What are the "16 trained models"? — Crypto ML Edge trains a separate LightGBM model for each coin (BTC, ETH, SOL, etc.) at each timeframe (1h, 4h). That means: 8 coins × 2 timeframes = 16 models. Each model is a specialist — the BTC-1h model only knows Bitcoin hourly patterns, while the SOL-4h model only knows Solana 4-hour patterns. Specialists beat generalists.
What is LightGBM? — LightGBM (Light Gradient Boosting Machine) is similar to XGBoost but faster. It builds decision trees one leaf at a time instead of one level at a time, making it better at finding complex patterns with less computing power. Think of it as a faster, more efficient version of XGBoost.
How does it pick features? — The system uses SHAP (SHapley Additive exPlanations) to automatically figure out which indicators actually matter for each coin. For BTC, maybe RSI and volume ratio are most important. For DOGE, maybe social sentiment matters more. SHAP removes useless features so the model doesn't get confused by noise.
How do picks appear here? — Every 30 min: fetch latest data → run all 16 models → each model outputs a confidence score (0-100%) → apply DSR validation gate → picks above threshold become active with TP/SL levels. The "Edge" tab shows model-driven picks. The "Quick" tab shows simpler rule-based signals that don't use ML.
Walk-forward validation: — Models are trained using a technique where they're always tested on data they've never seen. Imagine teaching someone with chapters 1-8 of a textbook, then testing them on chapter 9. This prevents "overfitting" (memorizing history instead of learning real patterns).
Loading dashboard data...