This course is part two of a two course series. After covering model use and maintenance aspects including validation and re-calibration in the first course, the focus is now on the original creation of risk prediction models.
While the topics in the first course relate more to the Business-As-Usual activities of a credit risk management team, model creation often takes the form of projects, which may sometimes be outsourced to third parties. Having said that, it is strongly advised for everybody working in this field to have a good understanding of all rating system aspects.
The first part of this course explains PD modelling in general and scorecard development in particular. We first cover the general preparatory steps of sampling and partitioning the data. Then we address the various challenges a multivariate model faces and present a variety of common model types that differ in the way they address these challenges.
We then focus on scorecard development in particular, starting with variable grouping and re-coding, moving on to the analysis of variable redundancy and the details of calculating the coefficients of a logistic regression model and finally ending with the calculation of scorepoints.
Two closely related additional topics end the first part of the course. Risk based segmentation uses variable grouping and re-coding from scorecard development to identify segments of cases that share similar value combinations of the most important risk drivers. Horizon-less scorecards address the restrictions imposed by a fixed prediction horizon and predict individual maturation curves instead.
The second part of the course covers LGD and EAD modelling. First, the definition of a workout LGD target variable and the requirements for specific models for already defaulted cases and for downturn adjustments are discussed.
Various LGD modelling approaches are then presented, ranging from simple segment averages to scenario based setups, in which predictions of a workout path are combined with predictions of a respective recovery rate. A large variety of statistical model types are presented that address the specific prediction tasks.
EAD modelling ends the course. Attention is given to the definition of the target variable, especially with regards to the choice of predicting original or undrawn limit usage.
