Ensemble Machine Learning Architecture
Swiss-Quant cryptocurrency forecasts are generated by an ensemble of XGBoost and LightGBM gradient boosting models, optimized through BayesSearchCV hyperparameter tuning. The system processes over 40 features extracted from multiple data sources to produce directional predictions with confidence scores for Bitcoin, Ethereum, Solana, Cardano, XRP, Dogecoin, and other major cryptocurrencies.
Feature Engineering Pipeline
The model ingests real-time OHLCV data at 15-minute, 1-hour, 4-hour, and daily timeframes. Technical features include RSI (14-period), MACD signal crossovers (12/26/9), Bollinger Band width and %B position, Stochastic oscillator (14,3), ADX trend strength, CCI momentum, and EMA crossover signals across 9/21/50/200-period moving averages. Each indicator is transformed into a continuous gradient score from -100 to +100, providing nuanced signal strength rather than binary buy/sell triggers.
On-Chain and Sentiment Integration
Beyond price-based features, the crypto forecast incorporates on-chain metrics including Bitcoin hash rate, mining difficulty, mempool congestion, exchange net flows, and the Fear and Greed Index. BTC dominance ratio and total crypto market capitalization momentum serve as regime indicators, helping the model adapt to risk-on versus risk-off environments.
Walk-Forward Validation
All predictions are validated using walk-forward testing on 200+ out-of-sample periods with purged cross-validation and a 2-day embargo gap to prevent data leakage. Models are retrained weekly on rolling 180-day windows to adapt to evolving market microstructure. Forecasts are generated 4 times daily at 06:00, 11:00, 16:00, and 21:00 CET.