Machine Learning · Risk Management
An end-to-end credit scoring and loss optimization tool for credit card default risk, designed to be accessible to both technical and non-technical users.
Demo: UI Tool walkthrough — predicting default probability for individual customers
This project develops an end-to-end credit scoring and loss optimization tool for credit card default risk. It combines machine learning model development, expected loss analysis, and an interactive UI — making the results accessible to both technical and non-technical users.
Three machine learning models were evaluated and benchmarked against each other using ROC-AUC, F1 score and recall on the default class.
| Model | F1 Score | Recall (Default) | ROC-AUC (validation) | |
|---|---|---|---|---|
| XGBoost | 0.46 | 0.42 | 0.75 | |
| Random Forest | 0.48 | 0.51 | 0.76 | |
| Neural Network | 0.51 | 0.62 | 0.79 | Selected |
The model Neural Network was chosen.
Using the default probabilities from the Neural Network, the project constructs a distribution of expected losses across the customer portfolio. Tail risk is quantified at the 95th and 99th percentiles, and mitigation strategies are proposed and evaluated based on their effectiveness in reducing these concentrations.
The project is finalized as an interactive interface built with Streamlit, allowing non-technical users to query the default probability of individual customers within a few clicks.
To launch the tool locally: