Diffusion models for dynamic volatility surface generation and data-driven hedging
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "Diffusion models for dynamic volatility surface generation and data-driven hedging".
Warren: Here's the gist of the paper. Diffusion models for dynamic volatility surface generation and data-driven hedging develop a diffusion-model framework for dynamic implied-volatility surface generation and evaluate its economic usefulness through data-driven hedging The proposed models generate coherent surface trajectories while capturing both cross-sectional and temporal dependence Framework Components The framework consists of two models, AD-Seq-Vol and AD-Seq-Vol-FT AD-Seq-Vol jointly learns the conditional evolution of the underlying asset return and the high-dimensional implied volatility surface, generating adapted multiperiod scenarios by sequentially updating the realized market history AD-Seq-Vol-FT further incorporates option-market structure through posttraining penalties for violations of static noarbitrage conditions Model Development and Constraints The AD-Seq framework is adapted to high dimensional implied volatility surfaces by learning a sequence of conditional generators adapted to evolving market information AD-Seq-Vol produces fewer and less severe staticarbitrage violations than the training data and the GANbased benchmark AD-Seq-Vol-FT reduces these violations to nearly zero by penalizing calendarspread, callspread, and butterflyspread constraints Economic Utility through Hedging The economic usefulness of the generated conditional scenarios is evaluated through data driven hedging using an optimizationbased hedging framework Compared to classical and datadriven benchmarks, the diffusionbased hedges maintain tracking errors near zero, substantially reduce tail risk, and exhibit particularly stable performance during the COVID19 market disruption Key Findings in Hedging Performance The AD-Seq-Vol model delivers the strongest and most consistent performance in hedging, with tracking errors staying centered near zero The fine tuned variant AD-Seq-Vol-FT attains the lowest 1% VaR in both windows, demonstrating a tail advantage that is not limited to the COVIDperiod alone The results show that diffusion models combine a nonadversarial scorematching objective with flexible historyconditioned generation, which may offer greater stability for highdimensional financial scenarios Future Directions A natural extension is to move beyond SPX to a joint, multiasset surface generator, spanning multiple indices and singlename options Arbitrageaware finetuning could be extended from static, persurface conditions to dynamic noarbitrage constraints enforced along the trajectory Methodologically, distillation or fewer-step samplers could reduce the cost of sequential diffusion, broadening the framework’s applicability to highfrequency settings References Beatrice Acciaio, Stephan Eckstein, and Songyan Hou.
Ruth: Timecausal vae: Robust financial time series generator.
Warren: SIAM Journal on Financial Mathematics, 17(1):245–279, 2026 Damien Ackerer, Natasa Tagasovska, and Thibault Vatter.
Ruth: Deep smoothing of the implied volatility surface.
Title and authors: Ruth: So we're looking at this paper, Diffusion models for dynamic volatility surface generation and data-driven hedging. It sounds like they're doing something pretty ambitious with these diffusion models to create a dynamic implied volatility surface and then using that for hedging.
Warren: Yeah, it’s about taking that high-dimensional surface and making it evolve over time, not just taking a snapshot of it, which is what most existing methods do.
Iris: The core idea seems to be using these diffusion models to learn how the implied volatility surface changes sequentially based on past market history.
Ruth: Exactly, they introduce two main models here: AD-Seq-Vol and then the AD-Seq-Vol with a finetuning component called AD-Seq-Vol FT.
Warren: The first one, AD Seq Vol, it learns the relationship between how asset returns evolve and how that high dimensional implied volatility surface changes as they go along.
Iris: And then they take that learned sequence of conditional generators and update it based on the realized market history to generate these multiperiod scenarios.
Ruth: That sequential updating part is key because it lets them capture both the cross sectional dependence—how different options move together—and the temporal dependence, which is how things change over time.
Warren: Right, so instead of just modeling one static surface, they're generating whole trajectories that reflect how the market itself moves and changes as you look further into the future.
Iris: And then there's this AD Seq Vol FT model which adds some structure back in by using post-training penalties to keep things close to reality.
Ruth: That’s where they enforce static noarbitrage conditions, like calendar spread, call spread, and butterfly spread constraints to make sure the generated surfaces are realistic.
Warren: And the paper shows that AD Seq Vol produces fewer and less severe static arbitrage violations compared to the training data or a standard GAN based benchmark.
Iris: The AD Seq Vol FT model is even better because it reduces those arbitrage violations down to nearly zero by penalizing those specific spread constraints.
Ruth: That’s a big step toward making these simulated surfaces more economically useful for actual trading decisions, not just statistically interesting.
Title and authors: Warren: And they don't stop there; they take these generated conditional scenarios and feed them into an optimization based hedging framework to see how good the model is actually at hedging.
Iris: The results show that when you use these diffusion based hedges, the tracking errors stay very close to zero compared to classical or even some other data driven benchmarks.
Ruth: And what’s particularly interesting is how this approach handles tail risk during events like the COVID-nineteen market disruption, showing particularly stable performance during that time <ref:2609.13402#pg1>.
Warren: The AD Seq Vol model itself delivers the strongest and most consistent hedging performance, with tracking errors centered near zero across the board.
Iris: And when you look at the fine tuned variant, AD Seq Vol FT, it achieves the lowest one percent value at risk in both windows of testing.
Ruth: That tail advantage isn't just limited to that COVID period; it shows up consistently across different market conditions.
Warren: The authors suggest that diffusion models combine a non adversarial scoring matching objective with flexible history conditioned generation, which seems to give them greater stability for these high dimensional financial scenarios.
Iris: They also point out that the framework can be integrated into an optimization based hedging procedure where you solve one period conditional hedging problem with a rebalancing cost penalty.
Ruth: That’s a practical application; it moves this from just generating data to actually creating a usable trading tool that accounts for the cost of rebalancing.
Warren: But the model can also generate multi day trajectories recursively, where each new surface you generate updates the history that conditions the next day's surface, which is different from simpler one step generators like VolGAN.
Iris: That recursive generation is what lets them build those coherent surface trajectories conditional on realized market history.
Ruth: So what are the real-world implications here? If this works as well as they claim, it means we can generate much more realistic volatility surfaces for pricing options and managing risk dynamically.
Warren: It means better hedging strategies that track closely to the actual market, which is valuable when things get choppy.
Title and authors: Iris: The next step they suggest is moving beyond just the SPX index to a joint multiasset surface generator spanning multiple indices and single name options.
Ruth: That would be a huge expansion of the framework, allowing us to model dependencies across different asset classes at once.
Warren: And they also mention that arbitrage aware finetuning could be extended from static conditions to dynamic noarbitrage constraints enforced along the trajectory itself.
Iris: Methodologically, they also explore distillation or fewer step samplers to reduce the computational cost of sequential diffusion, which would make it applicable in high frequency settings.
Ruth: So, to wrap up on Diffusion models for dynamic volatility surface generation and data-driven hedging: this work shows a way to generate evolving surfaces that respect market structure and perform well in hedging tests.
Warren: The main win is the strong performance of AD Seq Vol FT in terms of tracking errors near zero and reduced tail risk during stress periods.
Iris: It's important to remember, though, that they focus on SPX option data from two thousand to two thousand twenty-three for their evaluation <ref:2609.13402#pg1,SPX option data from 2000 to 2023>.
Ruth: And the authors acknowledge that the method does not necessarily cover every single possible financial scenario or market regime outside of those training conditions.
Warren: So, while this is a solid framework for dynamic surface generation and hedging evaluation, we have to keep an eye on how these models generalize when you step outside of that specific historical data range.
Iris: That’s the limitation they flag: it’s currently tied to the SPX option data from two thousand to two thousand twenty-three for their performance assessment <ref:2609.13402#pg1,SPX option data from 2000 to 2023>.
Ruth: It really is a sophisticated tool, moving away from static models towards something that evolves with the market history itself.
Warren: And for us, it suggests we need to look at how these diffusion based hedges perform when we introduce more complex dependencies or try to run them across a wider range of assets.
Iris: So, this paper provides a strong foundation for building more robust and dynamically consistent financial simulators going forward.
The paper's summary: Ruth: So, this paper is about using diffusion models to generate volatility surfaces that move over time, not just static pictures of them.
Warren: Right, they’re making these surfaces dynamic by feeding in the actual market history as things happen day by day.
Iris: The core idea is that the diffusion model learns how the underlying asset returns evolve and simultaneously learns how that high dimensional implied volatility surface changes with every step.
Ruth: And then they build a sequence of conditional generators based on all that past information to create these multi period scenarios.
Warren: It’s about capturing both the way things are related across different options right now and how those relationships change over time in a coherent way.
Iris: They also have this model called AD Seq Vol FT, which uses penalties to make sure the generated surfaces don't violate basic noarbitrage rules, like calendar spread or butterfly spread constraints.
Ruth: That’s an important part of it; they aren't just generating pretty pictures, they’re making sure these surfaces are economically plausible.
Warren: And the results show that this diffusion based hedging approach keeps tracking errors near zero when compared to traditional methods.
Iris: The paper found that AD Seq Vol FT achieves a very low one percent value at risk in both windows tested, which suggests a real tail risk advantage.
Ruth: So, what this means for us is that we might be able to generate much more realistic volatility surfaces for pricing options and managing risk dynamically.
Warren: It means better hedging strategies that track closely to the actual market when things get volatile.
Iris: The authors point out that they used SPX option data from two thousand to two thousand twenty-three for their testing, so we have to be careful about how well it works outside of those specific conditions.
Ruth: Exactly. It’s a strong framework for generating evolving surfaces and testing hedging performance under specific historical constraints.
Warren: Next, they look at how this whole system fits into an optimization based hedging procedure where you solve a one period conditional problem with a rebalancing cost penalty.
Iris: And they mention that the system can generate these trajectories recursively, using each new surface to update the history for the next day's generation.
Ruth: That recursive updating is what lets them build those coherent multi day paths, which is different from simpler generators that only make one step at a time.
Warren: So we’ve seen how they generate these surfaces and how they test their hedging performance under stress, but now we need to see if this works when you try to scale it up.
The paper's improvements: Tom: So, we're looking at how these authors think we can make this diffusion model even better after they show their initial results for dynamic volatility surface generation and hedging.
Ruth: They suggest a few ways to take this framework further that move beyond what they’ve already done.
Warren: Right, they look at making it more flexible and robust when applying it to different markets, not just one index like the SPX.
Iris: They propose moving beyond just the SPX to a joint multiasset surface generator that covers multiple indices and single name options at once.
Ruth: That would mean we could model how different asset classes are related to each other in a way that's much richer than focusing on just one index.
Warren: And they suggest that arbitrage aware finetuning could be extended from just static conditions to dynamic noarbitrage constraints enforced all along the entire trajectory.
Iris: That means instead of checking if the final surface is arbitrage free, you check for it at every single point as the model generates the path.
Ruth: That addresses a big assumption they made earlier about arbitrage being a fixed problem; this makes it a continuous challenge throughout time.
Warren: And on the engineering side, they talk about ways to reduce the cost of these sequential diffusion steps, like using distillation or fewer step samplers.
Iris: That’s smart because if generating one step takes too long, you can’t really run this in real time for high frequency trading.
Ruth: So they are trying to balance the need for a highly realistic model with the practical need for speed and efficiency.
Warren: This points toward making the whole system more applicable in faster environments where you need quick answers based on evolving market data.
Iris: It’s about finding that sweet spot between high fidelity simulation and computational feasibility.
Ruth: What this implies is that if we can make these models faster and more adaptable, they could become a standard tool for risk management across much wider financial markets.
Conclusion: Ruth: So to wrap up, we’ve seen how this paper uses diffusion models to generate dynamic volatility surfaces that evolve over time, and they test this generation through data driven hedging using a framework called AD Seq Vol and its fine tuned version AD Seq Vol FT.
Warren: The main point is that these models produce surfaces with tracking errors near zero when hedging against the market, which is pretty strong compared to classical methods.
Iris: I think it’s important to remember the constraints they put on those surfaces, specifically how they use penalties for spreads like calendar spread and butterfly spread to keep things arbitrage free.
Ruth: Exactly. The model assumes that these constraints are important because they reflect real market structure, but we have to ask what happens if those market structures change suddenly or if agents start ignoring them.
Warren: And from a trading side, the results show that this approach handles tail risk during major events like COVID-nineteen particularly well, which is crucial when you’re trying to figure out where the actual drawdown comes from.
Iris: I would add that since they trained on SPX option data from two thousand to twenty twenty-three, we have to be careful about how well this works when you try to apply it outside of that specific historical window.
Ruth: Right, so the paper shows a very solid way to generate these evolving surfaces and test hedging performance under those conditions.
Warren: It’s a good tool for seeing how deep the risk goes in complex scenarios, even if we can’t use it for every single market condition we might encounter.
Iris: I think the identification here is sound because they are checking the variation against realized market history, which is what really drives the change in those surfaces.
Ruth: It’s a significant step away from static models, moving toward something that reflects how markets actually behave moment by moment.
Warren: We've seen how they build these dynamic trajectories and how they use them to optimize a hedging cost penalty, which is where the practical application lies for running this strategy.
Iris: So the main implication is that we have a new way to simulate complex volatility dynamics and see if our risk management strategies hold up under those conditions.
Ruth: And it sets the stage for future work, like expanding it to multiasset surfaces or making the generation process faster for high frequency needs.
Warren: It’s definitely got us thinking about how much computational cost we can afford to run these complex simulations in a live trading system.
Stanford University · J.P. Morgan QTR AI Research
q-fin.CP, q-fin.PM, q-fin.RM
Submitted: 2026-09-11
Updated: 2026-11-10
Code: https://github.com/yinbinhan/volatility-surface-simulation
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 97/100
The gist: Diffusion models for dynamic volatility surface generation and data-driven hedging develop a diffusion-model framework for dynamic implied-volatility surface generation and evaluate its economic…
Key concepts
- Diffusion Models
- These are generative AI models that work by gradually adding noise to data until it becomes pure noise, and then learning to reverse this process. In this context, they are used to generate realistic sequences of implied volatility surfaces based on historical market data.
- AD-Seq-Vol
- This model jointly learns the conditional evolution of asset returns and the high-dimensional implied volatility surface. It generates multiperiod scenarios by sequentially updating the realized market history, allowing it to capture how both asset prices and volatility change over time in a dependent manner.
- AD-Seq-Vol-FT
- This is an enhanced version that incorporates constraints from option market structure, such as noarbitrage conditions. It uses posttraining penalties to force the generated surfaces to strictly adhere to constraints like calendar spread and butterfly spread, resulting in much more realistic and arbitrage-free trajectories.
- Data-Driven Hedging
- This involves using the generated scenarios to test hedging strategies. An optimization-based framework is used where the model's output helps solve a one-period conditional hedging problem, aiming to minimize rebalancing costs while maintaining accurate risk management.
Terminology
Summary
Diffusion models for dynamic volatility surface generation and data-driven hedging develop a diffusion-model framework for dynamic implied-volatility surface generation and evaluate its economic usefulness through data-driven hedging The proposed models generate coherent surface trajectories while capturing both cross-sectional and temporal dependence
Framework Components
The framework consists of two models, AD-Seq-Vol and AD-Seq-Vol-FT AD-Seq-Vol jointly learns the conditional evolution of the underlying asset return and the high-dimensional implied volatility surface, generating adapted multiperiod scenarios by sequentially updating the realized market history AD-Seq-Vol-FT further incorporates option-market structure through posttraining penalties for violations of static noarbitrage conditions
Model Development and Constraints
The AD-Seq framework is adapted to high dimensional implied volatility surfaces by learning a sequence of conditional generators adapted to evolving market information AD-Seq-Vol produces fewer and less severe staticarbitrage violations than the training data and the GANbased benchmark AD-Seq-Vol-FT reduces these violations to nearly zero by penalizing calendarspread, callspread, and butterflyspread constraints
Economic Utility through Hedging
The economic usefulness of the generated conditional scenarios is evaluated through data driven hedging using an optimizationbased hedging framework Compared to classical and datadriven benchmarks, the diffusionbased hedges maintain tracking errors near zero, substantially reduce tail risk, and exhibit particularly stable performance during the COVID19 market disruption
Key Findings in Hedging Performance
The AD-Seq-Vol model delivers the strongest and most consistent performance in hedging, with tracking errors staying centered near zero The fine tuned variant AD-Seq-Vol-FT attains the lowest 1% VaR in both windows, demonstrating a tail advantage that is not limited to the COVIDperiod alone The results show that diffusion models combine a nonadversarial scorematching objective with flexible historyconditioned generation, which may offer greater stability for highdimensional financial scenarios
Future Directions
A natural extension is to move beyond SPX to a joint, multiasset surface generator, spanning multiple indices and singlename options Arbitrageaware finetuning could be extended from static, persurface conditions to dynamic noarbitrage constraints enforced along the trajectory Methodologically, distillation or fewer-step samplers could reduce the cost of sequential diffusion, broadening the framework’s applicability to highfrequency settings
References
Beatrice Acciaio, Stephan Eckstein, and Songyan Hou. Timecausal vae: Robust financial time series generator. p. 2 Generative models for financial data. VAE-based methods have been used for multivariate timeseries and arbitrage-free option-surface simulation (Desai et al.Where the paper says it
Damien Ackerer, Natasa Tagasovska, and Thibault Vatter. Deep smoothing of the implied volatility surface. p. 2 We adapt it to the joint generation of underlying return and high-dimensional
surface, introduce arbitrage-aware post-training, and evaluate the resulting scenarios through datadriven hedging. Deep-learning-based implied-volatility and option simulation.Where the paper says it
Alexandre Alouadi, Baptiste Barreau, Laurent Carlier, and Huyˆen Pham. Robust time series generation via schr¨odinger bridge: a comprehensive evaluation. p. 2 , 2023; Cont et al. , 2026); Schr¨odinger-bridge methods provide a related stochasticcontrol formulation (Wang et al.Where the paper says it
Fischer Black and Myron Scholes. The pricing of options and corporate liabilities. p. 1 Specifically, we penalize violations of the calendar-spread,
call-spread, and butterfly-spread constraints induced by the generated surfaces. In our empirical
study of SPX options, AD-Seq-Vol already produces fewer and smaller static-arbitrage violations
than the training data and the GAN-based benchmark, while AD-Seq-Vol-FT reduces these
violations to nearly zero.Where the paper says it
Hans Buehler, Lukas Gonon, Josef Teichmann, and Ben Wood. Deep hedging. p. 1 Classical model-based approaches rely
on pre-specified asset-price dynamics and construct hedge positions using local sensitivities such as
delta and vega (Black and Scholes, 1973; Hull, 2022). Deep hedging methods instead learn hedging
policies from market scenarios by minimizing a risk measure of terminal hedging error (Buehler
et al.Where the paper says it
Hans Buehler, Blanka Horvath, Terry Lyons, Imanol Perez Arribas, and Ben Wood. Generating financial markets with signatures. p. 1 By updating
this conditioning information along each generated path, it produces adapted multi-period scenarios
that capture both the cross-sectional dependence across strikes and maturities and the temporal
dependence of surface dynamics. We introduce AD-Seq-Vol-FT, which incorporates the financial structure of option markets through arbitrage-aware fine-tuning.Where the paper says it
Hans Buehler, Blanka Horvath, Anastasis Kratsios, Yannick Limmer, and Raeid Saqur. SANOS: Smooth strictly arbitrage-free nonparametric option surfaces. p. 2 More
recently, Buehler et al. (2026a) construct smooth no-arbitrage option price surfaces with a nonparametric model, and Buehler et al. acero}@jpmorgan. com
1
arXiv:2609.Where the paper says it
Where the paper says it
Hans B¨uhler, Blanka Horvath, Terry Lyons, Imanol Perez Arribas, and Ben Wood. A data driven market simulator for small data environments. p. 2 , 2026; Ning et al. Signature-based simulators cover small-data market
generation and joint spot-option simulation (Buehler et al. A complementary line of work uses neural-SDE approaches to model arbitrage-free joint dynamics of vanilla
options (Cohen et al. , 2023), calibrate local stochastic volatility (Cuchiero et al.Where the paper says it
Where the paper says it
Borui Cai, Shuiqiao Yang, Longxiang Gao, and Yong Xiang. Hybrid variational autoencoder for time series forecasting. p. 0 1 Introduction
Generative models are widely used in finance to produce synthetic scenarios for market-risk evaluation (Flaig and Junike, 2022), tail-risk-event simulation (Cont et al. , 2026), time-series forecasting
(Vuleti´c et al.Where the paper says it
Haoyang Cao, Minshuo Chen, Yinbin Han, and Renyuan Xu. Diffusion models for adaptive sequential data generation. p. 1 Our work and main contributions. We propose a family of Adaptive Sequential Diffusion
Models tailored to dynamic implied-volatility surface generation. acero}@jpmorgan. com
1
arXiv:2609.Where the paper says it
Where the paper says it
Jay Cao, Jacky Chen, Soroush Farghadani, John Hull, Zissis Poulos, Zeyu Wang, and Jun Yuan. Gamma and vega hedging using deep distributional reinforcement learning. p. 3 Deep hedging (Buehler et al. , 2019), related
reinforcement learning methods (Cao et al.Where the paper says it
Vedant Choudhary, Sebastian Jaimungal, and Maxime Bergeron. Funvol: multiasset implied volatility market simulator using functional principal components and neural sdes. p. 3 In contrast, Cont and Vuleti´c (2025) learn a conditional market generator and obtain hedge
positions through one-period conditional risk minimization. We adopt this local formulation, using
AD-Seq-Vol and AD-Seq-Vol-FT to generate joint next-day scenarios for underlying returns
and implied-volatility surfaces and optimize hedge positions with a rebalancing-cost penalty.Where the paper says it
Samuel N. Cohen, Christoph Reisinger, and Sheng Wang. p. 2 , 2023; Cont et al. , 2026); Schr¨odinger-bridge methods provide a related stochasticcontrol formulation (Wang et al. (2021) use
normalizing flows to jointly simulate spot and option markets across multiple assets. A complementary line of work uses neural-SDE approaches to model arbitrage-free joint dynamics of vanilla
options (Cohen et al.Where the paper says it
Where the paper says it
Andrea Coletta, Joseph Jerome, Rahul Savani, and Svitlana Vyetrenko. Conditional generators for limit order book environments: Explainability, challenges, and robustness. p. 2 , 2026; Wiese et al. GAN-based methods span synthetic time series, conditional sequential generation, tailrisk simulation, trading-strategy generation, order-flow generation, and implied-volatility surfaces
(Mirza and Osindero, 2014; Esteban et al.Where the paper says it
Rama Cont and Jos´e da Fonseca. Dynamics of implied volatility surfaces. p. 7 4 Experiments
In this section, we first describe the data preprocessing, model specification, and training procedure of AD-Seq-Vol for implied-volatility surfaces. We then assess the generated trajectories
in terms of their qualitative surface dynamics and static-arbitrage violations, including the effect of arbitrage-aware fine-tuning.Where the paper says it
Rama Cont and Milena Vuleti´c. Data driven hedging with generative models. p. 1 , 2019; Fran¸cois et al. The data-driven hedging framework of Cont and Vuleti´c (2025)
directly connects scenario generation to hedge construction: it first trains a conditional generative
model on market data and then uses the generated scenarios to select hedging instruments and
optimize hedge ratios.Where the paper says it
Rama Cont, Mihai Cucuringu, Renyuan Xu, and Chao Zhang. Tail-gan: Learning to simulate tail risk scenarios. p. 0 We then integrate the generated conditional scenarios
into an optimization-based hedging framework. Compared to a range of classical and datadriven benchmarks, the diffusion-based hedges maintain tracking errors near zero, substantially
reduce tail risk, and exhibit particularly stable performance during the COVID-19 market disruption.Where the paper says it
Christa Cuchiero, Wahid Khosrawi, and Josef Teichmann. A generative adversarial network approach to calibration of local stochastic volatility models. p. 2 A complementary line of work uses neural-SDE approaches to model arbitrage-free joint dynamics of vanilla
options (Cohen et al. , 2023), calibrate local stochastic volatility (Cuchiero et al.Where the paper says it
Mark H. A. Davis and David G. p. 3 As shown in Corollaries 4. 3 from Davis and Hobson (2007), absence of static arbitrage among options with strikes and
maturities defined by (m, τ ) is equivalent to the following three conditions: 1 Implied volatility surfaces
Consider an asset and let St be the price of the underlying at time t, K the strike, T the expiration
date, and τ = T − t > 0 the time-to-maturity. For a call option with moneyness m = K/St and
market price ct(m, τ ) in the no-arbitrage range (St − Ke−rτ )+ < ct(m, τ ) < St
, its Black–Scholes
implied volatility is the unique σt(m, τ ) ∈ (0, ∞) such that
CBS(St
, K, τ, r, σt(m, τ )) = StN(d1) − Ke−rτN(d2) = ct(m, τ ),
where
d1 =
− ln m + τ
r + σ
2
t
(m, τ )/2
σt(m, τ )
√
τ
, d2 =
− ln m + τ
r − σ
2
t
(m, τ )/2
σt(m, τ )
√
τ
.Where the paper says it
Where the paper says it
Solveig Flaig and Gero Junike. Scenario generation for market risk models using generative neural networks. p. 1 , 2019; Fran¸cois et al. The data-driven hedging framework of Cont and Vuleti´c (2025)
directly connects scenario generation to hedge construction: it first trains a conditional generative
model on market data and then uses the generated scenarios to select hedging instruments and
optimize hedge ratios.Where the paper says it
Hans F¨ollmer and Martin Schweizer. Hedging of contingent claims under incomplete information. p. 1 ,
2026), which learns a sequence of conditional generators adapted to evolving market information. Hedging provides a natural decision-based evaluation of the resulting generator, since hedge
selection depends on the conditional joint evolution of the underlying asset and the volatility surface,
rather than on the realism of individual surfaces in isolation. As shown in Corollaries 4. 3 from Davis and Hobson (2007), absence of static arbitrage among options with strikes and
maturities defined by (m, τ ) is equivalent to the following three conditions:Where the paper says it
Where the paper says it
Pascal Fran¸cois, Genevi´eve Gauthier, Fr´ed´eric Godin, and Carlos Octavio P´erez Mendoza. Enhancing deep hedging of options with implied volatility surface feedback information. p. 2 We adapt it to the joint generation of underlying return and high-dimensional
surface, introduce arbitrage-aware post-training, and evaluate the resulting scenarios through datadriven hedging. Deep-learning-based implied-volatility and option simulation. acero}@jpmorgan. com
1
arXiv:2609.Where the paper says it
Where the paper says it
Rao Fu, Jie Chen, Shutian Zeng, Yiping Zhuang, and Agus Sudjianto. Time series simulation by conditional generative adversarial net. p. 0 1 Introduction
Generative models are widely used in finance to produce synthetic scenarios for market-risk evaluation (Flaig and Junike, 2022), tail-risk-event simulation (Cont et al. , 2026), time-series forecasting
(Vuleti´c et al.Where the paper says it
Jim Gatheral. The Volatility Surface: A Practitioner’s Guide. p. 0 Diffusion models for dynamic volatility surface generation and
data-driven hedging §
Yinbin Han∗ Jack Yuxiang Zhang† Manuel Torres‡ Fernando Acero‡
Renyuan Xu∗
September 18, 2026
Abstract
We develop a diffusion-model framework for dynamic implied-volatility surface generation
and evaluate its economic usefulness through data-driven hedging. The framework consists
of two models.Where the paper says it
Mohamed Hamdouche, Pierre Henry-Labordere, and Huyˆen Pham. Generative modeling for time series via Schr¨odinger bridge. p. 2 , 2023; Cont et al. , 2026); Schr¨odinger-bridge methods provide a related stochasticcontrol formulation (Wang et al. acero}@jpmorgan. com
1
arXiv:2609.Where the paper says it
Where the paper says it
Yinbin Han, Meisam Razaviyayn, and Renyuan Xu. Stochastic control for fine-tuning diffusion models: Optimality, regularity, and convergence. p. 1 We propose a family of Adaptive Sequential Diffusion
Models tailored to dynamic implied-volatility surface generation. The family consists of ADSeq-Vol, which learns the conditional evolution of the underlying asset and the implied-volatility
surface, and AD-Seq-Vol-FT, an arbitrage-aware variant obtained by fine-tuning the diffusion
model against static no-arbitrage violations.Where the paper says it
Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. p. 8 Optimization uses
AdamW for 1000 epochs, batch size 256, initial learning rate 7 × 10−5
, weight decay 10−2
, and a
warm-up cosine learning-rate schedule. For sampling, we use the classic denoising diffusion probabilistic model (DDPM) sampler (Ho et al.Where the paper says it
John C. Hull. Options, Futures, and Other Derivatives. p. 1 Specifically, we penalize violations of the calendar-spread,
call-spread, and butterfly-spread constraints induced by the generated surfaces. In our empirical
study of SPX options, AD-Seq-Vol already produces fewer and smaller static-arbitrage violations
than the training data and the GAN-based benchmark, while AD-Seq-Vol-FT reduces these
violations to nearly zero.Where the paper says it
James M. Hutchinson, Andrew W. p. 2 We then review learning-based approaches to hedging, distinguishing their
sources of training data and their local (one-step ahead) or global (full-horizon) optimization objectives. Early work of Hutchinson et al. Their score-matching objectives avoid adversarial optimization and have
demonstrated strong performance on modeling high-dimensional, multimodal distributions. We study whether diffusion models can capture the cross-sectional and temporal structure of
financial data, focusing on implied-volatility surfaces, a central input to derivative pricing, hedging
and risk management (Gatheral, 2011).Where the paper says it
Where the paper says it
Adriano Koshiyama, Nick Firoozye, and Philip Treleaven. Generative adversarial networks for financial trading strategies fine tuning and combination. p. 0 , 2024), and implied-volatility surface generation (Vuleti´c and Cont, 2024). Most prior
applications have relied on generative adversarial networks (GANs) and variational autoencoders
(VAEs) (Wiese et al.Where the paper says it
Shujian Liao, Hao Ni, Marc Sabate-Vidales, Lukasz Szpruch, Magnus Wiese, and Baoren Xiao. Sigwasserstein gans for conditional time series generation. p. 2 , 2026; Wiese et al. GAN-based methods span synthetic time series, conditional sequential generation, tailrisk simulation, trading-strategy generation, order-flow generation, and implied-volatility surfaces
(Mirza and Osindero, 2014; Esteban et al.Where the paper says it
Haksoo Lim, Minjung Kim, Sewon Park, and Noseong Park. Regular time-series generation using sgm. p. 0 , 2026), time-series forecasting
(Vuleti´c et al. , 2024), and implied-volatility surface generation (Vuleti´c and Cont, 2024). acero}@jpmorgan. com
1
arXiv:2609.Where the paper says it
Where the paper says it
Haksoo Lim, Jaehoon Lee, Sewon Park, Minjung Kim, and Noseong Park. Tsgm: Regular and irregular time-series generation using scorebased generative models. p. 0 1 Introduction
Generative models are widely used in finance to produce synthetic scenarios for market-risk evaluation (Flaig and Junike, 2022), tail-risk-event simulation (Cont et al. , 2026), time-series forecasting
(Vuleti´c et al.Where the paper says it
Yannick Limmer and Blanka Horvath. Robust hedging gans: Towards automated robustification of hedging strategies. p. 3 , 2019), related
reinforcement learning methods (Cao et al. , 2023), and robust deep-hedging frameworks (Limmer
and Horvath, 2024; L¨utkebohmert et al.Where the paper says it
Yuansan Liu, Sudanthi Wijewickrema, Ang Li, Christofer Bester, Stephen O’Leary, and James Bailey. Time-transformer: Integrating local and global features for better time series generation. p. 0 , 2026), time-series forecasting
(Vuleti´c et al. , 2024), and implied-volatility surface generation (Vuleti´c and Cont, 2024).Where the paper says it
Hang Lou, Siran Li, and Hao Ni. Pcf-gan: generating sequential data via the characteristic function of measures on the path space. In Advances in Neural Information Processing Systems, pages 39755–39781, 2023
Eva L¨utkebohmert, Thorsten Schmidt, and Julian Sester. Robust deep hedging. p. 3 , 2019), related
reinforcement learning methods (Cao et al. , 2023), and robust deep-hedging frameworks (Limmer
and Horvath, 2024; L¨utkebohmert et al.Where the paper says it
Oskari Mikkil¨a and Juho Kanniainen. Empirical deep hedging. p. 1 In our empirical
study of SPX options, AD-Seq-Vol already produces fewer and smaller static-arbitrage violations
than the training data and the GAN-based benchmark, while AD-Seq-Vol-FT reduces these
violations to nearly zero. We evaluate the economic usefulness of the generated conditional scenarios through datadriven hedging.Where the paper says it
Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. p. 0 , 2024), and implied-volatility surface generation (Vuleti´c and Cont, 2024). Most prior
applications have relied on generative adversarial networks (GANs) and variational autoencoders
(VAEs) (Wiese et al. acero}@jpmorgan. com
1
arXiv:2609.Where the paper says it
Where the paper says it
Ilan Naiman, Nimrod Berman, Itai Pemper, Idan Arbiv, Gal Fadlon, and Omri Azencot. Utilizing image transforms and diffusion models for generative modeling of short and long time series. p. 0 1 Introduction
Generative models are widely used in finance to produce synthetic scenarios for market-risk evaluation (Flaig and Junike, 2022), tail-risk-event simulation (Cont et al. , 2026), time-series forecasting
(Vuleti´c et al.Where the paper says it
Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. p. 8 Optimization uses
AdamW for 1000 epochs, batch size 256, initial learning rate 7 × 10−5
, weight decay 10−2
, and a
warm-up cosine learning-rate schedule. For sampling, we use the classic denoising diffusion probabilistic model (DDPM) sampler (Ho et al.Where the paper says it
Brian (Xin) Ning, Sebastian Jaimungal, Xiaorong Zhang, and Maxime Bergeron. Arbitrage-free implied volatility surface generation with variational autoencoders. p. 0 , 2024), and implied-volatility surface generation (Vuleti´c and Cont, 2024). Most prior
applications have relied on generative adversarial networks (GANs) and variational autoencoders
(VAEs) (Wiese et al.Where the paper says it
Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland Vollgraf. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. p. 2 Diffusion models for sequential data. Score-based diffusion models have been adapted to time-series
generation and forecasting (Rasul et al.Where the paper says it
Timur Sattarov, Marco Schreyer, and Damian Borth. Findiff: Diffusion models for financial tabular data generation. p. 2 , 2024; Yang et al. , 2026) and to tabular and sequential financial data (Sattarov et al.Where the paper says it
Martin Schweizer. A guided tour through quadratic hedging approaches. p. 1 Hedging provides a natural decision-based evaluation of the resulting generator, since hedge
selection depends on the conditional joint evolution of the underlying asset and the volatility surface,
rather than on the realism of individual surfaces in isolation. Classical model-based approaches rely
on pre-specified asset-price dynamics and construct hedge positions using local sensitivities such as
delta and vega (Black and Scholes, 1973; Hull, 2022). We study whether diffusion models can capture the cross-sectional and temporal structure of
financial data, focusing on implied-volatility surfaces, a central input to derivative pricing, hedging
and risk management (Gatheral, 2011). An implied-volatility surface summarizes option prices
across strikes and maturities .Where the paper says it
Where the paper says it
Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. p. 8 Optimization uses
AdamW for 1000 epochs, batch size 256, initial learning rate 7 × 10−5
, weight decay 10−2
, and a
warm-up cosine learning-rate schedule. For sampling, we use the classic denoising diffusion probabilistic model (DDPM) sampler (Ho et al.Where the paper says it
Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. p. 1 Their score-matching objectives avoid adversarial optimization and have
demonstrated strong performance on modeling high-dimensional, multimodal distributions. We study whether diffusion models can capture the cross-sectional and temporal structure of
financial data, focusing on implied-volatility surfaces, a central input to derivative pricing, hedging
and risk management (Gatheral, 2011).Where the paper says it
Pascal Vincent. A connection between score matching and denoising autoencoders. p. 5 (2026), whose original implementation is one-dimensional, and adapt it to dynamic implied-volatility surface generation. This yields two variants: AD-Seq-Vol, pretrained by denoising score matching to learn the
conditional dynamics of the market, and AD-Seq-Vol-FT, obtained by post-training AD-SeqVol with penalties for static-arbitrage violations.Where the paper says it
Milena Vuleti´c and Rama Cont. Volgan: A generative model for arbitrage-free implied volatility surfaces. p. 3 2 Diffusion Models for Dynamic Implied-Volatility Surface Generation
This section develops a diffusion-model framework for generating dynamic implied-volatility surfaces. We first review the structure of implied-volatility surfaces and the static no-arbitrage constraints imposed on each surface.Where the paper says it
Milena Vuleti´c, Felix Prenzel, and Mihai Cucuringu. Fin-gan: forecasting and classifying financial time series via generative adversarial networks. p. 0 1 Introduction
Generative models are widely used in finance to produce synthetic scenarios for market-risk evaluation (Flaig and Junike, 2022), tail-risk-event simulation (Cont et al. , 2026), time-series forecasting
(Vuleti´c et al.Where the paper says it
Gefei Wang, Yuling Jiao, Qian Xu, Yang Wang, and Can Yang. Deep generative learning via Schr¨odinger bridge. p. 2 , 2023; Cont et al. , 2026); Schr¨odinger-bridge methods provide a related stochasticcontrol formulation (Wang et al.Where the paper says it
Magnus Wiese, Lianjun Bai, Ben Wood, and Hans Buehler. Deep hedging: Learning to simulate equity option markets. p. 2 Deep-learning-based implied-volatility and option simulation. For equity-option markets, Wiese et al. acero}@jpmorgan. com
1
arXiv:2609.Where the paper says it
Where the paper says it
Magnus Wiese, Robert Knobloch, Ralf Korn, and Peter Kretschmer. Quant gans: deep generation of financial time series. p. 0 , 2026), time-series forecasting
(Vuleti´c et al. , 2024), and implied-volatility surface generation (Vuleti´c and Cont, 2024).Where the paper says it
Yiyuan Yang, Ming Jin, Haomin Wen, Chaoli Zhang, Yuxuan Liang, Lintao Ma, Yi Wang, Chenghao Liu, Bin Yang, Zenglin Xu, Shirui Pan, and Qingsong Wen. A survey on diffusion models for time series and spatio-temporal data. p. 2 Diffusion models for sequential data. Score-based diffusion models have been adapted to time-series
generation and forecasting (Rasul et al.Where the paper says it
Jinsung Yoon, Daniel Jarrett, and Mihaela van der Schaar. Time-series generative adversarial networks. p. 0 , 2024), and implied-volatility surface generation (Vuleti´c and Cont, 2024). Most prior
applications have relied on generative adversarial networks (GANs) and variational autoencoders
(VAEs) (Wiese et al.Where the paper says it
Xinyu Yuan and Yan Qiao. Diffusion-ts: Interpretable diffusion for general time series generation. p. 2 Diffusion models for sequential data. Score-based diffusion models have been adapted to time-series
generation and forecasting (Rasul et al.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive sequential diffusion models can generate coherent, history-dependent surface trajectories by sequentially updating realized market history, capturing both cross-sectional and temporal dependence of implied volatility surfaces.
-
Arbitrage aware fine tuning via AD-Seq-Vol-FT reduces static arbitrage violations to nearly zero by penalizing calendar spread, call spread, and butterfly spread constraints.
-
The resulting diffusion based hedges maintain tracking errors near zero and substantially reduce tail risk relative to classical sensitivity based hedges, exhibiting particularly stable performance during market disruptions like COVID-19.
-
The framework can be integrated into an optimization based hedging procedure where the goal is to solve a one period conditional hedging problem with a rebalancing cost penalty.
-
The system can generate multi day trajectories recursively, using each newly generated surface to update the history that conditions the next day’s surface, unlike one step generators like VolGAN.
Abstract
We develop a diffusion-model framework for dynamic implied-volatility surface generation and evaluate its economic usefulness through data-driven hedging. The framework consists of two models. AD-Seq-Vol jointly learns the conditional evolution of the underlying asset return and the high-dimensional implied-volatility surface, generating adapted multi-period scenarios by sequentially updating the realized market history. AD-Seq-Vol-FT further incorporates option-market structure through post-training penalties for violations of static no-arbitrage conditions. Using daily SPX option data from 2000 to 2023, we show that the proposed models generate coherent surface trajectories while capturing both cross-sectional and temporal dependence. AD-Seq-Vol produces fewer and less severe static-arbitrage violations than the training data and the GAN-based benchmark, while AD-Seq-Vol-FT reduces these violations to nearly zero. We then integrate the generated conditional scenarios into an optimization-based hedging framework. Compared to a range of classical and data-driven benchmarks, the diffusion-based hedges maintain tracking errors near zero, substantially reduce tail risk, and exhibit particularly stable performance during the COVID-19 market disruption. These results establish Adaptive Sequential Diffusion Models as a promising class of market-consistent and economically useful financial scenario generators. Our code is available at: https://github.com/yinbinhan/volatility-surface-simulation.
Sources
- SANOS Smooth strictly Arbitrage-free Non-parametric Option Surfaces
- DYSANOS Generative Dynamic Smooth Arbitrage-free Non-parametric Option Surfaces
- Diffusion Models for Adaptive Sequential Data Generation
- TimeVAE: A Variational Auto-Encoder for Multivariate Time Series Generation
- Real-valued (Medical) Time Series Generation with Recurrent Conditional GANs
- Enhancing Deep Hedging of Options with Implied Volatility Surface Feedback Information
- Generative modeling for time series via Schr{\"o}dinger bridge
- Regular Time-series Generation using SGM
- Conditional Generative Adversarial Nets
- Deep Hedging: Learning to Simulate Equity Option Markets
- Multi-Asset Spot and Option Market Simulation
Related papers
- AFNO: Universal Certifiably-Arbitrage-Free Operator Learning Framework for European Call-Price Surfaces
- Rethinking Synthetic Scenario Realism: Compatibility, Not Fidelity, Drives Hedging Performance
- CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining
- Diversification Across Predictors: Heterogeneous Ensemble Learning for Cross-Sectional Return Prediction
- DSINC beyond Heston: the 3/2 and other non-affine stochastic-volatility models
- Generative AI, Managerial Expectations, and Economic Activity