Bridging Stochastic Control and Deep Hedging: Structural Priors for No-Transaction Band Networks

1. Problem Statement

The paper addresses the problem of hedging and pricing European options under proportional transaction costs.

• Context: The standard Black-Scholes model assumes continuous rebalancing, which is impossible in practice due to transaction costs (bid-ask spreads, fees).
• Core Challenge: Under transaction costs, continuous delta-hedging is suboptimal. The optimal strategy involves a No-Transaction Band (NTB): a region around the optimal holding where it is rational to do nothing. Trading only occurs when the portfolio drifts outside this band.
• Gap: While stochastic control theory provides rigorous solutions (via Hamilton-Jacobi-Bellman Quasi-Variational Inequalities, HJBQVI), they suffer from the "curse of dimensionality" and are computationally expensive. Conversely, Deep Hedging (Deep Learning) approaches are flexible but often lack structural guidance, leading to slower convergence and less interpretable policies.

2. Methodology

The authors approach the problem from two complementary angles: Stochastic Control (analytical/numerical) and Deep Hedging (neural networks), aiming to bridge them.

A. Stochastic Control Framework

• Model: Geometric Brownian Motion (GBM) for the underlying asset with proportional transaction costs ($\lambda$).
• Utility: Constant Absolute Risk Aversion (CARA) utility, $u(x) = 1 - e^{-\gamma x}$.
• Dimension Reduction: A key theoretical step involves proving that under CARA utility, the value function can be factorized to eliminate the cash dimension. This reduces the problem from a 4D PDE to a 3D Quasi-Variational Inequality (QVI) dependent on time ($t$), stock price ($S$), and share holding ($y$).
• Solution: The authors solve the reduced QVI numerically using a backward induction scheme on a binomial tree for $S$ and a uniform grid for $y$.
• Asymptotic Approximation: They utilize the Whalley-Wilmott (WW) asymptotic formula for the NTB half-width in the limit of small $\lambda$:
  $$h_{WW} = \left( \frac{3\lambda \delta(t,T) S \Gamma_{BS}^2}{2\gamma} \right)^{1/3}$$
  where $\Gamma_{BS}$ is the Black-Scholes gamma.

B. Deep Hedging Framework

The authors propose three neural network architectures to learn the hedging policy, minimizing the entropic risk measure (equivalent to CARA utility maximization):

1. MLP Baseline: A standard Multi-Layer Perceptron that takes market features and the previous position as input to output the new hedge directly.
2. NTBN-$\Delta$ (No-Transaction Band Network): An architecture that explicitly outputs the lower and upper bounds of the NTB. Crucially, the authors formalize delta-centring: the band is centered around the Black-Scholes delta ($\Delta_{BS}$), with the network learning only the half-width corrections.
3. WW-NTBN (Whalley-Wilmott NTBN): The proposed novel architecture. It incorporates two structural priors:
   • Analytical Initialization: The initial half-width is set to the Whalley-Wilmott formula ($h_{WW}$). The network learns only a residual correction on top of this analytical prior.
   • Differentiable Soft Clamp: Replaces the non-differentiable "hard clamp" (which causes vanishing gradients when the position is outside the band) with a smooth tanh-based soft clamp, ensuring gradient flow during training.

3. Key Contributions

1. Unified Theoretical Exposition: The paper provides a self-contained derivation connecting the stochastic control HJBQVI to the deep hedging problem, specifically highlighting how the frictionless optimal holding ($\Delta_{BS}$) serves as the natural center for the NTB.
2. Architectural Innovation (WW-NTBN):
   • Embeds the Whalley-Wilmott asymptotic formula as a structural prior, allowing the network to start with the correct scaling regarding transaction costs ($\lambda$) and risk aversion ($\gamma$).
   • Introduces a differentiable soft clamp to solve the vanishing gradient problem inherent in hard-clamped band networks.
3. Bull Call Spread Analysis: The paper demonstrates that under transaction costs, the linearity of option pricing breaks down.
   • Hedging a portfolio (e.g., a bull call spread) jointly is strictly more efficient than hedging legs separately.
   • Joint hedging results in a wider effective NTB (due to gamma cancellation) and lower indifference prices compared to the "naive" sum of individual prices.
4. Comprehensive Numerical Validation: Extensive experiments comparing Stochastic Control (SC), MLP, NTBN-$\Delta$, and WW-NTBN across various transaction cost regimes.

4. Key Results

• Convergence & Performance:
  • WW-NTBN converges significantly faster than both the MLP and NTBN-$\Delta$ baselines.
  • It achieves the lowest entropic risk (terminal P&L variance) and the tightest bid-ask spreads among deep learning models.
  • The learned no-transaction bands in WW-NTBN quantitatively match the Stochastic Control solution more closely than other architectures, even at higher transaction costs where the asymptotic formula is less exact (the network learns the necessary corrections).
• Pricing:
  • All models correctly recover the Black-Scholes price when $\lambda = 0$.
  • As $\lambda$ increases, prices rise. WW-NTBN produces writer prices slightly lower than SC (due to SC's discretization conservatism) but higher than the MLP, striking a balance between optimality and flexibility.
• Bull Call Spread:
  • Joint vs. Naive: Joint hedging yields significantly lower indifference prices than naive (separate) hedging.
  • Mechanism: Naive hedging creates an "empty" intersection of individual bands near the money, forcing near-continuous trading. Joint hedging maintains a wide, active band, drastically reducing trade frequency and costs.
• Trade Metrics: Band-based architectures (NTBN-$\Delta$, WW-NTBN) trade significantly less frequently than the MLP, which struggles to learn the "do nothing" region without structural guidance.

5. Significance and Implications

• Bridging Theory and Practice: The paper successfully demonstrates that deep learning does not need to be a "black box." By injecting rigorous stochastic control insights (like the Whalley-Wilmott formula and delta-centring) as structural priors, neural networks can achieve superior performance, faster convergence, and better generalization.
• Practical Hedging: The results suggest that for practitioners, simply training a generic MLP is suboptimal. Incorporating domain knowledge (like the specific scaling of transaction costs) into the network architecture is critical for robust hedging in high-friction environments.
• Portfolio Management: The findings on the Bull Call Spread highlight a critical risk management insight: treating derivatives as a portfolio rather than a collection of individual instruments is essential to minimize transaction costs and improve pricing accuracy.
• Future Directions: The authors suggest extending WW-NTBN to non-GBM dynamics (e.g., stochastic volatility) by replacing the Black-Scholes gamma with model-specific sensitivities, and using the WW prior as a warm start for Reinforcement Learning.

In summary, this paper establishes that Deep Hedging is most effective when it respects the mathematical structure of the optimal control problem, specifically by using analytical approximations to guide the neural network's learning process.