AI-Augmented Financial Anomaly Detection

CFA × ML: Domain expertise meets machine learning for fraud detection. Built by a CFA charterholder who writes Python.

Key Results

Metric	Value
XGBoost AUC	0.987 (+8.9pp over CFA rule-based baseline)
CFA Rule-Based AUC	0.898 (surprisingly strong baseline)
CFA features in top 20 SHAP	8 of 20
Adversary-resistant floor	81% on synthetic data (system-controlled features only)
Data	100K synthetic PaySim transactions

Core insight: [On synthetic PaySim data] Domain expertise encoded as rules is a floor, not a ceiling. ML adds non-linear interaction detection that rules miss, but the CFA-informed features dominate SHAP importance rankings. Real financial data with adversarial dynamics would likely show different ratios; this project demonstrates the methodology (controllability analysis applied to fraud), not production-ready thresholds.

Quick Start

git clone https://github.com/rexcoleman/financial-anomaly-detection.git
cd financial-anomaly-detection
conda env create -f environment.yml
conda activate fin-anomaly

# Run full pipeline
python scripts/run_pipeline.py --seed 42

# Launch interactive dashboard
streamlit run app.py

Architecture

src/
  detection/          # Fraud detection models (XGBoost, Isolation Forest)
  features/           # CFA-informed feature engineering
  explainability/     # SHAP analysis
  core/               # Types and utilities
scripts/
  run_pipeline.py               # Full pipeline: data -> features -> models -> SHAP -> ACA
  generate_synthetic_data.py    # PaySim-style synthetic transactions
  generate_figures.py           # Publication-ready charts
app.py                          # Streamlit interactive dashboard

Methodology

This project validates the adversarial controllability analysis methodology (5th domain). Transaction features are classified by who controls them:

• System-controlled: account history, transaction frequency, institutional flags — 81% detection floor (on synthetic data)
• Attacker-controlled: transaction amount, timing, merchant category — manipulable by adversaries

See FINDINGS.md for detailed results.

Governed by govML

Built with reproducibility and decision traceability enforced across the entire pipeline.

License

MIT