Detect fraudulent credit card transactions in an extremely imbalanced dataset where fraud is rare. The key challenge is that accuracy is misleading: a model can achieve ~99.8% accuracy by predicting “not fraud” for every transaction while detecting zero fraud cases.

• Build a robust baseline fraud detection pipeline
• Evaluate models with metrics suitable for class imbalance (Precision, Recall, F1, PR-AUC)
• Perform decision threshold tuning based on business constraints (high recall vs. alert quality vs. cost minimization)
• Provide error analysis and a business-ready executive summary

This repository is structured like a real analytics / ML project:

• src/ contains reusable Python modules (metrics, thresholding, evaluation)
• notebooks/ contain narrative analysis (data understanding → modeling → business impact)
• reports/ contains an executive summary for non-technical stakeholders

Classic Kaggle credit card fraud dataset:

• Features: V1..V28, Time, Amount
• Target: Class (1 = fraud, 0 = normal)

Note: The dataset is not committed to GitHub.
Place creditcard.csv in data/raw/.

• data/ raw and processed data (ignored by git)
• notebooks/ narrative analysis notebooks
• src/ reusable Python modules (project core)
• reports/ executive summary and figures

1. Data quality & imbalance analysis

   • Class ratio, missing values, feature distributions
   • Demonstration of the “accuracy trap” in fraud detection

2. Baseline modeling

   • Logistic Regression with class weighting
   • Evaluation using Precision, Recall, F1, and PR-AUC

3. Threshold tuning

   • Selection of decision thresholds based on:
     • minimum precision
     • minimum recall
     • minimum expected business cost (false negative vs. false positive trade-off)

4. Error analysis

   • Detailed inspection of false positives and false negatives
   • Translation of model errors into business impact (alert volume, missed fraud loss)

• Precision, Recall, F1-score
• PR-AUC (Average Precision) — preferred under extreme class imbalance
• Confusion Matrix at the selected decision threshold

1. Create a virtual environment and install dependencies:

pip install -r requirements.txt

2. Place the dataset: data/raw/creditcard.csv

3. Run notebooks in order:

• notebooks/01_data_understanding.ipynb
• notebooks/02_modeling_thresholding.ipynb
• notebooks/03_error_analysis_business.ipynb

• Reusable evaluation and thresholding toolkit in src/
• Three notebooks with a clear analytical narrative
• reports/executive_summary.md with decision-ready conclusions

• Why accuracy fails in fraud detection
• Why PR-AUC and precision/recall are more informative
• How the decision threshold was selected using business cost trade-offs
• What patterns were observed in false positives / false negatives
• What would be done next (stronger models, calibration, monitoring)

🔙 Back to Portfolio