Adapting credit scores to evolving consumer behavior and data.

• Author: Huynh, Frederic
• Position: Symposium: Credit Reporting and Credit Scoring

1. INTRODUCTION

   Credit scores are a vital element of the lending ecosystem, providing lenders with an objective means to assess a consumer's creditworthiness. Broad-based credit scores, such as the FICO Score, are redeveloped periodically to capture changes in consumer risk patterns, leverage improvements in the reporting of information to the Credit Reporting Agencies (CRAs), and incorporate new technological enhancements into the score algorithm. (1)

   To demonstrate how consumer risk patterns have changed over time, we might consider the number of credit cards a consumer has. The earliest incarnations of the FICO Score typically incorporated a predictive variable that measured the number of credit cards a consumer possessed. In the nascent days of FICO Scores, having many credit cards was risky and having few credit cards actually represented good risk. Over time, the risk pattern associated with that variable fundamentally changed.

   In Figure 1, we observe a noticeable change in the number of credit cards and associated risk. In 1992, having many credit cards indicated a high level of credit risk; in fact, a consumer with eighteen or more credit cards was twice as risky as the total population. In 1998, consumers with many credit cards were actually slightly better credit risks than the total population. In both time periods, having no credit cards indicated a greater degree of credit risk, but elsewhere the risk pattern fundamentally changed. While most predictive characteristics employed by credit scores demonstrate more stable risk patterns, this is one example that demonstrates the benefit of redeveloping scores periodically in order to account for changing risk patterns.

   In addition to redeveloping the scores to accommodate for changing risk patterns, the building blocks of the score--the characteristics and the treatment of the primitive data elements--should also evolve. This Article presents three different examples of how the predictive characteristics of the score evolved to adapt for changes in the credit landscape. The first study focuses on enhancements introduced to an earlier generation of the FICO Score, released in the early 2000s. The second study focuses on enhancements that were incorporated into the FICO 8 models. The last study focuses on research that determines whether short sales and other codes signifying different types of mortgage stress events should be treated less harshly by the FICO Score.

2. CHANGING INQUIRY ASSESSMENT TO MORE EFFECTIVELY ACCOMMODATE RATE-SHOPPING

   Credit inquiries have long served as predictors in credit-scoring models. Although they contribute a relatively small percentage to the predictiveness of the final score, credit inquiries often attract a disproportionate amount of attention from consumers. This presumably has been due to the prominence they receive on credit reports. Since the mid-1990s, FICO Scores have employed logic in the treatment of inquiries that recognizes the presence of rate-shopping behavior. There are two components of this logic, a buffer and a deduplication window. The purpose of the buffer is to bypass any auto or mortgage inquiries made within the last thirty days. This prevents very recent auto or mortgage inquiries from influencing any current applications for credit. The deduplication window is a rolling timeframe in which multiple auto or mortgage inquiries, posted to the credit report during the deduplication window, will be counted as a single inquiry.

   The following table illustrates the general concepts behind the earlier inquiry logic. In this example, all of the auto inquiries occur within the last thirty days and are ignored. The two mortgage inquiries fall outside of the thirty-day buffer, and are eligible to be counted. However, since both mortgage inquiries fall within fourteen days of each other, only one inquiry will be counted. Even though the department-store inquiry occurs within fourteen days of the first mortgage inquiry, it is counted separately; only auto and mortgage inquiries are deduplicated. In this example, an earlier version of the FICO Score would count two inquiries.

   With the beginning of consumer score disclosure in 2000, the launch of MyFICO.com in the early 2000s, and the increase in financial advice to consumers through news media and the Internet, consumers became more aware of the benefits of shopping for the best rate. At the same time, lenders began to offer a wider array of credit products enabled by risk-based pricing. (2) As consumers became more financially savvy, their search for the best interest rates on a mortgage or auto loan often took longer than fourteen days. FICO suspected that consumers who were attempting to find the best rate could be penalized and that the inquiry logic could be improved upon. The company's scientists revisited the model's inquiry logic to determine if a fourteen-day deduplication window remained ideal for risk prediction. If the fourteen-day window was too short, too many inquiries were being counted, excessively penalizing the consumer and yielding slightly fewer predictive characteristics.

   To investigate the merits of broadening the deduplication window, FICO varied the length of the window and measured the resulting impact to predictiveness. In general, a longer deduplication window was proven to be more effective in evaluating inquiry information. Information value--a statistic that measures how well a given characteristic separates goods from bads--was used to determine if a change in the deduplication window was merited. (3)

   Figures 3 and 4 represent internal FICO research that assesses the analytic merit of different deduplication windows. They demonstrate that by expanding the length of the deduplication window, the characteristic is marginally better at predicting risk. Figure 3 is based on the performance of all credit accounts on the consumer's file. Figure 4 is based on the performance of new accounts. Inquiries can be more relevant in an originations context, so the same analysis was repeated based on the performance of new accounts--accounts opened within the six months following the scoring date. The patterns are fairly consistent in both contexts.

   For the total population, increasing the deduplication window leads to a slightly stronger characteristic. However, we found there is an upper limit to the length of the deduplication window. The improvement gained in using a sixty- or ninety-day deduplication window is marginal. For the clean population (roughly 70% of all consumers), the ideal deduplication window is approximately forty-five days when looking at performance metrics for both all accounts and new accounts separately. Interestingly, we see that the ideal deduplication window for the rest of the population--those who have at least one derogatory event on their credit history--may be greater than forty-five days. This is intuitive because consumers with blemished credit histories may require more time to find and secure credit.

   Only one deduplication window could be selected--it would be impractical and confusing to consumers if differing...