Go to main content

Interested in SA updates and events?



More Information»

Download PDF for Printing [252k]

More Information
+1 505-995-4720
Email

Watch Out When Using Credit Scores to Drive a Forecast!
Signals, Spring 2003

Most credit and behavior scores are ordinal (or rank order) measures of behavior. In other words, a score is simply a relative indicator of the likelihood of a future outcome. This is a powerful feature from the standpoint of score stability (one can feel confident that customers scoring 720 will always have less risk on average than those scoring 620 regardless of the environment), but a score does not by itself indicate the likelihood or magnitude of an event.The industry has relied upon "score-odds calibration" to align scores with their observed (historical) probability of occurrence, and this is one area where portfolio managers should exercise care and examine all assumptions in order to avoid surprises.

Figure 1

For instance, most credit scores are built from "ever-60+" (the odds that an account will become 60+ days past due) delinquency data measured over the past 18-24 months. That fact and sound engineering of the component parts will give it good stability but the period evaluated is often too short to encompass important lifecycle events. Thus a measure such as "ever-60+" potentially masks important segment behavior that can make a large impact on a growing portfolio, and won't say much at all about how the losses will come in over an extended period of time, or how such loss patterns will vary across groups of customers.

Figure 1 illustrates a potential distortion of this kind. Both Group A and Group B segments have the same average credit score—after 24 months of experience, both groups will register the same total level of delinquency. Yet, the portfolio implications of booking one group over the other could be striking. For example, if you choose to originate the majority of your new accounts among Group A, you had better understand the potential for early losses associated with this strategy. This example also has ramifications for pricing, since on a net present value basis Group B customers are more valuable if you look beyond 24 months. If you were relying on credit score alone as a proxy for loss (especially timing of loss) you could also be in for a surprise. More fundamentally, any sound forecast of the aggregate portfolio would have to recognize the fundamentally different loss patterns of the group, and ideally how those loss patterns might be affected by environmental or economic shifts that most certainly will occur.
Top of page