Financial Compliance Monitoring

Problem

OUC has thousands of transactions monthly that need to be monitored for compliance.

Data

The system processes transaction metadata including amounts, timestamps, merchant categories, and transaction types.

Approach

A hybrid approach combining gradient boosting models (XGBoost, LightGBM) as well as using anomaly detection models (Isolation Forest, Autoencoders) to identify unusual patterns in transaction flows.

Evaluation

Models are evaluated using precision, recall, F1-score, and false-positive rate. Before ML implementation, there was manual monitoring. After development, there has been an achieved weighted F1 score of 80%+.

Precision
83%
+23% from baseline
False-Positive Rate
1.2%
-23% from baseline
True Negatives
98.8%
Recall
72%
Maintained

Deployment

Infrastructure: Snowflake ML infrastructure.

Monitoring: Real-time monitoring of financial transactions with the intention to trigger human intervention as applicable.

Latency: Very low.

Impact

The ML system exists and is still in development.

What I Supported

  • Problem definition and scoping
  • Data exploration and feature engineering
  • Model development and training
  • Model evaluation and validation
  • Performance optimization and iteration
  • Documentation and knowledge sharing

What I Collaborate On

  • Infrastructure design with platform teams
  • Compliance requirements with compliance team
  • Product integration with data teams
  • Stakeholder alignment and prioritization
  • Model interpretation with compliance SMEs
  • User feedback collection and analysis

Safety & Compliance Checklist

  • Explainability: Feature importance analysis
  • Approvals: Model review process with stakeholders and compliance
  • Monitoring: Monthly review of F1 and false-positive rate
  • Documentation: Documentation for future iteration