mlflow-credit-risk

Credit risk prediction pipeline with MLflow experiment tracking. Compares Logistic Regression, Random Forest, XGBoost, and LightGBM on the German Credit dataset.

Results

Model	Accuracy	F1	AUC-ROC
Logistic Regression	0.735	0.521	0.764
Random Forest	0.760	0.567	0.791
XGBoost	0.775	0.594	0.803
LightGBM	0.780	0.601	0.809

After Optuna tuning, XGBoost reaches 0.812 AUC-ROC.

How it works

1. Data ingestion - fetch German Credit dataset, validate, save locally
2. Feature engineering - StandardScaler + OneHotEncoder pipeline, derived features (amount_per_month, age bins, credit_load)
3. Training - 4 models trained with all params/metrics logged to MLflow
4. Evaluation - model comparison, threshold tuning, SHAP feature importance
5. Registry - best model registered and promoted to Production stage
6. Serving - FastAPI endpoint loads Production model from MLflow

Setup

pip install -e .

Usage

# download and prepare data
make ingest

# train all models (logs to MLflow)
make train

# compare models and analyze
make evaluate

# start MLflow UI
make mlflow-ui

# serve predictions
make serve

Tech

MLflow, XGBoost, LightGBM, scikit-learn, Optuna, SHAP, FastAPI