Paper: arXiv 2510.04569

Authors: Jian’an Zhang

Abstract

We formulate option market making as a constrained, risk-sensitive control problem that unifies execution, hedging, and arbitrage-free implied-volatility surfaces inside a single learning loop. A fully differentiable eSSVI layer enforces static no-arbitrage conditions (butterfly and calendar) while the policy controls half-spreads, hedge intensity, and structured surface deformations (state-dependent rho-shift and psi-scale). Executions are intensity-driven and respond monotonically to spreads and relative mispricing; tail risk is shaped with a differentiable CVaR objective via the Rockafellar–Uryasev program. We provide theory for (i) grid-consistency and rates for butterfly/calendar surrogates, (ii) a primal–dual grounding of a learnable dual action acting as a state-dependent Lagrange multiplier, (iii) differentiable CVaR estimators with mixed pathwise and likelihood-ratio gradients and epi-convergence to the nonsmooth objective, (iv) an eSSVI wing-growth bound aligned with Lee’s moment constraints, and (v) policy-gradient validity under smooth surrogates. In simulation (Heston fallback; ABIDES-ready), the agent attains positive adjusted P&L on most intraday segments while keeping calendar violations at numerical zero and butterfly violations at the numerical floor; ex-post tails remain realistic and can be tuned through the CVaR weight. The five control heads admit clear economic semantics and analytic sensitivities, yielding a white-box learner that unifies pricing consistency and execution control in a reproducible pipeline.

Complexity vs Empirical Score

  • Math Complexity: 9.2/10
  • Empirical Rigor: 7.8/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a highly advanced mathematical framework integrating stochastic control, constrained reinforcement learning, and differential geometry of arbitrage-free eSSVI surfaces, with rigorous theoretical proofs (e.g., convergence rates, primal-dual theory, CVaR gradient estimators). Empirically, it uses a sophisticated agent-based simulator (ABIDES-ready) with Heston fallback, reporting concrete metrics like P&L, arbitrage violations, and tunable tail risk, indicating substantial implementation and backtesting effort.

Research Flowchart

  flowchart TD
  A["Research Goal: Unify execution, hedging, & arbitrage-free surfaces in risk-sensitive market making"] --> B{"Key Methodology: Constrained RL & Stochastic Control Bridge"}
  
  B --> C["Policy Control Heads<br>Half-spreads, Hedge intensity, Surface deformations"]
  B --> D["eSSVI Layer<br>Enforces no-arbitrage constraints<br>(Butterfly & Calendar)"]
  B --> E["Objective Function<br>Differentiable CVaR via Rockafellar-Uryasev"]
  
  C --> F["Computational Process<br>Constrained Policy Optimization<br>Primal-Dual & Gradient Flow"]
  D --> F
  E --> F
  
  F --> G["Simulations: Heston Fallback & ABIDES"]
  
  G --> H["Key Outcomes:<br>Positive P&L, Zero Calendar Violations, Tunable Tail Risk"]

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 constrained, risk‑sensitive option market‑making control framework that jointly learns half‑spreads, hedge intensity and structured eSSVI surface deformations while enforcing static no‑arbitrage via a fully differentiable eSSVI layer.

Source sentence
We formulate option market making as a constrained, risk-sensitive control problem that unifies execution, hedging, and arbitrage-free implied-volatility surfaces inside a single learning loop. A fully differentiable eSSVI layer enforces static no-arbitrage conditions (butterfly and calendar) while the policy controls half-spreads, hedge intensity, and structured surface deformations (state-dependent rho-shift and

Assumes 5

  • The action set is compact, rewards and penalties are bounded and locally Lipschitz in eSSVI parameters, and the Markov kernel is weakly continuous, ensuring well‑posedness of the CMDP. model
    Source sentence
    The action set is compact; r, ℓ, gj are bounded and locally Lipschitz in the eSSVI parameters under Assumption 1; and the Markov kernel P(·|s, a) is weakly continuous.
  • Static no‑arbitrage constraints (butterfly and calendar) are enforced only on a finite lattice; dynamic arbitrage across time is not ruled out. market
    Source sentence
    Static (not dynamic) arbitrage. The BF/CAL penalties enforce static no-arbitrage on a finite lattice (§3.2).
  • The simulation environment uses a Heston fallback model and ABIDES‑ready order‑flow intensities, abstracting away queue position, tick size, fees/rebates and latency. execution
    Source sentence
    Consequently, (i) order flow is exogenous and 16 (a) Per-step P&L distribution with tail markers (VaR5%, CVaR5%). Figure 2: Training dynamics, constraints, and behavior across episodes. summarized by smooth intensity maps (logistic in mispricing), (ii) queue position, tick discreteness, venue fees/rebates, and latency are abstracted away, and (iii) self-impact and cross-venue liquidity fragmentation are not modeled.
  • The eSSVI parameterization follows the classical SSVI butterfly‑free sufficient condition and is squashed into the open interval to guarantee static arbitrage‑free surfaces. model
    Source sentence
    The classical SSVI butterfly-free sufficient condition is 0 ≤ψm ≤ 2 1 + |ρm| for all m, (3) which we enforce by a smooth squashing into the open interval [0, ψmax(ρm)) with ψmax(ρ) = 2 1+|ρ| −εψ for a small εψ > 0 [14].
Show 1 more
  • The CVaR regularizer uses a per‑step downside CVaR proxy estimated via Monte‑Carlo scenarios with Poisson order volumes and Gaussian price perturbations. statistical
    Source sentence
    In our setting we estimate a per-step downside CVaR proxy by sampling execution and price scenarios conditioned on (st, at). For each scenario we compute ˜ PNLt = PNLquote t (˜v) + hedget ∆net t ˜∆S, 7 and plug into the smooth RU objective \ CVaR − q,t ≡min η∈R n η + 1 1 −q bE  sτ(η − ˜ PNLt)  o , (23) where sτ(x) is a temperature-τ softplus approximation of x+ = max{x, 0} and bE averages over the Monte Carlo

Data

  • Synthetic Heston simulation synthetic unknown
    Source sentence
    In simulation (Heston fallback; ABIDES-ready), the agent attains positive adjusted P\&L on most intraday segments while keeping calendar violations at numerical zero and butterfly violations at the numerical floor; ex-post tails remain realistic and can be tuned through the CVaR weight.

Costs

Modeled Costs are part of the reported headline results.

Reward includes an impact/fees term (− impact/fees) as part of the per‑step reward function.

Source sentence
The reward is Rt = spread revenue −impact/fees −λQ∥Qt+1∥2 −λarb Φarb(θt+1) | {z } butterfly & calendar surrogates and the training objective maximizes expected cumulative reward subject to a CVaR regularizer on episodic PnL: max π Eπ h X t γtRt i −λCVaR · CVaRq �−PnL(τ)  .

Finds 2

  • In simulation the agent achieves positive adjusted P&L on most intraday segments while keeping calendar violations at numerical zero and butterfly violations at the numerical floor. performance
    Source sentence
    In simulation (Heston fallback; ABIDES-ready), the agent attains positive adjusted P\&L on most intraday segments while keeping calendar violations at numerical zero and butterfly violations at the numerical floor; ex-post tails remain realistic and can be tuned through the CVaR weight.
  • Ex‑post tail risk remains realistic and can be tuned via the CVaR weight, demonstrating controllable tail‑risk shaping. performance
    Source sentence
    In simulation (Heston fallback; ABIDES-ready), the agent attains positive adjusted P\&L on most intraday segments while keeping calendar violations at numerical zero and butterfly violations at the numerical floor; ex-post tails remain realistic and can be tuned through the CVaR weight.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 7 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

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

Cited References (66)

  1. Optimal Dealer Pricing under Transactions and Return Uncertainty (1981)
  2. Bid, Ask and Transaction Prices in a Specialist Market with Heterogeneously Informed Traders (1985)
  3. Continuous Auctions and Insider Trading (1985)
  4. High-frequency trading in a limit order book (2008)
  5. Dealing with the Inventory Risk: A Solution to the Market Making Problem (2013)
  6. Algorithmic and High-Frequency Trading (2015)
  7. Buy Low, Sell High: A High Frequency Trading Perspective (2014)
  8. The Price Impact of Order Book Events: An Empirical Study (2014)
and 58 more
  1. The Flash Crash: High-Frequency Trading in an Electronic Market (2017)
  2. Empirical Market Microstructure (2007)
  3. Limit Order Books: A Survey (2013)
  4. Spectra of Some Self-Exciting and Mutually Exciting Point Processes (1971)
  5. Hawkes Processes in Finance (2015)
  6. Arbitrage-Free SVI Volatility Surfaces (2014)
  7. The Extended SSVI Volatility Surface (2019)
  8. No-Arbitrage SVI (2022)
  9. A Note on Sufficient Conditions for No Arbitrage (2005)
  10. Arbitrage-Free Implied Volatility Surfaces (2010) — link
  11. From Implied to Spot Volatilities (2009)
  12. The Moment Formula for Implied Volatility at Extreme Strikes (2004)
  13. The Pricing of Options and Corporate Liabilities (1973)
  14. Theory of Rational Option Pricing (1973)
  15. Prices of State-Contingent Claims Implicit in Option Prices (1978)
  16. A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options (1993)
  17. Deep Hedging (2019)
  18. Estimating Risks of Option Books Using Neural-SDE Market Models (2023)
  19. Arbitrage-Free Neural-SDE Market Models (2021) — link
  20. Deep Hedging: Learning to Simulate Equity Option Markets (2019) — link
  21. Optimization of Conditional Value-at-Risk (2000)
  22. Conditional Value-at-Risk for General Loss Distributions (2002)
  23. Some Remarks on the Value-at-Risk and the Conditional Value-at-Risk (2000)
  24. Algorithms for CVaR Optimization in MDPs (2014)
  25. Policy Gradient for Coherent Risk Measures (2015)
  26. Risk-Constrained Reinforcement Learning with Percentile Risk Criteria (2018)
  27. Near-Minimax-Optimal Risk-Sensitive Reinforcement Learning with CVaR (2023)
  28. Risk-Sensitive Reward-Free Reinforcement Learning with CVaR (2024)
  29. ABIDES: Towards High-Fidelity Market Simulation for AI Research (2020) — link
  30. ABIDES-Gym: Gym Environments for Multi-Agent Discrete Event Simulation of Limit Order Books (2021)
  31. Simulation of Arbitrage-Free Implied Volatility Surfaces (2023)
  32. Risk-Sensitive Markov Decision Processes (1972)
  33. A Mean Field Game Between Informed Traders and a Broker — → 2401.05257
  34. Brokers and Informed Traders: Dealing with Toxic Flow and Extracting Trading Signals (2025)
  35. Market Making with Exogenous Competition — → 2407.17393
  36. Proximal Policy Optimization Algorithms (2017)
  37. High-Dimensional Continuous Control Using Generalized Advantage Estimation (2016)
  38. On the Coherence of Expected Shortfall (2002)
  39. Coherent Measures of Risk (1999)
  40. Constrained Markov Decision Processes (1999)
  41. Markov Decision Processes: Discrete Stochastic Dynamic Programming (1994)
  42. Risk-Averse Dynamic Programming for Markov Decision Processes (2010)
  43. Lectures on Stochastic Programming: Modeling and Theory (2014)
  44. Constrained Policy Optimization (2017)
  45. Reward Constrained Policy Optimization (2019)
  46. A Long-Memory Model for Order Flow (2015)
  47. Trades, Quotes and Prices: Financial Markets Under the Microscope (2018)
  48. Monte Carlo Methods in Financial Engineering (2004)
  49. Trust Region Policy Optimization (2015)
  50. A Natural Policy Gradient (2002)
  51. Deep Reinforcement Learning that Matters (2018)
  52. Adam: A Method for Stochastic Optimization (2014) — link
  53. Options, Futures, and Other Derivatives (2018)
  54. The Complete Guide to Option Pricing Formulas (2007)
  55. Paul Wilmott on Quantitative Finance (2006)
  56. Variational Analysis (1998)
  57. Policy Gradient Methods for Reinforcement Learning with Function Approximation (2000)
  58. Simulation-Based Optimization of Markov Reward Processes (2001)

2 of 66 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.