Crypto Micro-Horizon Nowcasting (5-Minute)
Designed a real-time ML pipeline for short-term (5-minute) crypto price direction insights. Focused on live-style data ingestion, feature engineering, and model evaluation with walk-forward validation.
Objective
To develop a near-real-time prediction engine that generates actionable signals for crypto assets (e.g., BTC-USD) every 5 minutes. The model was built to simulate streaming inference, test latency handling, and validate predictive intervals under high-volatility conditions.
Pipeline Overview
- Data Source: Live market data (yfinance & exchange APIs) captured every 5 minutes.
- Feature Store: Rolling window metrics — returns, RSI, Bollinger Bands, volatility clustering.
- Modeling: Regression and classification ensemble (RandomForest, XGBoost, Ridge, LSTM baseline).
- Validation: Walk-forward expanding windows to preserve time dependency.
- Deployment: Modular Python pipeline with live-inference stub for next-step prediction.
Results
- Mean directional accuracy (MDA): 62–67% across major pairs (BTC, ETH, ADA).
- Feature importance indicated volatility and momentum lag features contributed most predictive power.
- Average inference latency: <1.5 seconds per 5-minute window update.
- Visualization module plotted confidence bands and live prediction streams for demo output.
