Credit Scoring Model (Internship Task)

This project implements an end-to-end credit scoring workflow with:

• model training (logistic, decision_tree, random_forest)
• model evaluation and metrics export
• FastAPI inference service (/predict)
• Streamlit dashboard for monitoring + manual predictions

Project Structure

• data/german_credit_data.csv - input dataset
• src/train_credit_model.py - training + evaluation pipeline
• src/api.py - FastAPI prediction API (uses models/random_forest.joblib)
• app/dashboard.py - Streamlit monitoring and prediction UI
• models/ - trained model artifacts
• reports/ - exported metrics per model
• requirements.txt - dependencies

Target Label Note

The dataset may not always include a true repayment/default label.

Supported modes:

1. Real target mode (recommended): provide a valid target column such as Risk.
2. Demo mode: generate a proxy label with --create-proxy-target using Credit amount and --proxy-threshold.

Proxy mode is for learning only and is not production-grade risk modeling.

Setup

Install dependencies:

pip install -r requirements.txt

Train Models

Logistic regression (real target)

python src/train_credit_model.py --data-path data/german_credit_data.csv --target-column Risk --model-type logistic

Decision tree (proxy target demo)

python src/train_credit_model.py --data-path data/german_credit_data.csv --create-proxy-target --model-type decision_tree

Random forest (proxy target demo)

python src/train_credit_model.py --data-path data/german_credit_data.csv --create-proxy-target --model-type random_forest

Optional threshold tuning at train time

python src/train_credit_model.py --data-path data/german_credit_data.csv --create-proxy-target --model-type random_forest --decision-threshold 0.45

Artifacts Generated

Per model:

• models/logistic.joblib
• models/decision_tree.joblib
• models/random_forest.joblib

Per model metrics:

• reports/metrics_logistic.json
• reports/metrics_decision_tree.json
• reports/metrics_random_forest.json

Run Inference API

Start FastAPI:

python src/api.py

Health check:

curl http://127.0.0.1:8000/

Predict endpoint:

curl -X POST http://127.0.0.1:8000/predict -H "Content-Type: application/json" -d "{\"Age\":35,\"Sex\":\"male\",\"Job\":2,\"Housing\":\"own\",\"Saving_accounts\":\"little\",\"Checking_account\":\"moderate\",\"Credit_amount\":5000,\"Duration\":24,\"Purpose\":\"radio/TV\"}"

Response shape:

{
  "risk_probability": 0.3124,
  "decision": "Approve"
}

Run Dashboard

Start Streamlit:

streamlit run app/dashboard.py

The dashboard supports:

• model selection (logistic, decision_tree, random_forest)
• threshold control from sidebar
• performance cards and confusion matrix
• class distribution and CV stability view
• manual prediction form (calls API at http://127.0.0.1:8000/predict)
• downloadable metrics JSON and HTML snapshot report

Evaluation Metrics

train_credit_model.py exports:

• Accuracy
• Precision
• Recall
• F1
• ROC-AUC
• PR-AUC
• Confusion Matrix
• Classification Report
• CV ROC-AUC mean/std

Recommended Local Run Order

1. Train at least one model (preferably all 3 for dashboard switching).
2. Start API server: python src/api.py.
3. Start dashboard: streamlit run app/dashboard.py.
4. Open Streamlit UI and test the Prediction page.