FastAPI backend that receives market candles from a MetaTrader 5 Expert Advisor, processes features, and serves predictions from an XGBoost model trained with walk-forward methodology and triple-barrier labeling.
The system automatically retrains weekly via APScheduler, runs containerized in Docker, and is only exposed on the local network — never directly to the internet. Training window sizing is dynamic, adjusted to the actual available history.
The project is explicitly treated as a validation of the data pipeline and infrastructure, not a profitability promise: the model's confidence threshold is calibrated conservatively while more validation history accumulates.
← Back to projects