Kalman filtering as a performance monitoring technique for a propensity scorecard

Bijak, Katarzyna (2008) Kalman filtering as a performance monitoring technique for a propensity scorecard. 28th International Symposium on Forecasting, Nice, France. 21 - 24 Jun 2008.

Record type: Conference or Workshop Item (Other)

Abstract

Propensity scorecards allow forecasting, which bank customers will soon be interested in new credits, through assessing their willingness to make application for new loans. Such scoring models are considered efficient tools to select customers for bank marketing campaigns. Kalman filtering can help monitoring scorecard performance. That technique is illustrated with an example of a propensity scorecard developed on the credit bureau data. Data coming from successive months are used to systematically update the baseline model. The updated scorecard is the output of the Kalman filter. As model parameters are estimated using commercial software dedicated to scorecard development, the estimator features are unknown. It is assumed that the estimator is unbiased and follows asymptotic normal distribution. The estimator variance is then derived from the bootstrap. The odds are defined as a measure of a customer’s willingness to apply for new loans, calculated as a ratio of the willing to the unwilling among customers having a given score. Once the scorecard is developed, an auxiliary linear model is estimated to find a relationship between the score and the natural logarithm of the odds for that score. That relationship is then used to determine the customer’s propensity level. Every month a new sample is scored with both the baseline and the updated scorecards. For each customer the log odds are estimated using the relationship between the baseline score and the willingness to apply for new loans. That estimate, which represents the customer’s propensity level provided that the baseline scorecard is still up-to-date, is then compared with the estimate computed using the relationship between the updated score and the log odds. The example demonstrates that a scorecard may become less and less up-to-date although the commonly used performance measures such as the Gini coefficient or the Kolmogorov-Smirnov statistic do not change considerably

This record has no associated files available for download.