Credit Models Are Degrading at a Clip Last Seen in the Great Recession

If you build or rely on statistical models to drive decisions for your lending business, you should be worried.  The models used by lenders since the Great Recession were built using increasingly sophisticated data sources during one of the most prosperous times in modern history.  As banks protect assets and help consumers, in part by offering forbearance programs for some COVID-19 impacted borrowers, data quality will deteriorate. Given the skew forbearance programs cause in reported bureau delinquencies, the credit data that is available can hardly be trusted.  Major economic indicators are moving by magnitudes never seen before.  Unemployment is quite literally off the charts[i], GDP is expected to be down 5% for every month of a partial shutdown[ii], and the Institute for Supply Management, which produces manufacturing indexes, indicates manufacturing production fell to its lowest level since January 1948[iii].  How lenders adapt to short-term implications of model deterioration will be a key differentiator in determining winners and losers during recovery and expansion in the long-term.   

The finely tuned models and data-driven segmentation strategies that banks and Fintechs spent the better part of the last decade building are rapidly degrading. These models were grounded in data from an environment which does not reflect the current situation, and as a result any data- or model-driven strategy is now misguided at best. AQN expects this issue will continue to compound in severity as variables rapidly gain or lose predictive power.  During the last recession, seemingly overnight, homeownership transitioned from a low risk to high risk indicator.   Based on April 2020 unemployment rates, even lenders who verify income and employment have a ~10% chance of a false signal from applications in March. Break points like these are surfacing in models used for decision making across all verticals in banks. This means firms must reevaluate their marketing, acquisitions, customer management, and loss mitigation strategies.

Source: U.S. Employment and Training Administration, Initial Claims [ICSA], retrieved from FRED, Federal Reserve Bank of St. Louis

Even if lenders want to rebuild their models, and eventually they should, latency and inconsistent reporting in credit performance data means current and accurate data is months away.  Moreover, the situation is changing too rapidly. For example, people with lower incomes may be returning to work in some form or receiving government relief while higher incomes customers are beginning to see growing impacts from layoffs driven by longer term concerns about demand. Since pausing lending until models can be rebuilt is not a realistic option, banks will need to think strategically in the short term.   

We anticipate many lenders will respond by stressing loss rate assumptions in their current models within their current segments and adjusting thresholds to regain equilibrium. This response is a sensible short-term response. However, it implicitly assumes that models still rank order accurately because they are built on the right signals with accurate weight given to each of these signals. Furthermore, this assumes that powerful new segmentations are not emerging amongst their prospects, applicants, and customers. These are bold assumptions given the rapid changes currently unfolding.

Potential Short-Term Solutions

• Calibrate models using alternative data or assumptions based on most recent data

• Create new monitoring plans, centered on COVID-specific segmentation strategies

• Augment models and segmentation with analysis-based expert opinion

Short-term strategies require a combination of judgment and expertise in understanding traditional and alternative data sources. Which data sources add value, and how to leverage them, will vary based on the asset class and customer segment. Most lenders will need to pull in new data or change the way they look at their current feed to avoid costly blunders. The speed and magnitude of change in the current environment requires meaningful and sustained investment in assessing, prioritizing, and addressing strategies impacted by modeled decision across the customer lifecycle.

Failing to quickly, and continuously, assess segmentation strategies and adjust models carries serious risks. Collections treatment groups based on pre-crisis data could further hamper agent teams already addled by the transition to remote work. Most lenders would not approve an application from an individual currently 60 days delinquent on a mortgage, but traditional bureau attributes may obscure customers’ inability to pay for the past two months.

The situation may seem dire, but it is not all bad news for lenders. The diminishing supply of available credit creates an enormous opportunity for lenders to acquire strong customers at an advantaged cost and on improved terms. However, seizing this opportunity will again require expertise, timing, and of course, highly predictive models.

[i] https://www.nytimes.com/interactive/2020/05/08/business/economy/april-jobs-report.html

[ii] https://www.mercatus.org/system/files/makridis-cost-covid-19-mercatus-v1.pdf

[iii] https://www.wsj.com/articles/u-s-factory-sector-contracted-in-april-11588343731