QuantPulse BTC

The Problem

- Quantitative trading strategy development typically happens in fragmented Jupyter Notebooks, making it difficult to simulate realistic market conditions like latency, execution slippage, and trading fees. - Standard sentiment analysis tools rely on generic NLP text models that fail to correctly capture domain-specific context in financial and cryptocurrency news headlines. - Transitioning historical trading models from backward-looking backtest engines to asynchronous live-paper trading simulations usually requires completely rewriting the codebase.

The Solution

Built a unified, asynchronous platform using FastAPI and PostgreSQL to bridge the gap between financial research and production engineering. Implemented an event-driven backtesting loop that replays historical candle data using vectorized Pandas operations while enforcing realistic fill logic, tracking portfolio balance states dynamically via SQLAlchemy ORM. Integrated an automated NLP ingestion pipeline utilizing a Hugging Face FinBERT model running as decoupled background tasks to score crypto news sentiment streams without blocking the main dashboard. Additionally, incorporated tree-based machine learning models (XGBoost Classifier/Regressor) to extract predictive regime-shifting and volatility insights from technical market indicators.

Key Features

• ✔ Event-driven historical candle replay engine with transaction fee and slippage modeling.
• ✔ Asynchronous paper trading simulator using a relational database persistence layer.
• ✔ News ingestion pipeline with Hugging Face FinBERT domain-specific sentiment scoring.
• ✔ XGBoost machine learning predictive models for market volatility and price direction.
• ✔ Dynamic data visualization dashboard powered by FastAPI endpoints and Chart.js graphics.
• ✔ Modular rule-based strategy interface built for extensible feature discovery.

Challenges Faced

- Challenge: Running heavy Transformer deep-learning inference models inside web request paths froze the single-threaded event loop. - Solution: Decoupled the FinBERT processing pipeline by wrapping news fetch routines into an isolated async background scheduler that commits pre-computed results directly to the Postgres layer. - Challenge: Handling messy time-series data streams risked producing data leakage or look-ahead bias during training indicator generation phases. - Solution: Enforced absolute temporal boundaries using specialized logical limits within Pandas indexing blocks alongside rolling window methods, strictly validating that feature state 't' only pulls historical arrays spanning 0 to t-1.