🇹🇷 Türkçe dokümantasyon için buraya tıklayın.
The primary goal of this project is to solve a critical problem in the banking sector: Credit Default Risk Prediction. Using historical data provided by Home Credit, I analyzed customer behaviors and built a Machine Learning model to predict whether a borrower will repay their loan or default.
This project goes beyond simple modeling; it involves deep SQL Analysis, Feature Engineering, and handling Imbalanced Data using advanced Boosting algorithms.
- Data Analysis: SQL (SQLite) & Python (Pandas)
- Visualization: Matplotlib & Seaborn
- Machine Learning: Random Forest & LightGBM (Gradient Boosting)
- Technique: Class Weighting for Imbalanced Data, Threshold Tuning
Before feeding the data into the model, I conducted a detailed analysis using SQL and Python to understand risk factors.
Observation: Younger clients (<30 years) are significantly riskier than older clients (>60 years).
- 20-30 Years Old: 11.46% Default Rate
- 60+ Years Old: 4.92% Default Rate
Observation: There is a correlation between the number of children and default risk. Families with 3+ children show the highest default probability (%10.04).
Observation: Surprisingly, people requesting loans 6+ times their income have a lower risk than those requesting 3-6 times. Interpretation: This indicates strict bank policies; only highly qualified customers are approved for such large loans (Survivor Bias).
The dataset was highly imbalanced (only ~8% defaults). A standard model would fail to detect risky customers (High Accuracy, Low Recall).
I used LightGBM with class_weight='balanced' and scale_pos_weight to solve this.
| Metric | Score | Description |
|---|---|---|
| ROC-AUC | %75.26 | Strong ability to distinguish between classes. |
| True Positives | 2,984 | Potential defaulters successfully detected. |
| False Negatives | 1,981 | Missed defaults (reduced significantly from baseline). |
The model successfully identified 2,984 risky customers who would have otherwise caused financial loss.
To provide business value, we must explain the "Why". The LightGBM model prioritized the following features:
- EXT_SOURCE (1, 2, 3): External credit scores are the strongest predictors.
- DAYS_BIRTH: Age is a critical demographic factor.
- AMT_ANNUITY: The burden of monthly payments affects repayment capability.
By transitioning from a baseline Random Forest to a tuned LightGBM model, we increased the detection of risky customers from near zero to ~60%. Implementing this model can help the bank:
- Minimize Non-Performing Loans (NPL).
- Optimize Interest Rates based on risk profiles.
- Automate the initial credit screening process.
Author: [Mehmetcan Mutlu]