Related Blogs


« Who are your strategic defaulters? | Main | The growth and evolution of ACH/wire fraud »

Predicting strategic default

The incidence of strategic default is on the rise. But before mortgage servicers can take action, they first must be able to identify strategic defaulters. Standard credit risk scores aren't sensitive to the quite different characteristics of these voluntary strategic defaulters, whose behavior is driven by incentive to pay rather than affordability.

FICO has recently published an Insights research paper describing a new analytic approach for finding those on the verge of strategic default—among delinquent accounts, as well as those who have not yet fallen behind on mortgage payments. Our research shows it's possible to know which 20% of accounts represents nearly 70% of your strategic default risk. This enables servicers to act sooner, to prevent losses and help customers avoid making a decision that will damage their credit standing.

The research leveraged a FICO-developed model aimed at teasing out the first signs of impending strategic default. Inputs to the model included credit bureau data and pooled master file data, as well as historical and forecasted property valuation data. This combination of data enables the model to analyze consumer credit behavior patterns with dynamic real estate market factors, such as relative change in property value over time and velocity of change.

The strategic default model identified 67% of strategic defaulters among the 20% of the nondelinquent mortgage population receiving the most risky scores. The most risky decile of borrowers were 100-times more likely to commit strategic default than the least-risky decile of borrowers.

Fig6-MortStrategicDefault 

The model was also precise in identifying strategic defaulters among a 30-to-180-day delinquent population (not yet written off). It found 76% of the strategic defaulters in the 30% of the population receiving the most risky scores. The analytic separation is again sharp—the most risky decile of borrowers are 200-times more likely to strategically default than the least risky decile of borrowers.

Fig7-MortStrategicDefault 

With this type of analytic, servicers can update their portfolio segmentation strategies to reflect the changing risk dynamics of today's marketplace. Segments can be created based on the dual dimensions of credit risk and strategic default risk.

First time on the Banking Analytics Blog?
Subscribe to the Banking Analytics Blog Feed or check out some other recent posts:

Comments

Daniel Lieb and Sasha Bucur


Thank you for sharing this article. We both have multiple years experience modeling similar mortgage data as well as other banking products and found it interesting.

One concern about your model: From figures 4-6, it appears that your KS is rather high (~0.45) and you capture 67% of strategic defaulters in the first two deciles. A naïve approach might look for such separation but further inspection of the data is revealing.

Typically when such dramatic separation occurs we find that we need to do initial segmentation before implementing the final model. Why? Because the results of the model are somewhat ‘artificial.’ The dramatic separation occurs because the model includes two distinct homogenous populations that cannot be combined and the unobserved homogeneity is actually driving the model.

More specifically, is there some other underlying variable that would account for the performance? Are almost all mortgage holders in the bottom five deciles non-delinquent? If nearly all non-delinquent mortgage holders can never be a strategic defaulter, then they should not be included in the model (segmented out) or if their incidence rate is an order of magnitude lower, a separate model should be built. What about LTV? Why would someone with a lot of equity in their house be a strategic defaulter? The same analysis applies.

The “strategic default” analysis should look at the definition of a strategic defaulter. The “strategy” assumes a benefit to the customer that decides to default. In such a scenario a person with high equity in their home (low loan-to-value ratio) could not be a “strategic defaulter.” A lender often does not allow loan modifications for people with equity and person’s equity in the home would be lost by defaulting. As such, there cannot be a “strategy” associated with it. On the average, any homeowner with LTV of 90% or lower would lose money under the “strategic default” definition. The rate of “strategic default” in the graphs shown (about 20% or so) is very close to the rate of homeowners that have little or negative equity in their households.

Under the definition described above, the propensity to be a “strategic defaulter” is very different between those with no or negative home equity versus those with positive home equity. As such, the lift chart shown and the model used to generate the ranking of “strategic defaulter” should be separated across the different population.

To visualize this further, consider the case where Amazon.com might try to sell their credit card to customers. A model might be able to achieve a high KS and dramatic separation where it to include all customers. But, if customers that seldom purchase from Amazon.com almost never accept a credit card offer, they should not be included in a model. In this case we segment out such customers and then build a model on the remaining population. Naturally, the KS will be lower, but this is irrelevant as measures such as KS are only applicable to the population on which you are modeling. The resulting model and all subsequent decisions based on the model will lead to a better understanding of the predicted and actual behavior.

The important lesson is that we should always look for segmentation first, and then model. If you first exclude a population that can never be a strategic defaulter, the resulting model will have a higher overall incidence rate and be both easier and more cost effective as well as make more business sense. Even more, the model definitions for the segments of the population may driven by different concepts which are otherwise muddled when non-homogeneous populations are used.

Daniel Lieb, Ph.D.
Sasha Bucur, Ph.D.

Joanne Gaskin

We completely agree about the importance of conducting segmentation analysis within the context of a model development effort. The strategic defaulter “model” is actually a collection of six models built on unique segments of the total mortgage-holding population. The segments are defined as permutations of a number of key variables, such as recent payment history on mortgage accounts and Current Loan To Value. In addition to validating the strategic defaulter model on the total mortgage-holding population, we also examined the efficacy of the score across various permutations of mortgage delinquency, CLTV, recourse vs. non-recourse states, etc. For example, we validated the model on a population with no recent mortgage delinquency but who were underwater (or very close to it) on their mortgage. Our finding was that the score proved to be a very effective tool for rank-ordering strategic default behavior in all of the segments that we examined.

HT

Excellent piece. It however boils down to what you define to be your target. Attrition, acceptance, default, strategic default... it could be anything....

I wrote about that here:
http://www.highstonetower.com/?p=901

The comments to this entry are closed.