Menu

Modeling Risk

The End of Credit Scores?

The pandemic is illuminating flaws in traditional consumer credit assessment methodologies. Will we therefore see the emergence of open-sourced algorithms and alternative, customized scoring mechanisms that yield more optimized scores?

Friday, September 4, 2020

By Cristian deRitis

Advertisement

Just as COVID-19 has accelerated trends in online shopping, remote working and videoconferencing, the measurement and provision of credit is bound to experience a transformation in coming years. Changes, in fact, are already underway, thanks in part to pandemic-driven policies that could lead to an overhaul of the traditional credit-scoring system.

The latest example of the impact that policy decisions can have on credit scores is related to COVID-19. As part of the CARES Act, the consumer credit bureaus have been instructed not to penalize borrowers who may be receiving payment forbearance. While well-intentioned, the suppression of information could have unintended consequences.

Cristian deRitis headshot
Cristian deRitis

Before breaking down the current system, and examining future possibilities, it's helpful to discuss how we got to where we are today.

The consumer credit score was one of the greatest innovations in risk management of the 20th century. Before credit scores were developed, loan applications were processed individually by a credit officer who subjectively considered the five C's of credit: character, capacity, capital, collateral, and conditions. At best, these evaluations were applied fairly but inefficiently. At worst, they were applied with explicit preferences for specific demographic groups.

Credit scores helped sharply reduce the overt discrimination that plagued the lending industry throughout much of its history. The new scoring algorithms were based solely on quantitative measures related to the propensity to default, such as the number of times a consumer had paid late on previous loans.

But credit scores are not perfect. Just as with other applications of machine learning, credit scores can reflect and even accentuate historical biases.

If a class of consumers has been discriminated against or disadvantaged historically, then the statistical methods used to estimate credit scores may codify these relationships and perpetuate them. For example, if borrowers from a certain geography had been more likely to be approved for credit in the past, a machine learning algorithm might continue to associate this geography with lower default risk in the future.

Problems with Traditional Methodologies

Ultimately, credit scores are only as good as the underlying data that goes into them. Erroneous data can lead to credit scores that are artificially low, while the suppression of data might create scores that are not as predictive as they could be.

Credit scores don't live in a vacuum. While borrowers may benefit from not seeing their scores change in the short run, lenders may be wary of using scores with diminished predictive power.

Lenders might resort to adopting average-based pricing as a result - potentially increasing interest rates on most borrowers to cover the losses on the higher-risk borrowers they can no longer detect. Alternatively, lenders may seek out other data to supplement the degraded credit score.

Another issue with traditional credit scoring models is their failure to consider the impact of the external environment on credit behavior. Ideally, we'd like credit scores to only capture individual behavior. This distinction not only provides a cleaner model but also allows us to clearly communicate the corrective actions that consumers could take to improve their credit.

The real world, however, is not so distinct. Borrower behavior is deeply entwined with the external economic environment. For example, a borrower who continues to make payments during a recession is presumably stronger than a borrower who makes payments during an economic expansion.

Failure to control for the external economic environment may cause us to incorrectly measure the importance of borrower-specific factors, such as credit balances or the number of late payments.

Indeed, if we look at the average credit score over time, we see a strong correlation between average scores and the credit cycle. That correlation could lead to artificial “inflation” of credit scores, yielding an underestimation of credit risk in good times that is exacerbated in bad times (see chart, below).

Cyclical Credit Scores May Mask True Risk
Cyclical Credit Scores May Mask True Risk headshot
Sources: Equifax, Moody's Analytics

Economically-Adjusted Scoring, Alternative Algorithms and Custom Models

Scores go up when consumers make fewer delinquent payments, but delinquent payments are tightly connected to the economy - not just individual attributes and attitudes toward credit. My research suggests that adjusting scores for economic conditions can reduce the impact of credit score inflation and offer a more realistic picture of underlying risk.

Moreover, these adjustments are countercyclical, leading lenders to increase the cost of credit when the economy is hot and to provide more credit during downturns. Historically, the best loan vintages have been originated in the worst of times - and vice versa. Economically-adjusted credit scores can help lenders to better assess the credit risk of applicants throughout the business cycle.

Although consumers now have greater access to their own credit reports and credit scores, they lack insight into the proprietary scoring algorithms used to calculate them. So, the opacity of credit scores is yet another concern.

Lenders may prefer to develop their own algorithms to better understand and explain the factors affecting their decisions. Ultimately, regulators hold lenders responsible for their actions - and owning the credit score algorithm gives the lenders greater control.

Even before the coronavirus hit, many lenders (particularly fintechs) were questioning the value provided by generic credit scores for decisioning and risk management processes - and began exploring alternatives.

One alternative is the creation of custom internal credit scores that are finely tuned to specific lending products. A score that is optimized for a personal loan portfolio may be superior to a generic score that considers performance across all loan products.

Other lenders have considered eliminating credit scores altogether from their credit modeling and decisioning processes. Rather than distilling numerous factors into a single score that is then input into a probability of default model, why not enter these factors directly into the default model itself?

We've needed credit scores to reduce the dimensionality of models, given limits on computing power in the past. But those limits no longer exist. We can now estimate models with hundreds or even thousands of parameters quickly and efficiently.

Parting Thoughts

As credit data becomes more ubiquitous and specialized firms evolve to collect it, more lenders will likely turn to alternative sources of data to develop their own optimized credit scores.

We may also see the emergence of open source credit scoring algorithms to compete with the proprietary “black box” models that dominate the marketplace currently. Smaller institutions may find value in pooling their data and resources to create transparent, open source standards to compete against larger lenders.

While credit scores still have their purpose, new technologies will enable lenders of all sizes to consider other options to expand credit availability, lower interest rates and manage their portfolios more effectively.

 

Cristian deRitis is the Deputy Chief Economist at Moody's Analytics. As the head of model research and development, he specializes in the analysis of current and future economic conditions, consumer credit markets and housing. Before joining Moody's Analytics, he worked for Fannie Mae. In addition to his published research, Cris is named on two U.S. patents for credit modeling techniques. He can be reached at cristian.deritis@moodys.com.




Advertisement

BylawsCode of ConductPrivacy NoticeTerms of Use © 2024 Global Association of Risk Professionals