This project demonstrates how machine learning can be applied to detect fraudulent credit card transactions in a real-world banking scenario.
Fraud detection is highly imbalanced: fraudulent transactions make up less than 0.2% of all data.
The challenge is to catch fraud while minimizing false alerts that overwhelm analysts and disrupt legitimate customers.

• 284,807 anonymized transactions from Kaggle: Credit Card Fraud Detection.
• Features V1...V28 are PCA-transformed for confidentiality.
• Added features Time and Amount.

1. Data Wrangling & Preprocessing

   • Checked for nulls, validated schema.
   • Scaled Amount and Time.
   • Stratified train/validation split to preserve fraud ratio.

2. Modeling

   • Logistic Regression (baseline).
   • Random Forest with calibrated probabilities.
   • Logistic Regression with SMOTE (oversampling minority class).

3. Evaluation

   • Metrics: Precision, Recall, F1, PR-AUC, ROC-AUC.
   • Focused on precision/recall trade-offs instead of accuracy.
   • Threshold tuning to balance fraud detection with false alert reduction.

4. Interpretability

   • Logistic Regression coefficients (top predictive components).
   • Random Forest feature importances.
   • Business-friendly explanations of model signals.

5. Deployment Awareness

   • Saved trained models and scaler with joblib.
   • Produced a comparison table for business decision-making.

The Random Forest model achieves a strong balance between recall (fraud caught) and precision (alerts that are correct).
The chart below shows how precision increases as recall decreases when the decision threshold is adjusted.

Key Insights:

• Logistic Regression: high recall (85%) but very noisy (52% precision).
• Random Forest: best balance — ~81% recall with ~90% precision.
• SMOTE + Logistic: better than baseline LR, but not as strong as Random Forest.

Business Takeaway:
A tuned Random Forest reduces false alerts by >90% while still catching >80% of fraud cases.
This makes fraud alerts more trustworthy for analysts and reduces unnecessary customer disruptions.

1. Download dataset from Kaggle.
2. Place creditcard.csv in your working directory.
3. Install requirements:

   pip install -r requirements.txt