I’ve been blogging about why you need a minimum amount of data, including recent data, to generate a valid FICO® Score. An estimated 15% of US consumer credit files don't meet our "minimum score criteria," and as a result, millions of consumers currently don't receive FICO Scores.
The idea of scoring more consumers is enticing. Lenders could reduce underwriting costs for consumers who would otherwise require manual underwriting. These savings can be passed on to consumers in the form of a lower cost of credit or greater availability of credit.
An obvious way to score more consumers is to simply eliminate or reduce the minimum score criteria. But in my last post, I discussed the dangers of doing so without deliberation and analytic rigor. Before the scoring algorithm can safely be expanded to cover currently unscorable consumers, three critical questions must be answered:
- Is the available credit information sufficiently predictive of a person’s repayment risk?
- Is there sufficient credit repayment history to predict future repayment behavior?
- Is the odds-to-score relationship appropriately aligned? In other words, does a 700 score represent the same risk level for both potential newly scorable and traditionally scorable groups?
We recently conducted research into each of these questions, and over the next few weeks, I’ll be sharing our findings on this blog.
I'll start with the first question, which essentially asks whether there is enough data available to calculate a robust and reliable score. To make that determination, consider the pie chart below. It shows the five categories of predictive credit information that are used to calculate a FICO® Score.
Some unscorable consumers have information in as few as one of these five categories, which amounts to very limited data. Our research confirmed that for consumers with such limited data available, the performance of a score built on that data was very weak.
Take, for example, consumers with one or more collections or adverse public records on file, but no account history. The Gini index of the score for these consumers was 0.147 – far less than the 0.600 to 0.800 range seen with traditionally scorable consumers. (A Gini index is a statistical measurement of predictive model effectiveness; the higher the number, the stronger the model. Gini indices range from 0 to 1, with 1 representing perfect risk discrimination.)
This research reinforces how critical it is to have enough credit data that's predictive of future repayment risk. A credit score built using limited data—for instance, using solely collections, public records or inquiries, and without active account history—could be very misleading or even downright inaccurate. Rather than indiscriminately scoring more records, we are committed to delivering a meaningful score that can facilitate a safe and responsible lending decision.
In my next couple posts, I’ll share our research findings with respect to the second and third questions above. And as a thinly veiled shameless plug, we are releasing an in-depth Insights white paper on this topic next week. I encourage you to subscribe to the Insights series so you are emailed a copy, or stay tuned to this blog for more details.