By: Kari Michel
How are your generic or custom models performing? As a result of the volatile economy, consumer behavior has changed significantly over the last several years and may have impacted the predictiveness of your models. Credit models need to monitored regularly and updated periodically in order to remain predictive.
Let’s take a look at VantageScore, it was recently redeveloped using consumer behavioral data reflecting the volatile economic environment of the last few years. The development sample was compiled using two performance timeframes: 2006 – 2008, and 2007 – 2009, with each contributing 50% of the development sample. This is a unique approach and is unlike traditional score development methodology, which typically uses a single, two year time window. Developing models with data over an extended window reduces algorithm sensitivity to highly volatile behavior in a single timeframe. Additionally, the model is more stable as the development is built on a broader range of consumer behaviors.
The validation results show VantageScore 2.0 outperforms VantageScore 1.0 by 3% for new accounts and 2% for existing accounts overall.
To illustrate the differences that were seen in consumer behavior, the following chart and table show the consumer characteristics that contribute to a consumer’s score and compare the characteristic contributions of VantageScore 2.0 vs VantageScore 1.0.
Payment History | Utilization | Balances | Length of Credit | Recent Credit | Available Credit | |
Vantage Score 2.0 | 28% | 23% | 9% | 8% | 30% | 1% |
Vantage Score 1.0 | 32% | 23% | 15% | 13% | 10% | 7% |
As we expect ‘payment history’ is a large portion driving the score, 28% for VantageScore 2.0 and 32% for VantageScore 1.0. What is interesting to see is the ‘recent credit’ contribution has increased significantly to 30% from 10%. There also is a shift with lower emphases on balances, 9% versus 15% as well as ‘length of credit’, 8% versus 13%.
As you can see, consumer behavior changes over time and it is imperative to monitor and validate your scorecards in order to assess if they are producing the results you expect. If they are not, you may need to redevelop or switch to a newer version of a generic model.