Deep Hedging Symbolic Distillation

---
name: deep-hedging-symbolic-distillation
description: "Methodology for auditing and distilling deep reinforcement learning hedging policies into interpretable symbolic formulas. Includes framework for analyzing delta corrections relative to Black-Scholes, symbolic regression distillation, and regime fragility stress-testing. Use when analyzing neural hedging strategies, quantitative risk management, options hedging with RL, or making black-box financial AI auditable."
---

# Deep Hedging Symbolic Distillation

## Description

Framework for making deep reinforcement learning (RL) hedging policies transparent and auditable. Combines empirical analysis of neural delta corrections with symbolic regression to distill complex policies into compact formulas, while systematically identifying regime conditions where learned hedges fail.

## Activation Keywords

- deep hedging
- symbolic distillation
- RL options hedging
- neural policy audit
- delta haircut analysis
- regime fragility testing
- quantitative finance RL
- hedge formula extraction

## Core Methodology

### Phase 1: Deep Hedging Baseline

1. **Train RL Agent**: Use TD3 or similar actor-critic algorithm for options hedging.
   - **State space**: Spot price, implied volatility, time-to-maturity, inventory.
   - **Action space**: Hedge ratio adjustments.
   - **Reward**: Local downside-shortfall or P&L-based utility.

2. **Walk-Forward Evaluation**: Train on rolling windows (e.g., 2015-2023) to prevent look-ahead bias.

3. **Baseline Comparison**: Compare against daily-updated Black-Scholes delta hedging on identical episodes.

### Phase 2: Delta Correction Analysis

1. **Compute Delta Haircuts**: Calculate `Agent Delta - BS Delta` for each time step.
   - Agents often learn systematic "haircuts" (adjustments) to BS deltas.
   - These corrections are typically driven by spot-implied-volatility co-movement.

2. **Feature Attribution**: Analyze when corrections occur.
   - Plot delta haircut vs. spot returns and vol changes.
   - Identify if the agent is hedging volatility risk that BS ignores.

3. **Performance Decomposition**:
   - Measure improvement in accumulated reward, terminal downside variance, and CVaR.
   - Attribute gains to specific delta adjustments.

### Phase 3: Symbolic Distillation

1. **Data Collection**: Generate state-action pairs from the trained agent.
   - Input features: Spot, IV, Tau, Inventory.
   - Target: Agent's hedging action.

2. **Symbolic Regression**: Use tools like `PySR` or `gplearn` to discover compact formulas.
   - **Objective**: Maximize reward correlation while minimizing formula complexity.
   - **Constraints**: Limit expression depth, prefer operations interpretable by traders.

3. **Validation**:
   - Trade the distilled formula out-of-sample.
   - Compare reward, downside variance, and CVaR against both RL agent and BS baseline.
   - **Success criteria**: Formula preserves >80% of RL advantage over BS.

### Phase 4: Regime Fragility Analysis

1. **Regime Identification**: Segment test period by market conditions.
   - Low vol vs. high vol regimes.
   - Spot-dominated P&L vs. vol-dominated P&L.
   - Adverse daily states (e.g., sharp drawdowns).

2. **Stress Testing**:
   - Evaluate distilled formula in each regime.
   - **Key finding**: Underhedging can raise variance when option P&L is spot-dominated and vol channel is weak.
   - Identify "fragility zones" where the hedge underperforms BS.

3. **Robustness Recommendations**:
   - If formula is regime-fragile, add regime-switching logic or fallback to BS in identified zones.
   - Monitor volatility channel strength in real-time to detect fragility conditions.

## Tools & Dependencies

- **RL Framework**: PyTorch, Stable Baselines3, or custom TD3 implementation.
- **Symbolic Regression**: `PySR` (recommended for speed/quality), `gplearn`.
- **Financial Data**: Options chains, implied volatility surfaces, spot prices.
- **Analysis**: Pandas, NumPy, Matplotlib for regime visualization.

## Error Handling & Pitfalls

### Regime Fragility
- **Symptom**: Distilled formula works in backtest but fails in live trading during stress periods.
- **Fix**: Implement regime detection; fallback to BS delta when volatility channel is weak or market is in adverse state.

### Overfitting Symbolic Formula
- **Symptom**: Formula fits training data perfectly but degrades OOS.
- **Fix**: Use cross-validation across multiple time windows; penalize complexity in symbolic regression objective.

### Computational Cost
- **Symptom**: Training RL agents for hedging is slow.
- **Fix**: Use parallelized simulation for option episodes; pre-compute BS deltas as features to reduce learning burden.

### Auditability Trade-off
- **Symptom**: Most accurate symbolic formulas are too complex to be interpretable.
- **Fix**: Constrain expression depth; prioritize formulas that map to intuitive risk factors (e.g., vega adjustments).

## Examples

### Example 1: Auditing a Neural Hedge
```
User: "Analyze what our RL hedging agent actually learned compared to Black-Scholes."

Agent Process:
1. Load agent weights and replay buffer.
2. Compute delta haircuts over 2015-2023 walk-forward window.
3. Plot haircut vs. spot-IV co-movement.
4. Report: "Agent learns systematic -5% delta haircut when IV rises with spot, improving CVaR by 12%."
```

### Example 2: Distilling to Formula
```
User: "Convert this neural hedging policy into a formula we can audit."

Agent Process:
1. Extract state-action pairs from agent.
2. Run PySR with complexity penalty.
3. Return formula: `Delta_BS * (1 - 0.05 * sign(dIV/dt) * |dS|)`.
4. Validate: "Formula preserves 88% of RL reward advantage, auditable by risk team."
```

## Resources

- **Source Paper**: [arXiv:2605.21696](https://arxiv.org/abs/2605.21696) - "What Does Deep Hedging Actually Learn? Delta Corrections, Regime Fragility, and Symbolic Distillation"
- **PySR Documentation**: https://pysr.readthedocs.io/
- **Black-Scholes Reference**: Standard options pricing models for baseline comparison.