🤖 What Is This?
Machine Learning model performance and prediction analytics. Shows how the LightGBM models are performing, which features are most important, and prediction accuracy metrics. Your window into the AI's 'thinking'.
⚙️ How Data Gets Here (Automatic)
- Training: Weekly | Predictions: Every 15 min | Metrics: Real-time
- The
ML Training & Prediction Pipeline job runs automatically - Models train weekly on historical data (6 months rolling window)
- Features: Price momentum, volume, OI changes, Greeks, patterns, FII flow
- Generates predictions every 15 minutes during market hours
- Tracks prediction accuracy against actual outcomes (T+1 hour)
- Updates feature importance rankings
- Dashboard shows live model performance
📊 Understanding The Panels
Model Performance:
- Accuracy: % of correct directional predictions (> 55% = edge)
- Precision: Of all BUY signals, % that were profitable
- Recall: Of all actual opportunities, % we caught
- Sharpe Ratio: Risk-adjusted returns from model signals
Feature Importance:
- Top Features: Which inputs influence predictions most
- Feature Trends: How importance changes over time (regime-dependent)
- Model Confidence: Average confidence of recent predictions
💡 How To Use This
✅ High Confidence Indicators:- Model accuracy > 58% (statistically significant edge)
- Recent predictions (last 100) maintaining > 55% accuracy
- Prediction confidence > 70% on individual signals
⚠️ Caution Zones:- Accuracy dropping below 52% (model degradation, retrain needed)
- Low sample size (< 50 predictions) in current session
- Feature importance shifting dramatically (regime change)
💡 Pro Tips:- Check model performance before trusting signals - models degrade over time
- Higher confidence predictions (70%+) have better hit rates
- Feature importance tells you what market is focused on right now
- Compare live accuracy vs training accuracy to detect overfitting
💭 Remember: ML models are probabilistic, not deterministic. Even 60% accuracy means 40% losing trades. Use confidence scores as filters and always combine with other analysis.