Loan Risk Scoring with Machine Learning

Overview

This project involves creating a model to calculate expected loss in a banking risk context. Using a loan dataset, different models are trained to estimate the capital at risk for that specific loan.

In addition, it includes a Streamlit web application that allows users to calculate loan risk using the trained models.

Risk Scoring Approach

The model follows the standard banking risk framework:

$$ EL = PD \times P \times EAD \times LGD $$

Where:

• EL = Expected Loss (€)
• PD = Probability of Default (%)
• P = Principal (€)
• LGD = Loss Given Default (%)
• EAD = Exposure at Default (%)

In order to calculate the expected loss, the models would predict from the personal data:

• PD with a Logistic Regression
• LGD and EAD with LightGBM

Typical elements of banking risk modeling (WOE, KS, information value, ...) were not used in order to keep this project more applicable to a general case

Process

1. Data preparation

   • Clean the data, group atypicals
   • Create new variables to get the target information for the estimation

2. Modeling

   • Optimize hyperparameters for the three models
   • Prepare the final pipelines for production

3. Production

   • Scripts for retraining and execution of the predictive models
   • Build the streamlit app based on the predictive models

Notebooks & Scripts

• Development notebooks
• Retraining script
• Execution script
• Streamlit app code

Installation

Clone the repository and install dependencies:

conda env create -f 01_Documents/risks.yml
conda activate risks