gogo2/MULTI_HORIZON_TRAINING_SYSTEM.md

Multi-Horizon Training System Documentation

Overview

The Multi-Horizon Training System addresses the core issues with your current training approach:

Problems with Current System

1. Immediate Training: Training happens right after trades close (couple seconds), often before meaningful price movement
2. No Profit Potential: Small timeframes don't provide enough movement for profitable trades
3. Reactive Training: Models learn from very short-term outcomes rather than longer-term patterns
4. Limited Prediction Horizons: Only predicts short timeframes that may not capture meaningful market moves

New System Benefits

1. Multi-Timeframe Predictions: Predicts 1m, 5m, 15m, and 60m horizons every minute
2. Deferred Training: Stores predictions and trains models when outcomes are actually known
3. Min/Max Price Prediction: Focuses on predicting price ranges over longer periods for better profit potential
4. Backtesting Capability: Can validate system performance on historical data
5. Scalable Storage: Efficiently stores model inputs for future training

System Components

1. MultiHorizonPredictionManager (core/multi_horizon_prediction_manager.py)

• Generates predictions for 1, 5, 15, and 60-minute horizons every minute
• Uses ensemble approach combining CNN, RL, and technical analysis
• Stores prediction snapshots with full model inputs for future training

Key Features:

• Real-time prediction generation
• Confidence-based filtering
• Automatic validation when target times are reached

2. PredictionSnapshotStorage (core/prediction_snapshot_storage.py)

• Efficiently stores prediction snapshots to disk
• SQLite metadata database with compression
• Batch retrieval for training
• Automatic cleanup of old data

Storage Structure:

• Compressed pickle files for snapshot data
• SQLite database for fast metadata queries
• Organized by symbol and prediction horizon

3. MultiHorizonTrainer (core/multi_horizon_trainer.py)

• Trains models when prediction outcomes are known
• Handles both CNN and RL model training
• Uses stored snapshots to recreate training scenarios

Training Process:

• Validates pending predictions against actual price data
• Trains models using historical prediction accuracy
• Supports batch training for efficiency

4. MultiHorizonBacktester (core/multi_horizon_backtester.py)

• Backtests prediction accuracy on historical data
• Validates system performance before deployment
• Provides detailed accuracy and profitability analysis

Backtesting Features:

• Historical data simulation
• Accuracy metrics by prediction horizon
• Profitability analysis
• Performance reporting

5. Enhanced DataProvider (core/data_provider.py)

• Added get_price_range_over_period() method
• Supports min/max price queries over specific time ranges
• Better integration with backtesting framework

Usage Examples

Running the System

# Run demonstration
python run_multi_horizon_training.py --mode demo

# Run backtest on 7 days of data
python run_multi_horizon_training.py --mode backtest --symbol ETH/USDT --days 7

# Force training session
python run_multi_horizon_training.py --mode train --horizon 60

# Run system for 5 minutes
python run_multi_horizon_training.py --mode run --runtime 300

Integration with Existing Code

from core.multi_horizon_prediction_manager import MultiHorizonPredictionManager
from core.prediction_snapshot_storage import PredictionSnapshotStorage
from core.multi_horizon_trainer import MultiHorizonTrainer

# Initialize components
prediction_manager = MultiHorizonPredictionManager(orchestrator=your_orchestrator)
snapshot_storage = PredictionSnapshotStorage()
trainer = MultiHorizonTrainer(orchestrator=your_orchestrator, snapshot_storage=snapshot_storage)

# Start the system
prediction_manager.start()
trainer.start()

# Get system status
status = prediction_manager.get_prediction_stats()
training_stats = trainer.get_training_stats()

Prediction Horizons

The system generates predictions for four horizons:

• 1 minute: Very short-term predictions for scalping
• 5 minutes: Short-term momentum predictions
• 15 minutes: Medium-term trend predictions
• 60 minutes: Long-term range predictions (focus area for meaningful moves)

Each prediction includes:

• Predicted minimum price
• Predicted maximum price
• Confidence score
• Model inputs for training
• Market state snapshot

Training Strategy

When Training Occurs

• Predictions are generated every minute
• Models are trained when prediction target times are reached (1-60 minutes later)
• Training uses the full context available at prediction time
• Rewards are based on prediction accuracy within the predicted price range

Model Types Supported

1. CNN Models: Trained on feature sequences to predict price ranges
2. RL Models: Trained with reinforcement learning on prediction outcomes
3. Ensemble: Combines multiple model predictions for better accuracy

Backtesting and Validation

Backtesting Process

1. Load historical 1-minute data
2. Simulate predictions at regular intervals
3. Wait for target time to check actual outcomes
4. Calculate accuracy and profitability metrics

Key Metrics

• Range Accuracy: How well predicted min/max ranges match actual ranges
• Confidence Correlation: How confidence scores relate to prediction accuracy
• Profitability: Simulated trading performance based on predictions

Performance Analysis

Expected Improvements

1. Better Profit Potential: 60-minute predictions allow for meaningful price moves
2. More Stable Training: Training occurs on known outcomes, not immediate reactions
3. Reduced Overfitting: Multi-horizon approach prevents overfitting to short-term noise
4. Backtesting Validation: Historical testing ensures system robustness

Monitoring

The system provides comprehensive monitoring:

• Prediction generation rates
• Training session statistics
• Model accuracy by horizon
• Storage utilization
• System health metrics

Configuration

Key Parameters

# Prediction horizons (minutes)
horizons = [1, 5, 15, 60]

# Prediction frequency
prediction_interval_seconds = 60

# Minimum confidence for storage
min_confidence_threshold = 0.3

# Training batch size
batch_size = 32

# Storage retention
max_age_days = 30

File Locations

• Prediction snapshots: data/prediction_snapshots/
• Backtest results: reports/
• Cache data: cache/

Integration with Existing Dashboard

The system is designed to integrate with your existing dashboard:

1. Real-time Monitoring: Dashboard can display prediction generation stats
2. Training Progress: Show training session results
3. Backtest Reports: Display historical performance analysis
4. Model Comparison: Compare old vs new training approaches

Migration Path

Gradual Adoption

1. Run in Parallel: Run new system alongside existing training
2. Compare Performance: Use backtesting to compare approaches
3. Gradual Transition: Move models to new training system incrementally
4. Fallback Support: Keep old system as backup during transition

Data Compatibility

• New system stores snapshots independently
• Existing model weights can be used as starting points
• Training data format is compatible with existing models

Troubleshooting

Common Issues

1. Low Prediction Accuracy: Check confidence thresholds and feature quality
2. Storage Issues: Monitor disk space and cleanup old snapshots
3. Training Performance: Adjust batch sizes and learning rates
4. Memory Usage: Use appropriate cache sizes for your hardware

Logging

All components use structured logging with consistent log levels:

• INFO: Normal operations and results
• WARNING: Potential issues that don't stop operation
• ERROR: Serious problems requiring attention

Future Enhancements

Planned Features

1. Advanced Ensemble Methods: More sophisticated model combination
2. Adaptive Horizons: Dynamic horizon selection based on market conditions
3. Cross-Symbol Training: Train models using data from multiple symbols
4. Real-time Validation: Immediate feedback on prediction quality
5. Performance Optimization: GPU acceleration and distributed training

Research Directions

1. Optimal Horizon Selection: Which horizons provide best risk-adjusted returns
2. Market Regime Detection: Adjust predictions based on market conditions
3. Feature Engineering: Better input features for price range prediction
4. Uncertainty Quantification: Better confidence score calibration

Conclusion

The Multi-Horizon Training System addresses your core concerns by:

1. Extending Prediction Horizons: From seconds to 60 minutes for meaningful profit potential
2. Deferred Training: Models learn from actual outcomes, not immediate reactions
3. Comprehensive Storage: Full model inputs preserved for future training
4. Backtesting Validation: Historical testing ensures system effectiveness
5. Scalable Architecture: Efficient storage and training for long-term operation

This system should significantly improve your trading performance by focusing on longer-term, more profitable price movements while maintaining rigorous training and validation processes.