We start by explaining why credit risk is important for financial institutions. We also define ground 0 terms, such as expected loss, probability of default, loss given default and exposure at default.

Here you will learn how to set up Python 3 and load up Jupyter. We’ll also show you what the Anaconda Prompt is and how you can use it to download and import new modules.

Our example focuses on consumer loans. Since there are more than 100 potential features, we've devoted a complete section to explain why some features are chosen over others.

Each raw datasets has its drawbacks. While most preprocessing is model specific, in some cases (like missing values imputation), we could generalize the data preparation.

Once we have completed all general preprocessing, we dive into model-specific preprocessing. We employ fine classing, coarse classing, weight of evidence and information value criterion to achieve the probability of default preprocessing. Conventionally, we should turn all variables into dummy indicators prior to modeling.

Having set up all variables to be dummies, we estimate the probability of default. The most intuitive and widely accepted approach is to employ a logistic regression.

Since each model overfits the training data, it is crucial to test the results on out-of-sample observations. Consequently, we find its accuracy, its area under the curve (AUC), the Gini coefficient and the Kolmogorov-Smirnov test.

In practice, banks don't really want a complicated Python-implemented model. Instead, they prefer a simple score-card which contains only yes and no questions that could be employed by any bank employee. In this section, we learn how to create one.

Model estimation is extremely important, but an often-neglected step is model maintenance. A common approach is to monitor the population stability over time using the population stability index (PSI) and revisit our model if needed.

To calculate the final expected loss, we need three ingredients: probability of default (PD), loss given default (LGD) and exposure at default (EAD). In this section, we preprocess our data to be able to estimate the LGD and EAD models.

LGD models are often estimated using a beta regression. To keep the modeling part simpler, we employ a two-step regression model, which aims to simulate a beta regression. We combine the predictions from a logistic regression with those from a linear regression to estimate the loss given default.

The exposure at default (EAD) modeling is very similar to the LGD one. In this section, we take advantage of a linear regression to calculate EAD.

After having calculated PD, LGD, and EAD, we reach the final step: computing expected loss (EL). This is also the number which is most interesting to C-level executives and is the finale of the credit risk modeling process.