San Francisco 2014
 
Blog:
blog home Home
 
       Follow:
RSS
 

The IFORS Distinguished Lecture – Credit Scoring and Financial Crisis

by Tengjiao Xiao on November 12th, 2014

The IFORS Distinguished Lecture was given by Dr. Lyn Thomas from University of Southampton. The lecture was both insightful and interesting. Dr. Thomas first introduced the history of credit scoring, which uses operations research and statistical models to assess default risk, for consumer lending. The fact that San Francisco is the birth place of credit scoring is a happy coincidence. Among the different approaches have been used in credit scoring, logistic regression is the most common one being used now. Besides logistic regression, classification trees and ensemble models are currently in use as well. Credit scoring for subprime mortgages played an important role in the financial crisis. There were two credit models mainly used in the financial crisis: bureau scoring models and rating agencies’ models. The objective of the Fannie Mae, Freddie Mac, and rating agencies such as Standard & Poor’s was to acquire loans and securitize them. Therefore, they strongly tended to give subprime loans very high scores so that they can sell the products to investors. Dr. Thomas suggested seven lessons for O.R. in credit scoring, and I quote:

1. Ensure model objective relevant to decision maker.

  1. Know if static model is sufficient or if situation dynamics needs to be included.
  2. Making models “public” means model will be gamed.
  3. Verify data used in model.
  4. If updated data becomes available, build model to use the updated data.
  5. Model extrapolation can be dangerous. Credit rating agencies used corporate model for consumer CDOs.
  6. If model disagrees with common sense, think before using the model.
1 Comment
  1. John Garrett, PhD permalink

    I disagree that the models built before the crisis were the primary problem, and I disagree that keeping the models secret would have helped reduce the crisis. From published models built for the Federal Reserve by Anthony Pennington-Cross on loan level data, it is clear that the loss predictions were extremely sensitive to housing prices. Anyone plugging the actual housing price collapse into those models would have known that there would be huge losses in subprime and prime mortgages in the Great Recession. Fannie Mae as early as 2001 realized that housing prices had risen above levels justified by fundamentals, yet housing prices rose even faster for several more years. With so much money being made by mortgage and home equity lending, few forecasters were willing to call a 30% decline in housing prices within 2 years, particularly since overall US housing prices had NEVER declined in national data from 1975 up until the onset of the Great Recession. This would put any price decline outside crude 95% confidence intervals, yet housing prices fell 30%.

    Keeping housing prices outside the models would have been worse: It would just would have meant that forecasters would have been surprised about losses when house prices fell.

    I don’t think gaming the models was the problem: This was a speculative boom funded by poor lending standards. Once housing prices began rising, even very bad loans could be paid back from the proceeds of rising home values, resulting in low observed loss rates. Once housing prices began to fall and lenders realized that they needed to tighten lending standards, a self-reinforcing process began that caused losses to skyrocket.

Leave a Reply to John Garrett, PhD

Note: XHTML is allowed. Your email address will never be published.

Subscribe to this comment feed via RSS