Files
gogo2/reports/AGGRESSIVE_TRADING_THRESHOLDS_SUMMARY.md
2025-07-02 00:52:50 +03:00

65 lines
2.7 KiB
Markdown

# Aggressive Trading Thresholds Summary
## Overview
Lowered confidence thresholds across the entire trading system to execute trades more aggressively, generating more training data for the checkpoint-enabled models.
## Changes Made
### 1. Clean Dashboard (`web/clean_dashboard.py`)
- **CLOSE_POSITION_THRESHOLD**: `0.25``0.15` (40% reduction)
- **OPEN_POSITION_THRESHOLD**: `0.60``0.35` (42% reduction)
### 2. DQN Agent (`NN/models/dqn_agent.py`)
- **entry_confidence_threshold**: `0.7``0.35` (50% reduction)
- **exit_confidence_threshold**: `0.3``0.15` (50% reduction)
### 3. Trading Orchestrator (`core/orchestrator.py`)
- **confidence_threshold**: `0.20``0.15` (25% reduction)
- **confidence_threshold_close**: `0.10``0.08` (20% reduction)
### 4. Realtime RL COB Trader (`core/realtime_rl_cob_trader.py`)
- **min_confidence_threshold**: `0.7``0.35` (50% reduction)
### 5. Training Integration (`core/training_integration.py`)
- **min_confidence_threshold**: `0.3``0.15` (50% reduction)
## Expected Impact
### More Aggressive Trading
- **Entry Thresholds**: Now require only 35% confidence to open new positions (vs 60-70% previously)
- **Exit Thresholds**: Now require only 8-15% confidence to close positions (vs 25-30% previously)
- **Overall**: System will execute ~2-3x more trades than before
### Better Training Data Generation
- **More Executed Actions**: Since we now store training progress, more executed trades = more training data
- **Faster Learning**: Models will learn from real trading outcomes more frequently
- **Split-Second Decisions**: With 100ms training intervals, models can adapt quickly to market changes
### Risk Management
- **Position Sizing**: Small position sizes (0.005) limit risk per trade
- **Profit Incentives**: System still has profit-based incentives for closing positions
- **Leverage Control**: User-controlled leverage settings provide additional risk management
## Training Frequency
- **Decision Fusion**: Every 100ms
- **COB RL**: Every 100ms
- **DQN**: Every 30 seconds
- **CNN**: Every 30 seconds
## Monitoring
- Training performance metrics are tracked and displayed
- Average, min, max training times are logged
- Training frequency and total calls are monitored
- Real-time performance feedback available in dashboard
## Next Steps
1. Monitor trade execution frequency
2. Track training data generation rate
3. Observe model learning progress
4. Adjust thresholds further if needed based on performance
## Notes
- All changes maintain the existing profit incentive system
- Position management logic remains intact
- Risk controls through position sizing and leverage are preserved
- Training checkpoint system ensures progress is not lost