Skip to main content

Quant Model

Fraud Risk Analyzer — Bayesian Network

Interactive Bayesian Belief Network for transaction fraud detection. Models probabilistic relationships between transaction attributes to calculate fraud risk scores with explainable factor contributions.

Methods

Python, Bayesian Inference, pgmpy, Flask, JavaScript

Impact

Demonstrates probabilistic graphical modeling for real-world risk scoring—translating uncertain evidence into actionable fraud probabilities.

A Bayesian Belief Network implementation for financial fraud risk scoring. Unlike black-box ML models, this approach uses probabilistic graphical modeling to capture domain knowledge about fraud indicators and their interdependencies.

What It Models

The network reasons about:

  • Account Age — New accounts carry higher baseline risk
  • Past Fraud History — Previous fraud is highly predictive
  • Transaction Amount — High-value transactions warrant scrutiny
  • Time of Day — Late-night activity correlates with fraud
  • Location — International transactions increase risk
  • Transaction Velocity — Rapid-fire transactions signal automation

Technical Approach

  • Bayesian Inference — Updates prior fraud beliefs with transaction evidence
  • Explainable Scoring — Shows which factors drive the risk assessment
  • Scenario Testing — Quick presets for common fraud patterns:
    • Routine Purchase (Low Risk ~2%)
    • Suspicious International (High Risk ~65%)
    • Late Night High-Value (Medium-High ~45%)
    • New Account Velocity Spike (Very High ~85%)

Why Bayesian Networks for Fraud

Traditional ML models can achieve high accuracy but lack transparency. Bayesian networks explicitly model causal relationships and provide:

  • Clear factor contribution breakdowns
  • Natural handling of missing evidence
  • Domain knowledge integration
  • Uncertainty quantification

This project demonstrates the quant analyst skill of building interpretable risk models for high-stakes decisions.