Subscribe to DSC Newsletter

Subprime crisis: did statistical models fail? Where to go from here?

Some people blame the crisis on poor statistical models, particularly credit scoring models predicting the chance that someone will be able to pay his/her mortgage. Supposedly, these models did not work well / were not enough tested on people with poor credit or stated income, subject to new exotic types or mortgage such as reverse mortgages. What are your thoughts on this issue?

Views: 249

Reply to This

Replies to This Discussion

There's an excellent article at
detailing Moody's approach to credit ratings
Basically the statistical models were excellent *but* no statistical model in the world can give you any *prediction* unless you also have an understanding into the why, the dynamics that drive the models, and an understanding that when the dynamics change the prediction must change.

If you can't promise that the dynamics are going to be stable in the future, or that you understanding how the dynamics can change and that the model can handle that change, then you have no business making a prediction.

Understanding the dynamics takes insight, experience, judgment , investigation, and can't be the result of hitting a data set with a technique.
According to my recent reading, it seems that there was a statistical model problem in the mortgage crisis; but the credit scoring models were not the critical issue.

In reading How Markets Fail, The Logic of Economic Calamities by John Cassidy, he points out that the President George W Bush program to reduce redlining was using minority diversity in completed mortgages as the metric of success. When he also allowed Wall Street to use mortgages as fodder for a financial instrument, the mortgage review due process evaporated in some lending institutions as the bubble grew. There were clearly mortgages issued where the parties could not pay the monthly interest that had to fail. Basically credit model scores were ignored in order to generate more fodder for Wall Street's mortgage bubble.

Pablo Triana in LECTURING BIRDS on FLYING, Can Mathematical Theories Destroy The Financial Markets published by John Wiley and Son, observes that the worst('toxic') mortgages were buried deep in the financial instruments, so the purchasers could not estimate the risk. Furthermore, Wall Street systematically under estimated their instrument risks because they used a portfolio correlation coefficient model which required the efficient market hypothesis. Although the individual instrument portfolios were diversified; there were large clusters of portfolios with similar diversification patterns.

When the first mortgages started failing, they reduced the liquidity in their cluster. As the supply and demand dynamics reduced housing prices and increased the rate of mortgage defaults, the lack of diversification eliminated the liquidity accelerating the positive feedback loop which was decreasing housing prices and increasing the default mortgage rate which eventually burst the bubble.

The need to monitor the loss of instrument liquidity as a function of the pattern of portfolio diversification was not expected. It would be interesting if datamining could approximate that function.
Both Thomas and Edmund's responses are right on the money. I have some inside information on what happened with subprime mortgages as I was analyzing the credit risk of WaMu's prime mortgage portfolio and just before I jumped ship in mid-2006 for greener pastures, I was given the responsibility of estimating the credit losses of Long Beach Mortgage's subprime portfolio (WaMu's subprime subsidiary). Here's what I found out:

1) Long Beach Mortgage (LBM) did not utilize credit models. FICO scores were next to useless when it comes to subprime borrowers. Let's face it - pretty much all applicants had FICO scores less than 600.

2) The major drivers behind application approvals were: housing price, loan fees, and some form of income verification. LTV was not much of a consideration given that many of those subprime mortgages were at 100% and even 105% of LTV (many were CRRA loans).

3) LBM acted more like a pawnshop than a mortgage lender - collateral was the underlying property. Default rates were expected to be high but LBM counted on rising property values and a liquid market to reduce the Loss Given Default (LGD).

Problem was, even before the financial crisis, LBM portfolio LGD was exceeding 25% and WaMu was forced to shut down LBM and subprime lending in 2006 and eat the losses. At one meeting, a senior manager commented that LBM was like an asylum run by the inmates. The poor performance of the portfolio could be traced back to extremely lax underwriting and looking to a rising housing market and economy to prevent any excessive losses.

It was no surprise when WaMu's prime mortgage portfolio imploded soon after LBM's demise but fortunately, I was not around to witness it. It would have been quite a carnage, to say the least.

Moral of the story? Subprime risk is a Frankenstein that could turn on you anytime. Modeling subprime risk using regular credit bureau data is next to impossible.


On Data Science Central

© 2021   TechTarget, Inc.   Powered by

Badges  |  Report an Issue  |  Privacy Policy  |  Terms of Service