad measures, CUs have had to individualize their use of these scores by drawing their own lines at levels of acceptability for approval, and by weighing these scores in relation to other known borrower metrics.
It’s not surprising that a one-size model will be more effective for some CUs, and less effective for others. While some common metrics may be predictive of borrower performance in all credit unions (e.g. debt-to-income, credit utilization), the weight of these metrics' predictive value in loan outcomes will differ. Further, a metric that at first blush appears irrelevant as a predictor could prove valuable when properly weighted, combined, and used in conjunction with metrics that are more often considered predictive. For instance, if a member’s home branch is “A” and the member works part time, the loan-to-income ratio could be weighted differently by one CU's well designed model than if the member’s home branch is “B” and he works full time.
The potential for combining a wider variety of predictors and varying the weights they are given is what can be most exciting about using individualized predictive models. By looking at the repayment behavior of a very specific subset of borrowers—one credit union’s members—the effects of hyper-local economic factors, community loyalty, and institutional loyalty on borrower repayment may be revealed.
How Much Data Is Big Data?
Many credit unions may think they are too small to get tangible value from individualized predictive modeling. They may think the historic data they have is too small to qualify as “big data.” Sure, some credit unions may indeed have too little data to produce a reliable model, but most CUs have more than enough.
“For lenders with relatively focused demographics and loan types, a few hundred loan samples spaced over the period of two or three years may be sufficient to build a useful model,” says Justin Washtell, Ph.D., chief technology officer and co-founder of ForecastThis, which sells automated predictive modeling solutions. “We ask that they provide data from more than one year so that we can detect possible seasonal and time-varying patterns.
“As part of the cross-validation process, we thoroughly measure the variability in the data and can infer from this whether and to what extent more data would improve the predictive quality of the model,” whether an effective model can be built with the available data, and to what extent additional data would be likely to improve the model.
Credit unions should not allow what may be their most effective means of predicting borrower performance lie dormant in their core processing systems. It should be a priority for every CU to learn to access its own data, and to look for the tools to leverage those data.
Credit unions should embrace their uniqueness. It doesn’t take a big institution to have big data, and it no longer takes a big institution to afford an individualized decisioning solution. Advances in technology have democratized solutions, so they are equally available to billion dollar and $5 million credit unions alike.
Matt Barbour is chair of $30 million Communicating Arts Credit Union in Detroit, and credit union advisor to LendingLens, Detroit, a custom predictive modeling product of ForecastThis.
