gavindiaz/TR_Agent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
b32b055ee3718c0e2250cc29036f5e43e081fd16
TR_Agent/DOCUMENTAZIONE_AGENTI.md
T

9.7 KiB
Raw Blame History

Multi-Agent Trading System - Interaction and Market Logic Documentation

Agent Interaction Flow

Overview

The multi-agent system operates through a structured interaction flow where each agent contributes to the final trading decision. This flow is divided into three main phases:

  1. Analysis Phase: Each agent analyzes market data and produces individual signals
  2. Interaction Phase: Agents communicate and combine their signals
  3. Orchestration Phase: The orchestrator makes the final trading decision

Detailed Interaction Flow

Phase 1: Analysis

Each agent independently analyzes the market data during the Analyze() method call:

  1. Hurst Agent (RegimeDetector):

    • Calculates Hurst exponent using DFA/R/S analysis
    • Updates SHARED_regimeH and SHARED_regimeZ
    • Period dynamically adjusted based on data length

  2. ADX Agent (RegimeADX):

    • Calculates ADX using native MT5 function
    • Normalizes ADX to z-score
    • Updates SHARED_adxZ and SHARED_adxRaw

  3. MA Agent (MAAgent):

    • Calculates price-to-MA ratio
    • Computes slope of MA
    • Combines signals using variance weighting
    • Updates lastZScore (local, not shared)

  4. Momentum Agent (MomentumAgent):

    • Calculates momentum and acceleration
    • Computes z-scores for both
    • Combines signals using variance weighting
    • Updates lastZScore (local, not shared)

  5. Consensus Agent (RegimeConsensus):

    • Called by orchestrator during Interact() phase
    • Combines Hurst and ADX signals
    • Calculates agreement between agents
    • Updates SHARED_regimeConsensus, SHARED_regimeAgreement
    • Detects trading patterns
    • Updates SHARED_patternCode, SHARED_patternName

  6. Pattern Hunter Agent (PatternHunter):

    • Called by orchestrator during Interact() phase
    • Analyzes historical patterns
    • Updates lastZScore (local, not shared)

Phase 2: Orchestration

The orchestrator combines agent signals using a sophisticated softmax gating mechanism:

  1. Signal Collection:

    • Collects z-scores from all agents
    • Filters out disabled agents

  2. Softmax Gating:

    • Calculates mean and variance of agent z-scores
    • Temperature parameter: temp = √(zVar) / √(agentCount)
    • Higher variance = higher temperature = more mixing
    • Lower variance = lower temperature = more selective

  3. Combined Signal:

    • Applies softmax to get weighted agent contributions
    • Applies tanh to normalize to [-1, 1]
    • Updates combinedZ and combinedZStats

Phase 3: Market Decision

  1. Entry Signal:

    • Combined z-score compared to minActionableZ
    • minActionableZ = combinedZStats.Std() (empirical)
    • Fallback: 0.5 (when stats not ready)

  2. Position Sizing:

    • Base lot: AdaptiveBaseLot()
    • Adjusted by entry z-score magnitude
    • Capped by broker limits

  3. Risk Management:

    • SL: AdaptiveSL() = returnStats.Std()
    • TP: AdaptiveTP() = AdaptiveSL() * RR ratio
    • RR ratio: 1.0 / (1.0 - winRate)

Market Entry/Exit Logic

Entry Conditions

The system enters a trade when:

  1. Signal Strength:

    • |combinedZ| > minActionableZ
    • minActionableZ = combinedZStats.Ready() ? combinedZStats.Std() : 0.5
    • This ensures we only trade when the signal is statistically significant

  2. Position Direction:

    • dir = 1 if combinedZ > thr
    • dir = -1 if combinedZ < -thr
    • thr = (overrideMinZ > 0) ? overrideMinZ : AdaptiveMinZ()

  3. Lot Size Calculation:

    riskFrac = wr * (1.0 - MathMin(maxDrawdown, 0.5)) * (0.5 + sharpeBonus * 0.5)
riskFrac = MathMax(0.002, MathMin(0.02, riskFrac))
lot = balance * riskFrac / 100000.0
lot = lot * MathAbs(fs.zScore)  // Scale by signal strength

Exit Conditions

Take Profit (TP)

  • Trigger: Price reaches TP level
  • TP Level: entryPrice + (atr * tpMult) (long) or entryPrice - (atr * tpMult) (short)
  • tpMult: AdaptiveTP() / atr

Stop Loss (SL)

  • Trigger: Price reaches SL level
  • SL Level: entryPrice - (atr * slMult) (long) or entryPrice + (atr * slMult) (short)
  • slMult: AdaptiveSL() / atr

Opposite Signal

  • Trigger: New trade signal in opposite direction
  • Logic: Close existing position before opening new one
  • Implementation: Check CountPositions() before opening new trade

Time-based Exit

  • Maximum Trades: 20 trades per session
  • Rollback Convergence: If weights stabilize for minTrades, system marks as converged

Risk Management Details

Drawdown Tracking

double eq = AccountInfoDouble(ACCOUNT_EQUITY);
if(eq > peakEquity) peakEquity = eq;
double dd = (peakEquity > 0) ? (peakEquity - eq) / peakEquity : 0;
if(dd > maxDrawdown) maxDrawdown = dd;

Weight Adaptation

if(combinedZStats.Ready() && returnStats.Count() > 0) {
   double rv = returnStats.Std();
   weightAlpha = rv / (1.0 + rv);
   weightAlpha = MathMax(minAlpha, weightAlpha);
   agents[i].weight = MathMax(weightMin, rho);
}

Sharpe Bonus Calculation

double sharpeBonus = MathMax(0, MathMin(1, rollingSharpe * 0.5));

Market Hours and Session Management

Trading Hours

  • Symbol-specific: Respects each symbol's trading hours
  • No weekend trading: No trades on weekends
  • Session limits: Maximum 20 trades per trading session

State Management

  • Model persistence: Saves state at intervals
  • Warm-up: Uses historical data for initial statistics
  • Recovery: Loads saved state on startup

Detailed Agent Interaction Examples

Example 1: Hurst and ADX Interaction

  1. Hurst Agent Output:

    • SHARED_regimeH = 0.6 (trending market)
    • SHARED_regimeZ = 0.8 (strong bullish signal)

  2. ADX Agent Output:

    • SHARED_adxZ = 0.7 (strong trend)
    • SHARED_adxRaw = 35.0 (strong trend strength)

  3. Consensus Agent Processing:

    • hurstZ = 0.8, adxZ = 0.7
    • hStd = 0.5, aStd = 0.5
    • combinedScale = √(0.5² + 0.5²) = 0.707
    • agreement = 1.0 - (0.1 / 0.707) = 0.86
    • SHARED_regimeConsensus = tanh((0.8 + 0.7) / 2 * 0.86) = 0.94
    • Pattern: strong_trend (pCode = 1)

Example 2: Pattern Hunter Interaction

  1. Pattern Hunter Inputs:

    • hurst = 0.8, adx = 0.7, ma = 0.3, mom = 0.9
    • tH = 0.5, tA = 0.5, tM = 0.4, tM2 = 0.6

  2. Pattern Detection:

    • maMomAgree = (0.3 * 0.9 > 0.4 * 0.6) = true
    • regimeTrend = (0.8 + 0.7) > 0.707 = true
    • momAccel = (0.9 > 0.3 + 0.4 * 0.15) && (0.3 > 0.3 + 0.4 * 0.15) = true
    • Pattern: trend_accel (pCode = 8)
    • pZ = (0.3 + 0.9) / comboNorm = 1.2 / 0.707 = 1.70
    • patternStrength = 1.0 - exp(-1.70) = 0.82

Example : Market Entry

  1. Combined Signal:

    • combinedZ = 0.94 (from consensus)
    • minActionableZ = 0.5 (from combinedZStats)
    • Entry: |0.94| > 0.5 ✓

  2. Position Sizing:

    • balance = 10000
    • winRate = 0.6, maxDrawdown = 0.2, sharpeBonus = 0.8
    • riskFrac = 0.6 * (1 - 0.2) * (0.5 + 0.8 * 0.5) = 0.6 * 0.8 * 0.9 = 0.432
    • riskFrac = MathMax(0.002, MathMin(0.02, 0.432)) = 0.02
    • lot = 10000 * 0.02 / 100000 = 0.2
    • lot = 0.2 * MathAbs(0.94) = 0.188

  3. SL/TP:

    • ATR = 0.02
    • SL = entryPrice - 0.02 * 1.5 = entryPrice - 0.03
    • TP = entryPrice + 0.02 * 2.5 = entryPrice + 0.05

Error Handling and Recovery

Data Issues

  1. Insufficient Data:

    • Wait for minimum samples before using statistics
    • Use fallback values when data insufficient

  2. Division by Zero:

    • Use DATA_EPS() for safe division
    • Check for zero variance before calculations

  3. Invalid Values:

    • Validate all input data
      • Use reasonable defaults for invalid values

Recovery Mechanisms

  1. Model Reload:

    • Automatic reload on startup if save file exists
    • Graceful degradation if save file corrupted

  2. State Reset:

    • Reset all statistics on error
    • Log errors for debugging

Performance Monitoring

Key Metrics

  1. Trading Performance:

    • Win rate
    • Average win/loss ratio
    • Maximum drawdown
    • Sharpe ratio

  2. System Performance:

    • Agent weight convergence
    • Signal strength statistics
    • Processing time
    • Memory usage

Health Monitoring

  1. Orchestrator Health:

    • PrintAgentStatus(): Shows all agent signals and weights
    • LogHealth(): Shows rolling performance metrics
    • LogBarHistory(): Stores historical signals

  2. Agent Health:

    • Each agent has SignalInfo() method
    • Shows agent name, z-score, and statistics
    • Helps identify underperforming agents

Future Enhancements

  1. Advanced Features:

    • Multi-timeframe analysis
    • Machine learning for pattern recognition
    • Natural language processing for news sentiment

  2. Risk Management:

    • Dynamic position sizing based on volatility
    • Correlation-based risk limits
    • Market regime-specific risk adjustments

  3. Monitoring:

    • Real-time performance dashboards
    • Automated alert system
    • Performance benchmarking

Conclusion

This documentation provides a comprehensive understanding of the multi-agent trading system's architecture, interaction patterns, and market logic. The system combines multiple analytical approaches to create a robust, data-driven trading strategy that adapts to changing market conditions.

Key strengths:

  • Diversification: Multiple analytical approaches reduce reliance on single indicators
  • Adaptability: All parameters are data-driven and adaptive
  • Robustness: Consensus-based decision making with lateral communication
  • Transparency: Clear documentation of all agent interactions and logic

The system is designed to be both powerful and maintainable, with clear separation of concerns between different analytical components.

Reference in New Issue View Git Blame Copy Permalink