How andrzejmiskow/tradeai-advancing-algorithmic-trading-systems-with-time-series-transformer-for-cryptocurrency-data works
Project type and purpose
• This is a full-stack algorithmic trading research system (“TradeAI”) for cryptocurrency data, combining:
• A React-based frontend UI (in UI/)
• A Python backend service serving ML predictions & trading logic (in prediction_service/)
• Data pipelines and experiments for model training (in src/)
• Versioned datasets managed by DVC (in data/)
• Containerization for local and Azure deployment (Dockerfile, docker-compose.yml)
File structure
• Top-level:
• UI/: React app for visualization and user controls
• prediction_service/: Flask/FastAPI-style service hosting ML models, DRL execution, portfolio management, test clients
• src/: Offline pipeline—data ingestion, feature building, model training, evaluation, notebooks, experiments
• data/: DVC-tracked raw and processed data organized by sampling frequency and split (train/test)
• docker-compose.yml, Dockerfile: orchestrate UI, API, potential data volumes
• Patterns:
• Separation of concerns: frontend vs. service vs. offline pipelines
• Use of DVC for data versioning and reproducible pipelines
• MLflow (train_and_evaluate_mlflow_template.py) for experiments tracking
• Transformer and DRL strategy modules under prediction_service/transformers & drl_strategy.py
README for additional insights
• Purpose: investigate transformer models + DRL for medium-frequency crypto trading
• Technologies: React, Flask (or similar), PyTorch/TensorFlow transformers, DVC, Docker, Azure
• Architecture notes: full-stack, microservice-friendly, MLOps aspirations
Detail
• Interactions (e.g., “DVC pull/push”, “MLflow logging”, “Docker build & deploy”)
• Avoid over-aggregation: break the backend into prediction, execution, portfolio, data handlers