Service overview

Scoring card calibration

A score is assigned based on an assessment of probability of some event such as a customer’s transition to 90+ DPD within the next 12 months or a refusal of some service, etc. We use a score predicting a borrower’s chances of going into a particular DPD category at some point.

The score is likelihood of a loan applicant’s becoming a “bad borrower.” It helps you range applicants so you only accept those most likely to pay back the loan. Past credit history lets the bank get a numerical estimate of future charge-offs.

Sooner or later a score’s quality (accuracy at predicting default) inevitably declines. This must happen as macroeconomics and client demographics change, the bank and competition redraw their policies and so on. But new information about a client, e.g., from a credit history company, can improve a score’s quality. This is why most banks routinely recompile score cards using latest and larger arrays trying to detect precisely the default possibility. But building a new score card and implementing it is usually difficult, and implementing it involves certain risks. It is much easier to amend a score by a process called “score card calibration”.

Calibration is basically a linear transformation to “fix” the score. For example, actual updated data on overdue interest are compared with the score’s predictions, a correction coefficient is set. Scoring card calibration via linear transformation is something RRAS excels at. The system can make use of all available statistics, including data on young vintages (mob <12).

Calibration, however, only solves some precision problems when estimating the default possibility. If a score’s ability to range clients (Kolmogorov–Smirnov statistics or, say, divergence value) declines, the score card will have to be rebuilt.



The author presents the methods for studying credit portfolio behavior in the Modeling and Stress-Testing Credit Portfolio Behavior are partly based on so-called “dual time dynamics” method. This work suggests using dual time dynamics not for decomposing scalar values but for decomposing matrices. The author considers a credit portfolio as a process described by a first-order heterogeneous Markov chain. Starting from this premise, the author uses vintage analysis and the theorem of strong convergence of modified fixed-point algorithms to arrive at transition matrix decomposition. This method makes highly accurate forecasts of credit portfolios possible. Reserves can be estimated with excellent precision and relevant values for stress-testing obtained.

In his later article The Theory and practice of Retail Credit the author considers some successful practical applications of his methods described in the article “Credit portfolios behavior modelling and stress-test”.