Credit Risk Modelling For Dummies: But With Fewer Dummies
Standard credit risk analysis utilises scorecards which are built using only datasets with categorical variables. This requires the continuous numerical features to be fine-classed (grouped into discrete sets) and converted to dummy variables. Although the point of the scorecard is to present the model in a simple way, this practice requires much convoluted pre-processing of the data, which greatly bloats the size of the dataset and makes it more susceptible to containing errors. Most importantly though, I find that, by retaining the numerical features, the predictive power of the data is great improved, with a Gini coefficient of 0.95 (cf. 0.40 with fine-classing) and Kolmogorov-Smirnov statistic of KS = 0.85 (cf. 0.30).
Although all of the processing was done in the python scripts mentioned in report.pdf, a simplified run-through is included as the Jupyter notebooks PP.ipynb and ML.ipynb