top of page
  • Writer's pictureRick Haskell

Data Mining Based on Cumulative Net Loss (CNL)—the Lending Executive’s Dream Tool

Updated: Jul 10

Ask 10 different managers for their “go to” variable when assessing borrower risk, and you’re likely to get 10 different answers. Having leadership duties over risk management for well over a decade, I came to learn that my most important, and often most difficult task, was to educate my colleagues regarding “true” predictive power in the variables they work with every day.

As they say, the numbers don’t lie; and the true way to know which variables are most predictive is to run a statistical characteristic analysis. It can be quite humbling to learn that your favorite predictor doesn’t even fall in the top 100. In fact, folks can start feeling threatened if this isn’t handled with care.

Short of statistical analysis, most people use intuition and experience when assessing risk in this way—and that’s a good thing, but it could be better.

Flipping a coin gets a 50/50 result. Use your experience and good judgment and you may increase predictive accuracy to well over 60%. Use statistical methods and now you’re up over 70%—possibly 80%. These estimates are just generalities, but you get the idea. Optimization over decision-making has taken over just about every industry out there, and for good reason—companies are getting more done, more accurately, and with fewer staff!

Sadly, performing this sort of analysis doesn’t come through intuition and experience. Statisticians have developed procedures for these sorts of predictions; but virtually all are unit-based in their root algorithms. But the financial services industry deals with dollars—not units, so we need a better solution.

Lendisoft has developed a data mining algorithm built around Cumulative Net Dollars (CNLs). CNLs are the principal dollars lost due to defaults in a loan portfolio. As a risk manager at an enterprise-class auto finance company, I would routinely run CNL-based performance reports across hundreds of portfolio segments. Eventually I created a binning algorithm that replicates the unit-based algorithms commonly used in ML model building. In time, this algorithm was refined and matured in its use and effectiveness.

Today, this algorithm can spin through hundreds of variables in a matter of minutes and produces output that breaks down every variable into its best and worst performing continuous ranges. For categorical variables it breaks down every value variation and groups them according to similar CNL performance—producing priceless learning for your entire management team.

Included are interactive dashboards where all of these results are summarized into views. One view gives you the Top 100 breakdown, with each variable showing its precise power statistic. This is perfect for rank-ordering your most predictive variables, but also for comparing precise predictive power between variables—with everything based on CNL performance as its basis.

Finally, we provide a campaign builder view where you can group together multiple variables of your choice and see instantly how many loans it constitutes (units and dollars), and the forecasted CNL performance for that population. This is a great way that business leaders can quickly experiment with campaign building for future loan programs, marketing promotions, or anything else you can imagine. Who needs a risk team anymore—you now have the power at your fingertips! For a demo of our no-code Data Mining tool, contact Lendisoft today!

bottom of page