Project Objective

Given attributes about a loan applicant, decide the credibility of the borrower with the Probability of Default (PD) and financial stress score and justify the decision using XAI techniques.

Dataset: German Credit Data (UCI Machine Learning Repository)

Stack: Python, Pandas, Plotly, Scikit-learn, XGBoost, SHAP, Streamlit, MLflow , Statsmodels, SciPy, Jupyter notebook

---

Run Locally

# Clone repo
git clone https://github.com/Abhinav289/credit-risk-project.git
cd credit-risk-project

# Create environment
conda create -n credit_risk python=3.11
conda activate credit_risk

# Install dependencies
pip install -r requirements.txt

# Run Streamlit app
cd app
streamlit run streamlit_app.py

---

Key EDA Findings

Default Borrower Profile

A borrower most likely to default shows:

Feature	High Risk Signal
Checking account	Negative or absent balance
Credit history	No credits / all paid
Savings	Less than 100 DM
Property	No known collateral
Other parties	Co-applicant
Duration	45+ months
Age	25–30 years

Hypothesis Testing

All features validated with statistical tests (α = 0.05):

• Mann-Whitney U for numeric features
• Chi-Square for categorical features
• Wilcoxon Signed-Rank for model comparison

Feature Engineering

3 domain-driven features engineered:

Feature	Logic	Impact
financial_stress_score	checking + savings stress combined	Score 0→6: 8.87%→52.05% default rate
monthly_burden	credit_amount / duration	Captures repayment affordability
high_risk_interaction	negative checking AND no credit history	69.12% default rate when flagged

---

Design Patterns Applied

Pattern	Implementation
Rebalancing	SMOTE applied on training set only — never on test
Reframing	Probability calibration via CalibratedClassifierCV
Checkpoints	XGBoost early stopping (stopped at round 107)
Explainable Predictions	SHAP waterfall per individual prediction
Repeatable Splitting	Stratified train-test split

---

Experiment tracked with MLflow:

Parameter	Value
Model	XGBoost + StandardScaler
n_estimators	107 (early stopping)
max_depth	4
learning_rate	0.1
Calibration	Isotonic (cv=5)
Threshold	0.612

---

Performance

Model	CV ROC-AUC	F1 (threshold=0.612)
Logistic Regression (baseline)	0.856 ± 0.013	0.55
XGBoost	0.890 ± 0.019	0.614

Statistical comparison: Wilcoxon Signed-Rank test (p=0.0625)
XGBoost outperformed LR on every single fold.
Threshold optimized at 0.612 to maximize F1 score.

---

References

• Hofmann, H. (1994). German Credit Data. UCI ML Repository
• Lakshmanan et al. — Machine Learning Design Patterns (O'Reilly)

• Demšar, J. (2006). Statistical Comparisons of Classifiers