One of the most exciting things for us at Zest is finding new ways for our finance customers to use machine learning (ML) to boost revenue and reduce losses. Recently, one lender asked if it would be possible to use Zest Automated Machine Learning (ZAML) to examine an existing group of borrowers and, without selling these customers anything new, improve revenue coming in from that group.
Challenge accepted. Usually, our lending clients use ZAML to either increase the dollar amount of loans given while holding losses steady over time or to reduce losses while holding loan dollars steady over time. But our data scientists felt confident they could use ZAML modeling to help the client price more accurately for risk by better understanding its existing customers.
After all, ZAML uses thousands of variables to transparently improve the risk profile of new customers far better than traditionally built models can. So it should be better at evaluating customers a bank would be approving under any model. The bulk of the benefit from this type of model comes from swapping certain borrowers in and out of the approved applicant population. But among the borrowers that get approved by both models, some are swapped into higher and lower pricing tiers based on their credit risk. The net result is that borrowers get loan terms more closely aligned to their risk, and lenders get more fairly compensated for the risk they are taking.
To verify how this might work in a real world example, we tested this idea, by looking at customers who were both approved for loans under the bank’s old non-ML model and through ZAML.
With most loan products, lenders usually place their borrowers into different pricing levels to reflect how risky they are, like how holders of the same credit card may have different annual percentage rates (APRs) on their accounts. As this diagram of a four-tiered pricing plan shows (with the A tier being lowest risk), applying any new model to the same group of customers will result in some being assessed as less risky, others as riskier and some similar enough that they wouldn’t change pricing tiers.
In our experiment, the lender’s old model relied on a handful of traditional variables for assessing and sorting borrowers, including credit score and loan-to-value. The new ZAML model uses more than 3,000 variables to find unseen connections that indicate how likely someone is to repay their loan.
The results of our experiment showed some existing customers were more likely to repay their loans than previously indicated. Moving them into a less risky pricing tier meant those customers get a price cut — which is nice for them but doesn’t particularly help the bottom line of the lender.
But at the other end of the scale, ZAML found a large number of customers who were assessed to be more risky than the benchmark model deemed them to be. As a result, these customers should be moved to a higher pricing tier. In this lender’s example there were many more customers that should be moved into higher pricing tiers than customers moving down to lower ones. The net effect is that ZAML can increase portfolio yield – that is, a bank’s revenue – without any other change to the customer mix.
Now, Economics 101 tells you that any increase in price should result in a drop in demand, but the size of that drop depends on the elasticity, or price sensitivity. Academic research has shown that credit tends to have pretty inelastic demand, meaning most customers won’t change their behavior even in the face of increasing prices. There are myriad reasons why customers might be price insensitive but whatever the reason, the result is that improving a lender’s risk assessment of this population of customers approved by both models can result in real boosts to revenue.
How much of a boost can lenders see? Actual returns will vary depending on the specific characteristics of the loan portfolio, but our analysis suggests gains of $500,000 to $4 million are feasible. The table here reflects ZAML’s revenue boost from the specific dataset we tested on. These results are based on a set of assumptions that include average loan amount, the number and spread of APR tiers, and customer price sensitivity.
While this was just an experiment, the results suggest that since ZAML can quickly and efficiently reevaluate customers as variables change, lenders will eventually want to run dynamic, continuous risk modeling to enhance their loan portfolio yield. That’s one of the features built into ZAML.
Contact us at email@example.com for a full demo.