COB WS fix
This commit is contained in:
Dobromir Popov
@ -1369,15 +1369,31 @@ class TradingOrchestrator:
|
||||
reasoning['models_aggregated'] = [pred.model_name for pred in predictions]
|
||||
reasoning['aggregated_confidence'] = best_confidence
|
||||
|
||||
# Apply confidence thresholds for signal confirmation
|
||||
# Calculate dynamic aggressiveness based on recent performance
|
||||
entry_aggressiveness = self._calculate_dynamic_entry_aggressiveness(symbol)
|
||||
|
||||
# Adjust confidence threshold based on entry aggressiveness
|
||||
# Higher aggressiveness = lower threshold (more trades)
|
||||
# entry_aggressiveness: 0.0 = very conservative, 1.0 = very aggressive
|
||||
base_threshold = self.confidence_threshold
|
||||
aggressiveness_factor = 1.0 - entry_aggressiveness # Invert: high agg = low factor
|
||||
dynamic_threshold = base_threshold * aggressiveness_factor
|
||||
|
||||
# Ensure minimum threshold for safety (don't go below 1% confidence)
|
||||
dynamic_threshold = max(0.01, dynamic_threshold)
|
||||
|
||||
# Apply dynamic confidence threshold for signal confirmation
|
||||
if best_action != 'HOLD':
|
||||
if best_confidence < self.confidence_threshold:
|
||||
logger.debug(f"Signal below confidence threshold: {best_action} {symbol} "
|
||||
f"(confidence: {best_confidence:.3f} < {self.confidence_threshold})")
|
||||
if best_confidence < dynamic_threshold:
|
||||
logger.debug(f"Signal below dynamic confidence threshold: {best_action} {symbol} "
|
||||
f"(confidence: {best_confidence:.3f} < {dynamic_threshold:.3f}, "
|
||||
f"base: {base_threshold:.3f}, aggressiveness: {entry_aggressiveness:.2f})")
|
||||
best_action = 'HOLD'
|
||||
best_confidence = 0.0
|
||||
reasoning['rejected_reason'] = 'low_confidence'
|
||||
else:
|
||||
logger.info(f"SIGNAL ACCEPTED: {best_action} {symbol} "
|
||||
f"(confidence: {best_confidence:.3f} >= {dynamic_threshold:.3f}, "
|
||||
f"aggressiveness: {entry_aggressiveness:.2f})")
|
||||
# Add signal to accumulator for trend confirmation
|
||||
signal_data = {
|
||||
'action': best_action,
|
||||
@ -1418,8 +1434,7 @@ class TradingOrchestrator:
|
||||
except Exception:
|
||||
memory_usage = {}
|
||||
|
||||
# Calculate dynamic aggressiveness based on recent performance
|
||||
entry_aggressiveness = self._calculate_dynamic_entry_aggressiveness(symbol)
|
||||
# Get exit aggressiveness (entry aggressiveness already calculated above)
|
||||
exit_aggressiveness = self._calculate_dynamic_exit_aggressiveness(symbol, current_position_pnl)
|
||||
|
||||
# Create final decision
|
||||
@ -1440,6 +1455,9 @@ class TradingOrchestrator:
|
||||
f"entry_agg: {entry_aggressiveness:.2f}, exit_agg: {exit_aggressiveness:.2f}, "
|
||||
f"pnl: ${current_position_pnl:.2f})")
|
||||
|
||||
# Trigger training on each decision (especially for executed trades)
|
||||
self._trigger_training_on_decision(decision, price)
|
||||
|
||||
return decision
|
||||
|
||||
except Exception as e:
|
||||
@ -2032,6 +2050,253 @@ class TradingOrchestrator:
|
||||
logger.error(f"Error calculating enhanced reward: {e}")
|
||||
return base_pnl
|
||||
|
||||
def _trigger_training_on_decision(self, decision: TradingDecision, current_price: float):
|
||||
"""Trigger training on each decision, especially executed trades
|
||||
|
||||
This ensures models learn from every signal outcome, giving more weight
|
||||
to executed trades as they have real market feedback.
|
||||
"""
|
||||
try:
|
||||
# Only train if training is enabled and we have the enhanced training system
|
||||
if not self.training_enabled or not self.enhanced_training_system:
|
||||
return
|
||||
|
||||
symbol = decision.symbol
|
||||
action = decision.action
|
||||
confidence = decision.confidence
|
||||
|
||||
# Create training data from the decision
|
||||
training_data = {
|
||||
'symbol': symbol,
|
||||
'action': action,
|
||||
'confidence': confidence,
|
||||
'price': current_price,
|
||||
'timestamp': decision.timestamp,
|
||||
'executed': action != 'HOLD', # Assume non-HOLD actions are executed
|
||||
'entry_aggressiveness': decision.entry_aggressiveness,
|
||||
'exit_aggressiveness': decision.exit_aggressiveness,
|
||||
'reasoning': decision.reasoning
|
||||
}
|
||||
|
||||
# Add to enhanced training system for immediate learning
|
||||
if hasattr(self.enhanced_training_system, 'add_decision_for_training'):
|
||||
self.enhanced_training_system.add_decision_for_training(training_data)
|
||||
logger.debug(f"🎓 Added decision to training queue: {action} {symbol} (conf: {confidence:.3f})")
|
||||
|
||||
# Trigger immediate training for executed trades (higher priority)
|
||||
if action != 'HOLD':
|
||||
if hasattr(self.enhanced_training_system, 'trigger_immediate_training'):
|
||||
self.enhanced_training_system.trigger_immediate_training(
|
||||
symbol=symbol,
|
||||
priority='high' if confidence > 0.7 else 'medium'
|
||||
)
|
||||
logger.info(f"🚀 Triggered immediate training for executed trade: {action} {symbol}")
|
||||
|
||||
# Train all models on the decision outcome
|
||||
self._train_models_on_decision(decision, current_price)
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Error triggering training on decision: {e}")
|
||||
|
||||
def _train_models_on_decision(self, decision: TradingDecision, current_price: float):
|
||||
"""Train all models on the decision outcome
|
||||
|
||||
This provides immediate feedback to models about their predictions,
|
||||
allowing them to learn from each signal they generate.
|
||||
"""
|
||||
try:
|
||||
symbol = decision.symbol
|
||||
action = decision.action
|
||||
confidence = decision.confidence
|
||||
|
||||
# Get current market data for training context
|
||||
market_data = self._get_current_market_data(symbol)
|
||||
if not market_data:
|
||||
return
|
||||
|
||||
# Train DQN agent if available
|
||||
if self.rl_agent and hasattr(self.rl_agent, 'add_experience'):
|
||||
try:
|
||||
# Create state representation
|
||||
state = self._create_state_for_training(symbol, market_data)
|
||||
|
||||
# Map action to DQN action space
|
||||
action_mapping = {'BUY': 0, 'SELL': 1, 'HOLD': 2}
|
||||
dqn_action = action_mapping.get(action, 2)
|
||||
|
||||
# Calculate immediate reward based on confidence and execution
|
||||
immediate_reward = confidence if action != 'HOLD' else 0.0
|
||||
|
||||
# Add experience to DQN
|
||||
self.rl_agent.add_experience(
|
||||
state=state,
|
||||
action=dqn_action,
|
||||
reward=immediate_reward,
|
||||
next_state=state, # Will be updated with actual outcome later
|
||||
done=False
|
||||
)
|
||||
|
||||
logger.debug(f"🧠 Added DQN experience: {action} {symbol} (reward: {immediate_reward:.3f})")
|
||||
|
||||
except Exception as e:
|
||||
logger.debug(f"Error training DQN on decision: {e}")
|
||||
|
||||
# Train CNN model if available
|
||||
if self.cnn_model and hasattr(self.cnn_model, 'add_training_sample'):
|
||||
try:
|
||||
# Create CNN input features
|
||||
cnn_features = self._create_cnn_features_for_training(symbol, market_data)
|
||||
|
||||
# Create target based on action
|
||||
target_mapping = {'BUY': [1, 0, 0], 'SELL': [0, 1, 0], 'HOLD': [0, 0, 1]}
|
||||
target = target_mapping.get(action, [0, 0, 1])
|
||||
|
||||
# Add training sample
|
||||
self.cnn_model.add_training_sample(cnn_features, target, weight=confidence)
|
||||
|
||||
logger.debug(f"🔍 Added CNN training sample: {action} {symbol}")
|
||||
|
||||
except Exception as e:
|
||||
logger.debug(f"Error training CNN on decision: {e}")
|
||||
|
||||
# Train COB RL model if available and we have COB data
|
||||
if self.cob_rl_agent and symbol in self.latest_cob_data:
|
||||
try:
|
||||
cob_data = self.latest_cob_data[symbol]
|
||||
if hasattr(self.cob_rl_agent, 'add_experience'):
|
||||
# Create COB state representation
|
||||
cob_state = self._create_cob_state_for_training(symbol, cob_data)
|
||||
|
||||
# Add COB experience
|
||||
self.cob_rl_agent.add_experience(
|
||||
state=cob_state,
|
||||
action=action,
|
||||
reward=confidence,
|
||||
symbol=symbol
|
||||
)
|
||||
|
||||
logger.debug(f"📊 Added COB RL experience: {action} {symbol}")
|
||||
|
||||
except Exception as e:
|
||||
logger.debug(f"Error training COB RL on decision: {e}")
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Error training models on decision: {e}")
|
||||
|
||||
def _get_current_market_data(self, symbol: str) -> Optional[Dict]:
|
||||
"""Get current market data for training context"""
|
||||
try:
|
||||
if self.data_provider:
|
||||
# Get recent data for training
|
||||
df = self.data_provider.get_historical_data(symbol, '1m', limit=100)
|
||||
if df is not None and not df.empty:
|
||||
return {
|
||||
'ohlcv': df.tail(50).to_dict('records'), # Last 50 candles
|
||||
'current_price': float(df['close'].iloc[-1]),
|
||||
'volume': float(df['volume'].iloc[-1]),
|
||||
'timestamp': df.index[-1]
|
||||
}
|
||||
return None
|
||||
except Exception as e:
|
||||
logger.debug(f"Error getting market data for training: {e}")
|
||||
return None
|
||||
|
||||
def _create_state_for_training(self, symbol: str, market_data: Dict) -> np.ndarray:
|
||||
"""Create state representation for DQN training"""
|
||||
try:
|
||||
# Create a basic state representation
|
||||
ohlcv_data = market_data.get('ohlcv', [])
|
||||
if not ohlcv_data:
|
||||
return np.zeros(100) # Default state size
|
||||
|
||||
# Extract features from recent candles
|
||||
features = []
|
||||
for candle in ohlcv_data[-20:]: # Last 20 candles
|
||||
features.extend([
|
||||
candle.get('open', 0),
|
||||
candle.get('high', 0),
|
||||
candle.get('low', 0),
|
||||
candle.get('close', 0),
|
||||
candle.get('volume', 0)
|
||||
])
|
||||
|
||||
# Pad or truncate to expected size
|
||||
state = np.array(features[:100])
|
||||
if len(state) < 100:
|
||||
state = np.pad(state, (0, 100 - len(state)), 'constant')
|
||||
|
||||
return state
|
||||
|
||||
except Exception as e:
|
||||
logger.debug(f"Error creating state for training: {e}")
|
||||
return np.zeros(100)
|
||||
|
||||
def _create_cnn_features_for_training(self, symbol: str, market_data: Dict) -> np.ndarray:
|
||||
"""Create CNN features for training"""
|
||||
try:
|
||||
# Similar to state creation but formatted for CNN
|
||||
ohlcv_data = market_data.get('ohlcv', [])
|
||||
if not ohlcv_data:
|
||||
return np.zeros((1, 100))
|
||||
|
||||
# Create feature matrix
|
||||
features = []
|
||||
for candle in ohlcv_data[-20:]:
|
||||
features.extend([
|
||||
candle.get('open', 0),
|
||||
candle.get('high', 0),
|
||||
candle.get('low', 0),
|
||||
candle.get('close', 0),
|
||||
candle.get('volume', 0)
|
||||
])
|
||||
|
||||
# Reshape for CNN input
|
||||
cnn_features = np.array(features[:100]).reshape(1, -1)
|
||||
if cnn_features.shape[1] < 100:
|
||||
cnn_features = np.pad(cnn_features, ((0, 0), (0, 100 - cnn_features.shape[1])), 'constant')
|
||||
|
||||
return cnn_features
|
||||
|
||||
except Exception as e:
|
||||
logger.debug(f"Error creating CNN features for training: {e}")
|
||||
return np.zeros((1, 100))
|
||||
|
||||
def _create_cob_state_for_training(self, symbol: str, cob_data: Dict) -> np.ndarray:
|
||||
"""Create COB state representation for training"""
|
||||
try:
|
||||
# Extract COB features for training
|
||||
features = []
|
||||
|
||||
# Add bid/ask data
|
||||
bids = cob_data.get('bids', [])[:10] # Top 10 bids
|
||||
asks = cob_data.get('asks', [])[:10] # Top 10 asks
|
||||
|
||||
for bid in bids:
|
||||
features.extend([bid.get('price', 0), bid.get('size', 0)])
|
||||
for ask in asks:
|
||||
features.extend([ask.get('price', 0), ask.get('size', 0)])
|
||||
|
||||
# Add market stats
|
||||
stats = cob_data.get('stats', {})
|
||||
features.extend([
|
||||
stats.get('spread', 0),
|
||||
stats.get('mid_price', 0),
|
||||
stats.get('bid_volume', 0),
|
||||
stats.get('ask_volume', 0),
|
||||
stats.get('imbalance', 0)
|
||||
])
|
||||
|
||||
# Pad to expected COB state size (2000 features)
|
||||
cob_state = np.array(features[:2000])
|
||||
if len(cob_state) < 2000:
|
||||
cob_state = np.pad(cob_state, (0, 2000 - len(cob_state)), 'constant')
|
||||
|
||||
return cob_state
|
||||
|
||||
except Exception as e:
|
||||
logger.debug(f"Error creating COB state for training: {e}")
|
||||
return np.zeros(2000)
|
||||
|
||||
def _check_signal_confirmation(self, symbol: str, signal_data: Dict) -> Optional[str]:
|
||||
"""Check if we have enough signal confirmations for trend confirmation with rate limiting"""
|
||||
try:
|
||||
|
||||
Reference in New Issue
Block a user