Case study / project evidence
Credit Risk Scoring
The main challenge of this project was to address the risk of loan defaults. Based on data from 466,285 customers, around 50,968 customers (11%) failed to repay their loans. The goal of this project was to develop a credit risk prediction model that minimizes default risk, identifies potential target markets, and determines key attributes that influence credit scores.
01 / Process
How the work unfolded
The dataset contained customer information from a loan company covering the period 2007–2014, with 74 columns of mixed data types (float, integer, and string) and about 460,160 missing values. The target variable was loan_status (0 = good, 1 = bad).
The process included:
- Data Preprocessing: Handling outliers, cleaning invalid data, managing missing values (dropping some columns and filling others with median/mode), feature engineering, scaling numeric features, encoding categorical features, and oversampling to address class imbalance.
- Model Training: Implemented several machine learning models including Logistic Regression, Decision Tree, Random Forest, and XGBoost.
- Model Evaluation: XGBoost achieved the best performance with an average accuracy of 87%, showing strong precision and recall in identifying high-risk loans.
- Feature Importance: Key factors influencing credit risk included recoveries (repayment plans), total_rec_late_fee, and int_rate.
- Validation: Performance was confirmed using AUC and KS metrics, both meeting industry standards (AUC > 0.7, KS > 0.3).
To make the findings more understandable, I created visual dashboards and feature importance plots that explained the drivers behind default risks.
The resulting predictive model not only provided an effective tool to reduce financial losses but also generated valuable insights for designing better loan products, targeting reliable market segments, and improving customer segmentation strategies.
03 / Product proof
Key features
- Large dataset with over 466K records and 74 features.
- Extensive data preprocessing pipeline for cleaning, feature engineering, and handling imbalances.
- Comparison of multiple machine learning models.
- Validation using standard financial modeling metrics (AUC, KS).
Continue exploring