This project develops and evaluates machine learning models to classify loan applicants as Good (1) or Bad (0) credit risks, with a strong focus on business-aligned decision making rather than just model accuracy. It demonstrates how credit risk models can be translated into lending policies through threshold tuning, risk banding, and lending policy simulations. The analysis is implemented in Jupyter notebook, written as a narrative workflow with short cells and inline commentary explaining each modeling choice, threshold decision, and business insight.
| Aspect | Summary |
|---|---|
| Problem | Predict whether an applicant is a good or bad credit risk. |
| Dataset | German Credit Dataset (1,000 records, mixed numerical & categorical features). |
| Algorithms | Logistic Regression, Decision Tree, Random Forest, XGBoost. |
| Final Model | XGBoost with tuned decision threshold. |
| Key Metrics (Test Set) | 73% accuracy, 68% recall for Bad class, 85% precision for Good class. |
| Business Layer | Risk banding and lending policy simulation on test set. |
Financial institutions need to balance risk (avoiding defaults) and growth (approving more loans).
This project builds a credit risk model and then connects it to practical decision frameworks that a bank could use in production.
Core objectives:
- Build interpretable models to classify applicants as Good (1) or Bad (0).
- Improve detection of bad-credit applicants using threshold tuning.
- Segment applicants into risk bands (Low, Medium, High).
- Simulate lending policies under different risk appetites.
- German Credit Dataset
- Mix of numerical and categorical features
- Target variable: Risk
0→ Bad credit risk1→ Good credit risk
Due to the limited dataset size, the project focus is on robust methodology, interpretability, and decision frameworks rather than overfitting for marginal metric gains.
Key analyses performed:
- Descriptive statistics and value counts
- Missing value handling (treated as a separate category where applicable)
- Correlation heatmap
- Group-wise analysis (credit amount by sex, job, housing, etc.)
- Risk-wise visualizations:
- Box plots and violin plots
- Scatter plots (Credit Amount vs Age, sized by Duration)
- Count plots for categorical features split by risk class
EDA was used to understand customer behavior and risk patterns before modeling.
- One-hot encoding: purpose, sex
- Ordinal encoding:
- Saving accounts:
none < little < moderate < rich < quite rich - Checking account:
none < little < moderate < rich
- Saving accounts:
- Target variable encoded using label encoding
- Train-test split applied after encoding
The following models were trained and evaluated:
- Logistic Regression (baseline)
- Decision Tree
- Random Forest
- XGBoost
Evaluation focused on:
- Accuracy
- Precision & Recall (especially for Bad class)
- Classification analysis
- Confusion matrix
Instead of relying on the default 0.5 threshold, custom threshold tuning was applied to improve detection of bad credit applicants, which is critical in financial risk scenarios.
- Model probabilities were analyzed
- Threshold adjusted to optimize Bad-class recall
- Final decision threshold chosen based on business risk tolerance
| Metric | Bad (0) | Good (1) |
|---|---|---|
| Recall | 68% | 74% |
| Precision | 53% | 85% |
| Overall Accuracy | 73% |
This configuration achieves a strong balance between identifying risky applicants and maintaining loan approval quality.
XGBoost feature importance highlights:
- Credit Amount
- Age
- Duration
- Checking Account
- Saving Account
- Job
- Purpose
This supports model interpretability and trust.
| Policy | Approved | Expected Defaults | Default Rate |
|---|---|---|---|
| Approve All | 200 | 60 | 30.0% |
| Low + Medium Risk | 181 | 46 | 25.41% |
| Low Risk Only | 146 | 28 | 19.18% |
Default Rate:
The proportion of approved loan applications that are classified as bad risk, indicating loans that are expected to result in default or serious repayment issues.
- Risk banding enables controlled reduction in default rates
- Banks can choose policies based on risk appetite
- Demonstrates how ML outputs translate into real-world decisions
- Accuracy alone is insufficient in credit risk modeling
- Threshold tuning is critical for aligning models with business goals
- Risk banding adds significant real-world value
- Simple datasets still allow strong insights when approached methodically
- Python
- Pandas, NumPy
- Matplotlib, Seaborn
- Scikit-learn
- XGBoost
Through interpretability-driven modeling, threshold optimization, and policy simulation, this project bridges the gap between data science and practical credit risk management. It reflects workflows used in modern financial analytics teams.
This project focuses on building an interpretable and business-oriented credit risk model using a limited set of applicant-level features. Possible future enhancements include:
-
Richer Feature Availability
- Incorporating additional financial attributes such as income, existing liabilities, or historical repayment behavior could improve risk discrimination.
- With the current dataset, feature engineering options are intentionally limited.
-
Policy-Level Optimization
- Further analysis could be performed to identify optimal approval thresholds based on different business objectives, such as maximizing approved volume under a fixed default-rate constraint.
-
Out-of-Sample Validation
- Testing the framework on a larger or more recent credit dataset would help assess model stability and generalizability across populations.
-
Deployment-Oriented Monitoring
- In a real-world setting, model performance and default rates would need continuous monitoring to detect performance drift over time.
These extensions focus on data and decision-level improvements, which are often more impactful than additional model complexity in practical credit risk systems.