AN ALGORITHM FOR SCORE CALIBRATION BASED ON CUMULATIVE BAD RATES

There is a great need in production for an agorithm which automates the task of discovering problems in the distribution of g (X) arise as the result of the calibration process, and this paper proposes such a fully automatic algorithm.

Abstract

In credit scoring, it is sometimes desirable to implement a new score in place of an existing score. Let Y denote the existing score and let X denote the new score. It is almost always the case that the score ranges for X and Y are different. For example, the score range for X might be 0 to 100, while the score range for Y might be 300 to 600. It follows from this difference in score ranges that a major difficulty in implementing any new score will be in training those who use the existing score to use the new score. Score calibration is the process of constructing a calibration function g (·) in such a way that the calibrated score, g (X), is a score which mimics the behavior of Y. In applications, it is commonly found that problems in the distribution of g (X) arise as the result of the calibration process. There is a great need in production for an agorithm which automates the task of discovering these problems and correcting them. In this paper, we propose such a fully automatic algorithm. This algorithm has been successfully tested in many applications.