Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reinforcement Learning for Option Hedging: Static Implied-Volatility Fit versus Shortfall-Aware Performance

Ziheng Chen, Minxuan Hu, Jiayu Yi, Wenxi Sun

TL;DR

The paper addresses the disconnect between static IV-fitting and dynamic hedging performance under market frictions. It extends the QLBS framework by incorporating risk aversion and transaction costs and introduces Replication Learning of Option Pricing (RLOP), a forward-looking replication approach. Using SPY and XOP data across crisis and calm regimes, the authors show Adaptive-QLBS improves static implied-volatility fit, while RLOP achieves superior dynamic hedging by reducing shortfall probability and trading costs. By decoupling pricing calibration from hedging performance and leveraging a bidirectional RL architecture, the work highlights practical paths to robust hedging in frictional markets.

Abstract

We extend the Q-learner in Black-Scholes (QLBS) framework by incorporating risk aversion and trading costs, and propose a novel Replication Learning of Option Pricing (RLOP) approach. Both methods are fully compatible with standard reinforcement learning algorithms and operate under market frictions. Using SPY and XOP option data, we evaluate performance along static and dynamic dimensions. Adaptive-QLBS achieves higher static pricing accuracy in implied volatility space, while RLOP delivers superior dynamic hedging performance by reducing shortfall probability. These results highlight the importance of evaluating option pricing models beyond static fit, emphasizing realized hedging outcomes.

Reinforcement Learning for Option Hedging: Static Implied-Volatility Fit versus Shortfall-Aware Performance

TL;DR

The paper addresses the disconnect between static IV-fitting and dynamic hedging performance under market frictions. It extends the QLBS framework by incorporating risk aversion and transaction costs and introduces Replication Learning of Option Pricing (RLOP), a forward-looking replication approach. Using SPY and XOP data across crisis and calm regimes, the authors show Adaptive-QLBS improves static implied-volatility fit, while RLOP achieves superior dynamic hedging by reducing shortfall probability and trading costs. By decoupling pricing calibration from hedging performance and leveraging a bidirectional RL architecture, the work highlights practical paths to robust hedging in frictional markets.

Abstract

We extend the Q-learner in Black-Scholes (QLBS) framework by incorporating risk aversion and trading costs, and propose a novel Replication Learning of Option Pricing (RLOP) approach. Both methods are fully compatible with standard reinforcement learning algorithms and operate under market frictions. Using SPY and XOP option data, we evaluate performance along static and dynamic dimensions. Adaptive-QLBS achieves higher static pricing accuracy in implied volatility space, while RLOP delivers superior dynamic hedging performance by reducing shortfall probability. These results highlight the importance of evaluating option pricing models beyond static fit, emphasizing realized hedging outcomes.
Paper Structure (20 sections, 2 theorems, 4 equations, 4 figures, 10 tables)

This paper contains 20 sections, 2 theorems, 4 equations, 4 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Replication Pricing and Reinforcement Learning Framework
  3. Two Reinforcement Learning Paradigms for Option Pricing
  4. Adaptive QLBS: A Backward Value-Based RL Framework
  5. RLOP: A Forward Replication Learning Framework
  6. Learning Algorithms and Numerical Verification
  7. Empirical Evaluation on Market Data
  8. Data Description and Experimental Design
  9. Data
  10. Bucketing and moneyness
  11. Daily calibration and fitting
  12. Evaluation
  13. Static Pricing Accuracy: Implied Volatility Fit
  14. Dynamic Hedging Performance
  15. ATM benchmark (Table \ref{['tab:hedging_28d_atm']})
  16. ...and 5 more sections

Key Result

Proposition 1

For sufficiently large $\epsilon$ that appears in the linear transaction cost assumption $\text{TC}(\Delta u, S)=\epsilon|\Delta u| \,S$, the option price $C(S_0) := -\max_{\pi\in\mathbf{\Pi}} V_0^\pi$ is monotonically increasing in both $\lambda$ and $\epsilon$.

Figures (4)

  • Figure 1: The adaptive-QLBS method takes a backward, value-based approach.
  • Figure 2: The RLOP method takes a forward, replication-based approach.
  • Figure 3: Price under Adaptive-QLBS model (left) and RLOP model (right) given different parameters of volatility. The common setup uses maturity $T=2$ months, strike $K=1$, interest rate $r=4\%$.
  • Figure 4: Price under Adaptive-QLBS model given different levels of hyperparameters: drift $\mu$ (left), risk aversion intensity $\lambda$ (middle), and friction $\epsilon$ (right).

Theorems & Definitions (5)

  • Definition 1
  • Proposition 1
  • Definition 2
  • Proposition 1
  • proof