ML Transaction Prediction Project

Data Requirements:
- Accumulate historical transaction data with fraud labels from core banking systems
- Store transaction data in a distributed data store optimized for ML (e.g graph DB or in memory DB)
- Implement streaming architecture to ingest and extract features from real-time transactions

Modeling Requirements:
- Train supervised models for both labeled and unlabeled data to predict fraud probabilities
- Train unsupervised models like autoencoders to detect anomalous transactions
- Develop rule-based models to encode human expert knowledge and handle high velocity features
- Perform feature engineering to extract relevant attributes from transaction data

Model Deployment Requirements:
- Containerize models and standardize APIs for real-time prediction
- Build horizontally scalable serving infrastructure to handle high throughput
- Optimize models for low latency predictions (e.g. quantization)
- Implement failover and redundancies for high availability

Monitoring Requirements:
- Log predictions and performance metrics like accuracy, latency and drift
- Visualize trends and model metrics on monitoring dashboards
- Retrain models periodically with updated data
- Implement alerts for model degradation and failures

Integration Requirements:
- Expose model APIs for integration with core banking systems
- Ingest real-time transaction data from upstream message queues/Kafka
- Propagate predicted fraud probabilities and alerts to downstream systems
- Secure API endpoints and data transfers with encryption, authentication, RBAC

Security requirements:
- The entire system should be built and tested against OWAPS top 10