Paper: arXiv 2512.12420

Authors: Travon Lucius, Christian Koch, Jacob Starling, Julia Zhu, Miguel Urena, Carrie Hu

Abstract

We present a reinforcement-learning (RL) framework for dynamic hedging of equity index option exposures under realistic transaction costs and position limits. We hedge a normalized option-implied equity exposure (one unit of underlying delta, offset via SPY) by trading the underlying index ETF, using the option surface and macro variables only as state information and not as a direct pricing engine. Building on the “deep hedging” paradigm of Buehler et al. (2019), we design a leak-free environment, a cost-aware reward function, and a lightweight stochastic actor-critic agent trained on daily end-of-day panel data constructed from SPX/SPY implied volatility term structure, skew, realized volatility, and macro rate context. On a fixed train/validation/test split, the learned policy improves risk-adjusted performance versus no-hedge, momentum, and volatility-targeting baselines (higher point-estimate Sharpe); only the GAE policy’s test-sample Sharpe is statistically distinguishable from zero, although confidence intervals overlap with a long-SPY benchmark so we stop short of claiming formal dominance. Turnover remains controlled and the policy is robust to doubled transaction costs. The modular codebase, comprising a data pipeline, simulator, and training scripts, is engineered for extensibility to multi-asset overlays, alternative objectives (e.g., drawdown or CVaR), and intraday data. From a portfolio management perspective, the learned overlay is designed to sit on top of an existing SPX or SPY allocation, improving the portfolio’s mean-variance trade-off with controlled turnover and drawdowns. We discuss practical implications for portfolio overlays and outline avenues for future work.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced stochastic optimization and neural network-based reinforcement learning (high math), while providing a fully implemented, deterministic pipeline with extensive backtesting, robustness checks, and public code/data for reproducibility (high rigor).

Research Flowchart

  flowchart TD
  A["Research Goal: RL-based dynamic hedging<br>for options under realistic constraints"] --> B["Data & State Construction"]
  B --> C["Simulator & Environment Design"]
  C --> D["Stochastic Actor-Critic Agent"]
  D --> E["Training & Validation"]
  E --> F["Testing & Baseline Comparison"]
  
  subgraph Data_Inputs ["Inputs"]
      B1["SPX/SPY Implied Vol Surface"]
      B2["Macro Rate Context"]
      B3["Realized Volatility"]
  end
  
  subgraph Processes ["Computational Processes"]
      C1["Leak-free environment"]
      C2["Cost-aware reward function"]
      C3["Transactional constraints"]
  end
  
  subgraph Outcomes ["Key Findings"]
      F1["Improved Sharpe vs Baselines"]
      F2["Robust to 2x transaction costs"]
      F3["Controlled turnover & drawdowns"]
  end
  
  B --- Data_Inputs
  C --- Processes
  F --- Outcomes

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a reinforcement‑learning framework for dynamic hedging of equity index option exposures with realistic transaction costs and position limits.

Source sentence
We present a reinforcement-learning (RL) framework for dynamic hedging of equity index option exposures under realistic transaction costs and position limits.

Assumes 4

  • Trading occurs on daily end‑of‑day bars, ignoring intraday inventory and execution costs. execution
    Source sentence
    The present study operates on daily bars, which ignore intraday inventory management and execution costs
  • Option Greeks are not used as state features; only surface‑level implied volatility, term structure and skew are used. model
    Source sentence
    For simplicity, we do not feed option Greeks directly into the state because of their sensitivity to intra‑day moves and model assumptions; instead we use more stable surface‑level features such as implied volatility levels, term structure, and skew.
  • Transaction costs consist of a fixed 10 bps per unit of |Δposition| plus proportional slippage of 8 bps. execution
    Source sentence
    with transaction costs of 10 bps per unit of |∆position|, proportional slippage of 8 bps
  • The market data span the 2005–2023 window for SPX/SPY implied volatility, realized volatility and rates. market
    Source sentence
    The figure below illustrates the macro/volatility regimes that motivate this work, highlighting how VIX, realized volatility, and rates evolve across the 2005–2023 window.

Data

Backtest sample: S&P 500 index / SPY · options · daily

  • SPX/SPY panel data options unknown
    2005–2023 · SPX and SPY
    Source sentence
    daily end-of-day panel data constructed from SPX/SPY implied volatility term structure, skew, realized volatility, and macro rate context

Costs

Modeled Costs are part of the reported headline results.

10 bps per unit of |Δposition| plus 8 bps proportional slippage, with a position limit of 2.0 and rebalance cadence of every 25 steps

Source sentence
with transaction costs of 10 bps per unit of |∆position|, proportional slippage of 8 bps, a position limit of pos_limit = 2.0, and rebalance cadence rebalance_every = 25

Finds 5

Reported performance (GAE deep hedging policy): Sharpe 0.50 · max drawdown 3.0%

  • The GAE policy achieves deterministic Sharpe ratios of 0.48 (train), 0.77 (validation) and 0.50 (test) under the specified cost and cadence settings. performance
    Source sentence
    deterministic evaluation yields Train Sharpe 0.48, Validation 0.77, and Test 0.50
  • Turnover stays below one full notional rotation per day on average, indicating low churn. performance
    Source sentence
    Turnover stays below one full notional rotation per day on average
  • Maximum drawdown on the test split for the standalone overlay‑equity series remains inside -3%, lower than unhedged or simple volatility‑targeting approaches. performance
    Source sentence
    Maximum drawdown on the test split for the standalone overlay-equity series remains inside -3%, materially lower than an unhedged or simple volatility targeting approach
  • Only the GAE overlay’s test‑sample Sharpe is statistically distinguishable from zero, though confidence intervals overlap with a long‑SPY benchmark. statistical
    Source sentence
    only the GAE overlay’s test Sharpe is statistically distinguishable from zero, but confidence bands versus the long‑SPY benchmark remain wide enough
  • Robustness checks show modest seed‑to‑seed variance (base test Sharpe average 0.45 ± 0.13, shifted test Sharpe 0.52 ± 0.05) and similar drawdowns when shifting the training window forward. performance
    Source sentence
    remains modest: base test Sharpe averages 0.45 (std 0.13) and shifted test Sharpe 0.52 (std 0.05). Mean portfolio-level max drawdowns for the 50/50 blend cluster around −20% (base/test) and −16% (shifted/test)

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 23 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

https://github.com/tlucius16/deep-hedging-rl — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.