01 Council Mixture of Experts

01. The 5-Agent Council Architecture (Mixture-of-Experts)

The fundamental failure point of retail algorithmic trading is single-model non-stationarity. Financial time series exhibit evolving statistical distributions: an LSTM or XGBoost model trained on trending bull markets suffers catastrophic drawdown the moment the market shifts into high-volatility sideways mean-reversion.

To overcome single-model fragility, FinRL-X implements a 5-Expert Mixture-of-Experts (MoE) architecture. Every 5-minute candle close, five decoupled specialists evaluate market state independently before casting weighted votes into an NSGA-III Pareto consensus gate.

---

🧠 The 5 Council Specialists

┌─────────────────────────────────────────────────────────────────────────────┐
│                       THE 5 SPECIALIST AGENTS                               │
├───────────────────┬───────────────────────────────────┬─────────────────────┤
│ Specialist        │ Underlying Engine                 │ Primary Mandate     │
├───────────────────┼───────────────────────────────────┼─────────────────────┤
│ E1: DRL Trader    │ Soft Actor-Critic (SAC)           │ Direction Conviction│
│ E2: Regime Master │ 3-State Gaussian HMM              │ Market Regime Gate  │
│ E3: Prophet       │ TimesFM 2.5 Zero-Shot Transformer │ Volatility Corridor │
│ E4: Analyst       │ XGBoost + SHAP Tree Explainability│ Order-Flow Delta    │
│ E5: Chief Actuary │ PyMC Bayesian Credible Intervals  │ Risk & Stop Pricing │
└───────────────────┴───────────────────────────────────┴─────────────────────┘

---

Expert 1: DRL Trader (Soft Actor-Critic)

• Engine: PyTorch continuous-action Soft Actor-Critic (SAC) with twin Q-networks and automated entropy temperature adjustment ($\alpha$).
• State Space: 48 normalized features including M5 return autocorrelation, normalized ATR, Bollinger Band bandwidth, RSI momentum, and tick volume velocity.
• Action Space: Continuous directional conviction $a_t \in [-1.0, +1.0]$:
  • $a_t > +0.30 \implies \text{Long Conviction}$
  • $a_t < -0.30 \implies \text{Short Conviction}$
  • $-0.30 \le a_t \le +0.30 \implies \text{Neutral Hold}$
• Reward Formulation: $$R_t = \frac{r_t}{\sigma_t} - \lambda_{\text{DD}} \cdot \mathbb{I}{\text{DD}} - \lambda{\text{cost}} \cdot (\text{spread} + \text{slippage})$$ The agent is explicitly penalized for drawdown duration and trade friction, discouraging over-trading in noisy sessions.

---

Expert 2: Regime Master (Gaussian Hidden Markov Model)

• Engine: 3-state Gaussian HMM trained on rolling log returns and normalized high-low spread variance.
• States & Position Sizing Multipliers:
  1. State 0 (Bullish Momentum): Low variance, positive mean drift. Multiplier: 1.00x (full allocation).
  2. State 1 (High-Volatility Bear/Chop): High variance, negative skew. Multiplier: 0.25x (defensive dampening).
  3. State 2 (Sideways Mean-Reversion): Low volatility, zero drift. Multiplier: 0.60x (tight targets).
• Veto Authority: If the posterior state probability $P(\text{State } 1) > 0.65$, all breakout expansion trades are vetoed to protect against bull-traps.

---

Expert 3: Prophet (TimesFM 2.5 Zero-Shot Transformer)

• Engine: Google Research TimesFM 2.5 foundation time-series transformer.
• Function: Generates multi-horizon forecasts for the next 12 M5 intervals (1 hour) without task-specific fine-tuning.
• Volatility Band Calibration: Projects 90th and 10th percentile expected price envelopes: $$\text{Corridor Upper} = \hat{y}{t+k}^{(90)}, \quad \text{Corridor Lower} = \hat{y}{t+k}^{(10)}$$
• Prevents entries when price is already extended into the 90th percentile exhaustion zone.

---

Expert 4: Quantitative Analyst (XGBoost + SHAP)

• Engine: Extreme Gradient Boosting (XGBoost) classifier with real-time TreeSHAP contribution values.
• Input Vectors: Bid/ask volume delta, CVD (Cumulative Volume Delta) divergence, Tick Imbalance Ratios, and VWAP displacement.
• Explainability Filter: Trades require positive SHAP attribution from order-flow momentum. If technical indicators indicate long but SHAP reveals institutional volume distribution, the trade is suppressed.

---

Expert 5: Chief Risk Actuary (PyMC Bayesian Inference)

• Engine: Markov Chain Monte Carlo (MCMC) sampling via PyMC.
• Function: Calibrates dynamic Stop Loss and Take Profit levels by computing the 95% Bayesian Highest Density Interval (HDI) of adverse excursion.
• Mandatory Enforcement:
  1. Applies an upfront 60.0-point index breathing floor on NAS100/US30.
  2. Computes the maximum volume allowed under the strict 0.50% capital risk ceiling.
  3. Clamps volume with math.floor and blocks trades where minimum broker lot exceeds the risk budget.

---

⚖️ The Consensus Deliberation Protocol

At each M5 candle close, the Council engine executes the following voting algorithm in src/council/council.py:

# Council Consensus Score Calculation
consensus_score = (
    w_sac * e1_action +
    w_regime * e2_direction * e2_confidence +
    w_prophet * e3_direction +
    w_analyst * e4_signal * e4_prob
)

# High Conviction Gate
if abs(consensus_score) >= 0.70:
    candidate_direction = "BUY" if consensus_score > 0 else "SELL"
    
    # Rule 7: Macro Trend Governor Check
    if candidate_direction == "BUY" and current_bid < h1_ema50:
        logger.warning("VETO: Long trade below H1 EMA 50 (Rule 7)")
        return None
    if candidate_direction == "SELL" and current_ask > h1_ema50:
        logger.warning("VETO: Short trade above H1 EMA 50 (Rule 7)")
        return None
        
    return approve_trade(candidate_direction, consensus_score)

---

🛡️ Rule 7: The H1 Macro Trend Governor

Even if short-term M5 momentum displays a strong signal, counter-trend trades against the higher timeframe structural trend have a failure rate exceeding 68%.

The Council enforces Rule 7:

• BUY Trades: Permitted only if current Bid price is strictly above the 50-period Exponential Moving Average on the H1 timeframe (Bid > H1_EMA50).
• SELL Trades: Permitted only if current Ask price is strictly below the 50-period Exponential Moving Average on the H1 timeframe (Ask < H1_EMA50).

This single rule eliminates whipsaw fakeouts during macro trend days.