In one of my previous posts, it was shown that how one can obtain PiT PD from TTC PD using the Kalman filter. I came across an interesting alternative approach on Quantitative Finance. In particular, assuming that TTC PD estimates are already available at the portfolio level and rating levels, a forecast of the portfolio level PiT PD is all that required by the approach. More concretely, under the approach, the PiT PD for rating i is given by the following equation:

In a predictive modeling project, one of the issues that an analyst always encounter is over what time period should a variable be derived. For example, should the analyst derive 3-month historical utilization rate to predict future default, or 6-month historical utilization rate is more appropriate. Usually, when deriving variables over time period makes sense, the analyst would settle on 4 time periods: 3-month, 6-month, 9-month, and 12-month. And then, based on some predictive performance metrics such as information-value (IV) or AUC, one variant of the derivation is short-listed for the next stage of the model development process.

Prior to the adoption of International Financial Reporting Standard 9 (IFRS9), provisioning were made only after exposures had turned delinquent. This “after-the-fact” shortcoming was heavily criticized for painting rosy pictures of the health of financial institutions before the 2008 Global Financial Crisis. IFRS9 addressed the shortcoming by introducing the concept of expected credit loss (ECL). The calculation of ECL is quite daunting. One of the inputs that required in the calculation is the default probability of an obligor given a certain economic state.

This is the final part of a 3-parts series. 6.0 Regression Analysis The ground is now set for developing a credit scorecard. The technique, as widely documented in the literature, is based on logistic regression. To obtain a parsimonious logistic regression, one approach is to use the stepwise method. This method seeks to minimize the AIC by allowing variables to enter or to exit iteratively. Each type of method has its own pros and cons and this will not be discussed here.

This is the second part of a 3-parts series. 4.0 Univariate Analysis 4.1 Fine Classing Fine classing is a technique that groups a variable’s values into a number of fine bins. Using these bins, a measure of the variable’s predictive power, known as information value (IV), can be computed. Also from these fine bins, further grouping can be carried out to result in coarse classing. As will be shown in the section below, bins from coarse classing are the bins that will be used in a credit scorecard.