Searcharxiv⌕ Search

arXiv · 2611.10956

Risk-Sensitive Reinforcement Learning for an Options Market Maker: An Empirical Study with Real S&P 500 Options

Abstract

Options market makers absorb client order flow and must hedge the resulting inventory risk under trading frictions. We study whether an option hedging reinforcement learning (RL) agent trained on historical market data outperforms three alternatives in tail-risk protection: (i) traditional delta hedging, (ii) simple hedge-ratio rules, and (iii) the same agent trained in simulated markets. Most RL hedging agents are trained in model-generated markets that do not capture the crashes and volatility regimes present in real data. We introduce an empirical replay environment using S&P 500 option data from 1996 to 2017 and compare against a matched geometric Brownian motion (GBM) environment, a stochastic-volatility (SABR) simulator, and a hybrid that values historical paths with Black-Scholes, all sharing the same portfolio mechanics, stochastic arrivals, costs, and test period. We train a distributional Twin Delayed Deep Deterministic Policy Gradient (TD3) agent and evaluate it out of sample over 2018-2023. The agent trained on empirical data reduces Conditional Value-at-Risk (CVaR) tail loss by 3.1% compared to delta hedging and outperforms every hedge-ratio rule chosen during the training period and every alternative training environment. The reduction reaches 8.9% in volatile and falling markets, but reverses to a cost of 7-10% in rising markets. Empirical training induces a state-dependent short net delta: the policy under-hedges negative-delta books, most strongly after prices have fallen. This tilt disappears with GBM training and is weaker with SABR or hybrid training. These findings suggest that exposure to historical market dynamics can materially change a tail-risk hedging policy, although the evidence rests on a single test path and the edge is regime-conditional.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Jacky Chen, Ing-Haw Cheng, John Hull, Landis Li, Zissis Poulos, Justin Shi, Andreas Veneris, Zeyu Wang. 2026-11-07. Risk-Sensitive Reinforcement Learning for an Options Market Maker: An Empirical Study with Real S&P 500 Options. https://searcharxiv.com/abs/2611.10956

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Loss-Given-Default Modeling by Post-Last Passage Time Process

We bridge the dichotomy between structural and intensity-based credit models by defining default via the last passage time ($λ_α$) of a linear diffusion. Because $λ_α$ is not a stopping time, this approach naturally generates "surprise" defaults within an asset-value framework without artificial filtration enlargements. Mathematically, we establish two general results for transient linear diffusions: we derive explicit Laplace transforms for occupation times and characterize post-passage dynamics. Practically, our two-parameter model calibrates seamlessly to default probabilities and Credit Default Swap (CDS) spreads. We obtain firm-specific Loss-Given-Default (LGD) densities - resolving a long-standing open question - and quantify sentiment divergence between equity and credit markets, yielding a unified valuation framework.

q-fin.RM↗

An axiomatic foundation for optimized certainty equivalents

Optimized Certainty Equivalents (OCEs) are an important class of risk measures for optimization and risk management that includes the Mean--Variance criterion, Expected Shortfall, the Entropic Risk Measure, and expectation. Despite their importance, economic axioms have not characterized OCEs. We propose a new axiom called Fair-Coin Compensation, which characterizes OCEs together with Strong Risk Aversion and Normalization. We interpret Fair-Coin Compensation through the fair assignment of two losses by a coin toss: transfers between two participants can remove the additional risk created by the random assignment. Adding further axioms yields sharper characterizations. Monotonicity yields standard OCEs; Positive Homogeneity, together with Prudence, yields Expected Shortfall; and Fair-Coin Independence yields the Entropic Risk Measure. We also axiomatize the Mean--Variance criterion using our novel axiom of Equal-Mean Randomization, which states that random assignment of two losses with the same expected value does not produce a risk measure above the average of their risk measures.

q-fin.RM↗

On the Structure of Risk Contribution: A Leave-One-Out Decomposition into Inherent and Correlation Risk

Risk Contribution (RC) is a widely used risk attribution metric that assigns each position a share of portfolio volatility, with contributions that sum to total portfolio volatility. This paper proposes a leave-one-out decomposition of RC into two interpretable components at the position level: an inherent term (position's own scaled variance) and a correlation term (covariance with the other holdings). Unlike the X-sigma-rho framework, which uses the position's correlation with the full portfolio, ICD uses covariance with the other holdings, yielding qualitatively different diagnostic signals when a position is large relative to the portfolio. The contribution is the interpretive framing and an exact identity linking the components to incremental volatility (iVol), which allows the discrete condition under which a position reduces portfolio volatility to be stated in closed form. That condition differs from the marginal condition and the two can disagree for over-hedged or levered positions. The decomposition also inherits RC's strict additivity. Empirical illustrations on three multi-asset portfolios show the components shift differently across regimes, distinguishing stand-alone volatility spikes from correlation breakdown, and demonstrate value for risk reporting, stress testing, and attribution.

q-fin.RM↗