New Scores in Mortgage Models
Recently the Federal Housing Finance Agency (FHFA) announced some upcoming changes related to the use of new credit scores, FICO 10T and VantageScore 4.0 by Fannie Mae and Freddie Mac. “FHFA expects that implementation of FICO 10T and VantageScore 4.0 will be a multiyear effort. Once implemented, lenders will be required to deliver both FICO 10T and VantageScore 4.0 credit scores with each loan sold to the Enterprises”. This announcement will impact the entire mortgage ecosystem.
In this blog, I will discuss some of the challenges that come with transforming the analytical models used to value and manage the risk of mortgage loans and securities. Behavioral models are used to forecast the probability of prepayment, delinquency, default, and losses given default. Credit scores are typically inputs into these models. The models in use today are all calibrated using the classic FICO score.
Getting the models ready to run with FICO 10T and VantageScore 4.0 will mean that every single model using these scores will have to be refitted, tested, and validated before they can be put into production. Let us look at what all this entails.
Two populations will be affected by this change:
- Population of new loans
- Population of existing loans and securities
Let’s start with the first population. All new loans will have to be analyzed using FICO 10T and VantageScore 4.0, which means that any origination or risk models for these loans will need to be estimated using the new scores. Underwriters will need to understand the nuances of the new scores and will probably need a mapping from old scores to new scores.
For the second population, we will need to refit the existing prepayment and credit models used by the industry. The new models should take the new scores as inputs. We will need historical data for FICO 10T and VantageScore 4.0 going back at least to the financial crisis of 2008, along with the loan and collateral information. Having data from various economic cycles will be important to parametrize the models with the new scores and validate the sensitivity of the various factors in the models that use the new scores. It would be good to have data from the period leading up to and following the crisis. Pre-crisis will let us quantify the “bad” loans that led to the crisis, whereas post-crisis will let us quantify the delinquencies, defaults, and losses. We also need data for periods when rates went down and when rates went up. With its historically low rates, the pandemic period is a unique period and probably less important from a historical perspective.
Trended data gives us information about a borrower's financial position or liquidity at a given time. A borrower who pays the minimum payment on their credit card debt is called a “Revolver,” while a borrower who makes a full payment is called a “Transactor”. A limitation of the trended data available today is that some major credit card issuers do not report the trended information to the bureaus, which means that the trended scores would have limited training data sets. Is there a way to overcome this bias? Utility data is also not readily available for most borrowers. Fannie Mae and Freddie Mac are now using rental data, but there is no good source of rental data for industry participants.
As we look at new scores, we should also consider how the scores can be made more useful. We know that a borrower’s trended data affects every loan transition throughout the loan lifecycle. Loans could transition from being Current to Delinquent to being Seriously Delinquent to becoming Real-Estate Owned (REO) and finally terminate (and could also transition to prior states). These loan transitions have increased predictive power if we use trended data. The question is, how can industry participants use this information?
This will be a multi-year effort for the industry. It would be better if, for the population of existing loans (about $11 Trillion), we could find an easy way to bridge the existing scores with trended and utility data. Also, as we start using rental, utility, and telecom data, we will have previously unscored loans coming into the ecosystem. There will be a need for frequent model updates as we get additional history about the behavior of these borrowers in various stress environments.
A solution is to use the classic FICO score and use other variables that are orthogonal to the classic FICO score to obtain metrics that are much more predictive through the entire loan lifecycle. A benefit of doing it this way is that we can use loan and collateral information which is not available in a credit score alone. For example, LTV or loan-to-value significantly impacts borrower behavior in stress situations.
We currently do not know a lot about the transition pathway to FICO 10T and VantageScore 4.0 in models used by the mortgage industry. However, one thing is clear. It will take many years before the market is positioned to utilize the advances in analytics coming from new data and new models.
A big question for all market participants is, who will provide the historical data required to recalibrate the models? It is not enough to just have access to the new scores. There should be a way to merge the scores with the collateral and loan data. Fannie Mae and Freddie Mac would be good sources for loans sold to the enterprises, but we would also need data for FHA/VA loans and loans held in bank balance sheets. Also, for the agencies, we would need data for all loans and not just for the loans in the CRT (Credit Risk Transfer) reference data set.
We at Andrew Davidson & Co., Inc. have been working with trended data from Equifax and have found interesting ways to link our prepayment and credit models with the available trended data. It is almost like the next frontier in mortgage prepayment and credit modeling. Models evolve with the availability of new data. Bringing borrower credit bureau data into the modeling process will help us understand and forecast borrower behavior in a much more meaningful way.
FICO 10T, VantageScore 4.0, and Equifax are trademarks of Fair Isaac Corporation, VantageScore Solutions, LLC, and Equifax, Inc., respectively.