This project presents an Agent-Driven Machine Learning workflow to predict credit risk using the German Credit Data. The model is built to classify loan applicants as either 'Good Risk' or 'Bad Risk,' a crucial task in FinTech.
To develop a robust binary classification model that maximizes the identification of high-risk customers, prioritizing metrics like Recall and F1-Score for the minority class ('Bad Risk').
- Agent Framework: AI Agent-Driven Development
- Languages: Python
- Libraries: Pandas, NumPy, Scikit-learn
- Modeling: Logistic Regression, Random Forest
The data was found to be highly imbalanced (70% Good Risk / 30% Bad Risk). After training Logistic Regression and Random Forest models, the following results were obtained for the critical 'Bad Risk' class (Class 1):
| Model | Accuracy (Overall) | F1-Score (Bad Risk) | Recall (Bad Risk) |
|---|---|---|---|
| Logistic Regression | 78.50% | 0.60 | 0.53 |
| Random Forest | 77.00% | 0.50 | 0.38 |
The Logistic Regression model was selected as superior due to its higher F1-Score (0.60) in identifying the minority class ('Bad Risk'). This indicates a better balance between precision and recall, which is paramount for financial risk assessment. Future work should focus on data augmentation techniques like SMOTE or cost-sensitive learning to improve the model's ability to capture all instances of bad risk.