The objective of this project is to develop a supervised machine learning model to predict loan default based on a variety of observable and legally usable client and transaction characteristics. This is a critical task in the banking sector, which seeks to minimise risk and ensure financial stability by accurately forecasting which loans are most likely to default.
We utilise the Berka Dataset, a comprehensive, anonymised financial dataset from a Czech bank, which includes information on clients, accounts, transactions, and loans.
The dataset comprises multiple relational files capturing different facets of banking activity:
| File Name | Observations | Description |
|---|---|---|
account.asc |
4,500 | Static characteristics of bank accounts |
client.asc |
5,369 | Demographic characteristics of clients |
disp.asc |
5,369 | Mapping between clients and accounts |
order.asc |
6,471 | Information on credit card payment orders |
transaction.asc |
1,056,320 | Detailed account transactions |
loan.asc |
682 | Records of loans granted to accounts |
card.asc |
892 | Details of credit cards issued |
district.asc |
77 | Demographic statistics of districts |
-
Load and Explore the Data
Load selected datasets and perform initial exploratory analysis to understand distributions, missing data, and relationships. -
Preprocessing & Feature Engineering
Merge datasets (e.g.,client,account,loan,transaction), handle missing values, and create features relevant to loan risk (e.g., average balance, transaction count, district unemployment rate). -
Model Training
Apply classification models such as Logistic Regression, Random Forest, and Gradient Boosting. Tune hyperparameters using cross-validation. -
Evaluation
Use accuracy, precision, recall, F1-score, and ROC-AUC to evaluate model performance on a hold-out validation set.
Target Variable: loan default (yes/no)
Goal: Predict whether a loan will default based on financial and demographic characteristics prior to or at loan issuance.
- Python 3.x
- pandas, numpy
- matplotlib, seaborn
- scikit-learn
- xgboost (optional)
- Jupyter Notebook