Neuro-Symbolic Trading: Combining Neural Networks with Symbolic Rules

The Problem with Pure Neural Trading

Pure neural network trading systems have a fundamental problem: they're black boxes. You can't inspect why they made a decision, you can't enforce risk rules, and they can fail catastrophically when market conditions change.

The Solution: Neuro-Symbolic Architecture

Neuro-symbolic trading combines:

• Neural networks for pattern recognition and prediction
• Symbolic rules for risk management and decision gating

The key insight: symbolic rules are checked BEFORE the neural network. The neural network cannot override them.

Architecture

Market Data
    |
    v
[Symbolic Rules] -- checked first
    |  |
    |  v
    |  [Block] if rules violated
    |
    v
[Neural Network] -- prediction
    |
    v
[Decision] -- final

Symbolic Rules (Python)

class SymbolicRules:
    def check(self, signal, position, account):
        # Rule 1: Never exceed max position size
        if position.size + signal.size > account.max_position:
            return False, "Position too large"

        # Rule 2: Never trade without stop loss
        if signal.type == "open" and signal.stop_loss is None:
            return False, "Stop loss required"

        # Rule 3: Never trade in unknown regime
        if market.regime == "unknown":
            return False, "Unknown market regime"

        # Rule 4: Max drawdown check
        if account.drawdown > account.max_drawdown:
            return False, "Max drawdown exceeded"

        return True, "OK"

Neural Network (Prediction)

class NeuralPredictor:
    def predict(self, features):
        # 60% technical indicators
        technical = self.technical_model(features.rsi, features.macd, features.bollinger)

        # 40% MoP-JEPA (Multi-output Prediction with Joint-Embedding Predictive Architecture)
        mopa = self.mopa_model(features.candles, features.orderbook)

        # Weighted combination
        signal = 0.6 * technical + 0.4 * mopa
        return signal

Combined Decision

def make_decision(market_data, position, account):
    # Step 1: Neural prediction
    signal = neural_predictor.predict(market_data)

    # Step 2: Symbolic rule check (CANNOT be overridden)
    approved, reason = symbolic_rules.check(signal, position, account)
    if not approved:
        log(f"Trade blocked: {reason}")
        return HOLD

    # Step 3: Execute
    return signal.action

Benefits

1. Safety: Symbolic rules prevent catastrophic losses
2. Explainability: You can inspect why a trade was blocked
3. Adaptability: Neural network adapts to market changes
4. Reliability: Rules ensure consistent risk management

Results

• Win rate: 55.8% (BTC+ETH+SOL backtest)
• Max drawdown: controlled by symbolic rules
• No catastrophic losses (rules prevent them)

Next Steps

• Add more symbolic rules (regime detection, correlation limits)
• Improve neural predictor with more features
• Add continual learning for regime adaptation

---

This is a project from Nexus Trading - neuro-symbolic trading research.