AFNO: Universal Certifiably-Arbitrage-Free Operator Learning Framework for European Call-Price Surfaces
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: "AFNO: Universal Certifiably-Arbitrage-Free Operator Learning Framework for European Call-Price Surfaces".
Warren: Here's the gist of the paper. Neural operators (NO) are deep learning solutions intended to resolve computational bottlenecks in solving families of problems, but standard NO architectures are insufficient for predicting European call-price surfaces because they cannot guarantee static no-arbitrage conditions and can lead to negative PnL in trading applications.
Ruth: The paper introduces an arbitrage-free (AF) and universal encoder-decoder NO architectural framework.
Warren: This framework continuously maps inputs from a Reproducing Kernel Hilbert Space (RKHS) H over any fixed subset of
0, infinity) to a convex combination of arbitrage-free European call-price surfaces by continuously resolving an approximate nearest neighbor problem. [Ruth: The key contributions establish the certifiability and universality of this approach.
Paper summary: Warren: So, looking at the whole "AFNO: Universal Certifiably-Arbitrage-Free Operator Learning Framework for European Call-Price Surfaces," what does this actually mean for people working in finance or quantitative trading?
Ruth: It means they've built a system where you can use deep learning to generate price surfaces that are guaranteed to obey the no-arbitrage rules, which is a huge deal because standard methods fail that test.
Iris: The universality claim is key here, suggesting this framework isn't just good for one specific set of inputs, but it works for any regular map you feed it.
Warren: If this holds up across the board, it means we can use these neural operators to create pricing models that are much more reliable for risk calculations because we have a certificate of admissibility built into the structure itself.
Ruth: The authors are claiming this architectural enforcement is necessary precisely because unconstrained models can produce off-grid violations, and this AFNO retains its certificate even on unseen grids.
Iris: The authors also point out that this approach addresses a gap in approximation theory where existing theorems for neural operators don't guarantee membership in these specific nonlinear, nonclosed subsets of spaces like the space of arbitrage-free call surfaces.
Warren: So, the implication is that we can move from just getting an accurate curve to getting a curve that is mathematically sound for pricing and trading right out of the box.
Ruth: It's less about finding a magic formula and more about building a framework where the economic constraints are baked into how the network produces its output.
Conclusion: Ruth: So we're wrapping up on AFNO, and what that title really means is they’ve created a deep learning tool for pricing options that is guaranteed to follow no-arbitrage rules.
Warren: It sounds like a big claim, Ruth. They’re saying this operator can predict prices while keeping the math sound across all possible inputs.
Iris: The authors are building an arbitrage-free neural operator framework that works on European call price surfaces and they've shown it’s universal, meaning it can approximate any regular map into those surfaces with arbitrary accuracy.
Ruth: That universality is what makes me think this could be a big deal for risk management. If we can use this to build models that are mathematically sound from the start, that simplifies things a lot for traders who have to worry about bad PnL later on.
Warren: From my side, I’m thinking about the cost of running something like this. If it requires this complex encoder-decoder structure with all those layers, how much computational power are we talking about just to get that certificate?
Iris: The framework is structured in three phases—a sample-based encoder on a kernel space, a transformer backbone, and then arbitrage-free decoders—which suggests the complexity comes from enforcing that constraint during the decoding step.
Ruth: It’s all about how they managed to encode no-arbitrage without having to manually program every single constraint into the architecture itself. That’s what they call a certificate of admissibility baked into the range of the map.
Warren: So, it’s not just predicting a price curve; it’s predicting one that is economically admissible by construction, which addresses those negative results where standard neural operators produce things that break no-arbitrage conditions.
Iris: They prove this works by showing that every parameter setting of the AFNO produces a surface with controlled right-wing decay, and they can approximate any smooth function into the arbitrage-free space to any desired precision.
Ruth: It shifts the focus from just "how accurate is this prediction?" to "is this prediction valid in a real market?" which is a much harder question for anyone in finance.
Warren: The implication here is that we might start seeing models where the primary concern isn't just minimizing error, but ensuring the model stays within the bounds of what’s possible under risk-neutral pricing theory.
Iris: This framework is compatible with risk-neutral pricing uniformly across all parameter settings, meaning you can query it at any maturity or strike without losing that critical static no-arbitrage condition for consistent calculations.
Ruth: So, they're offering a way to get the flexibility of deep learning prediction with the mathematical rigor required for actual financial products.
Warren: It’s definitely an architectural approach that encodes the economic constraints into its very structure so it provides a certificate while still being flexible enough to make predictions.
Iris: And they’ve shown that this structural enforcement is necessary because unconstrained models can produce violations on unseen grids, meaning this AFNO retains its certificate even when you ask it for something new.
Ruth: This moves the needle from just getting a good fit on training data to having a model that is fundamentally sound for trading and pricing.
Warren: It opens up a new way to think about using neural operators in finance, provided we can manage the computational overhead of this complex structure.
Iris: So next time you hear about neural operators, keep an eye out for these types of frameworks that focus on certifying the economic properties of their outputs.
Anastasis Kratsios, Yannick Limmer, Hans Buehler, Blanka Horvath, Raeid Saqur
Mathematical Institute, University of Oxford · Oxford-Man Institute for Quantitative Finance, University of Oxford
q-fin.CP, cs.AI, cs.CE
Submitted: 2026-10-13
Updated: 2026-10-13
Comments: 63 pages, 9 figures, 7 tables
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: Neural operators (NO) are deep learning solutions intended to resolve computational bottlenecks in solving families of problems…
Key concepts
- Neural Operators (NO)
- Deep learning solutions designed to solve complex families of problems. Standard NOs are insufficient for pricing because they cannot guarantee static no-arbitrage conditions, which is critical for financial applications.
- Arbitrage-Free (AF) Framework
- A specific architectural approach where the model is designed to continuously map inputs into a convex combination of surfaces that are guaranteed to be arbitrage-free. This ensures that every output price surface is economically admissible.
- Universal Encoder-Decoder
- The AFNO structure involves an encoder, a backbone for calculating weights via softmax, and a decoder. This setup allows the model to approximate any sufficiently regular map from input data into the space of arbitrage-free call surfaces with arbitrary precision.
Terminology
Summary
Neural operators (NO) are deep learning solutions intended to resolve computational bottlenecks in solving families of problems, but standard NO architectures are insufficient for predicting European call-price surfaces because they cannot guarantee static no-arbitrage conditions and can lead to negative PnL in trading applications. p. 0 AFNO: Universal Certifiably-Arbitrage-Free Operator Learning
Framework for European Call-Price Surfaces
Anastasis Kratsios3,4∗ Yannick Limmer5 Hans Buehler1
Blanka Horvath1,2 Raeid Saqur1,4
1Mathematical Institute, University of Oxford
2Oxford-Man Institute for Quantitative Finance, University of Oxford
3Department of Mathematics, McMaster University
4Vector Institute
5DRW
Abstract
Neural operators (NO) have been introduces as deep learning-powered solutions to the computational bottleneck of resolving families of problems; since, once trained an NO can rapidly
output approximate solutions, thus avoiding the need to re-run classical solvers. While standard
NO architectures resolve this computation trouble they are inadequate for predicting European
call-price surfaces from market or model features cannot guarantee static no-arbitrage conditions
and, as we verify experimentally, can yield negative PnL in downstream trading tasks. While standard
NO architectures resolve this computation trouble they are inadequate for predicting European
call-price surfaces from market or model features cannot guarantee static no-arbitrage conditions
and, as we verify experimentally, can yield negative PnL in downstream trading tasks. We introduce a certifiably arbitrage-free (AF) and universal encoder-decoder NO architectural
framework—continuously mapping inputs from a Reproducing Kernel Hilbert Space (RKHS) H over
any fixed subset of [0,∞)
2
to a convex combination of arbitrage-free European call-price surfaces;
by continuously resolving an approximate nearest neighbour problem. While standard
NO architectures resolve this computation trouble they are inadequate for predicting European
call-price surfaces from market or model features cannot guarantee static no-arbitrage conditions
and, as we verify experimentally, can yield negative PnL in downstream trading tasks. We introduce a certifiably arbitrage-free (AF) and universal encoder-decoder NO architectural
framework—continuously mapping inputs from a Reproducing Kernel Hilbert Space (RKHS) H over
any fixed subset of [0,∞)
2
to a convex combination of arbitrage-free European call-price surfaces;
by continuously resolving an approximate nearest neighbour problem.Where the paper says it
Where the paper says it
Where the paper says it
The key contributions establish the certifiability and universality of this approach. p. 2 These results distinguish empirical constraint enforcement from architectural certification. Penalties may reduce arbitrage violations, and post-processing may repair
particular outputs, but neither approach, by itself, guarantees that the resulting operator maps
every input into the space of arbitrage-free call surfaces. 1 Primary Contribution
On the positive side, we prove that every parameter configuration of an AFNO produces an arbitragefree European call-price surface with controlled right-wing decay (Theorem 4. Moreover, AFNOs
are universal: every Lipschitz map from R
d
into the corresponding class of arbitrage-free call
2 1 Primary Contribution
On the positive side, we prove that every parameter configuration of an AFNO produces an arbitragefree European call-price surface with controlled right-wing decay (Theorem 4. Moreover, AFNOs
are universal: every Lipschitz map from R
d
into the corresponding class of arbitrage-free call
2Where the paper says it
Where the paper says it
Where the paper says it
The framework is structured in three phases:
(Phase 1) Sample-Based Encoder on RKHS:
(Phase 2) Transformer Backbone:
(Phase 3) Arbitrage-Free Decoders:
The AFNO architecture is defined as an encoder that extracts finitely many real-valued input features, a backbone that computes simplex weights via softmax, and a decoder that combines a separate dictionary of arbitrage-free output surfaces using these weights. This structural approach is motivated by negative results showing that predicted surfaces need not admit a nearest arbitrage-free approximation (Theorem 5. p. 2 avoiding re-running a linear program as with [BHK+26] or
calibrating an SDE-type model, but they simply can simply evaluate a forward pass of our AFNO
model. 2 Secondary Contributions
On the negative side, a predicted surface need not admit a nearest arbitrage-free approximation
(Theorem 5. 1); thus, no-arbitrage cannot, in general, be restored by post-processing. Moreover,
standard encoder-decoder neural operators and Fourier neural operators admit parameter configurations whose outputs violate static no-arbitrage (Propositions 5. This motivates enforcing no-arbitrage
directly through the architecture. Practically, our experiments demonstrate the consequences
of forgoing such a certificate: a downstream learning agent may focus on spurious arbitrage
opportunities created by defects in the predicted call-price surfaces rather than learn genuinely
profitable trading strategies.Where the paper says it
Where the paper says it
Where the paper says it
The framework is compatible with risk-neutral pricing theory uniformly across all parameter specifications and remains expressive enough to approximate any sufficiently regular map from an RKHS into the space of arbitrage-free European call-price surfaces. p. 0 While standard
NO architectures resolve this computation trouble they are inadequate for predicting European
call-price surfaces from market or model features cannot guarantee static no-arbitrage conditions
and, as we verify experimentally, can yield negative PnL in downstream trading tasks. We introduce a certifiably arbitrage-free (AF) and universal encoder-decoder NO architectural
framework—continuously mapping inputs from a Reproducing Kernel Hilbert Space (RKHS) H over
any fixed subset of [0,∞)
2
to a convex combination of arbitrage-free European call-price surfaces;
by continuously resolving an approximate nearest neighbour problem. AFNO: Universal Certifiably-Arbitrage-Free Operator Learning
Framework for European Call-Price Surfaces
Anastasis Kratsios3,4∗ Yannick Limmer5 Hans Buehler1
Blanka Horvath1,2 Raeid Saqur1,4
1Mathematical Institute, University of Oxford
2Oxford-Man Institute for Quantitative Finance, University of Oxford
3Department of Mathematics, McMaster University
4Vector Institute
5DRW
Abstract
Neural operators (NO) have been introduces as deep learning-powered solutions to the computational bottleneck of resolving families of problems; since, once trained an NO can rapidly
output approximate solutions, thus avoiding the need to re-run classical solvers. While standard
NO architectures resolve this computation trouble they are inadequate for predicting European
call-price surfaces from market or model features cannot guarantee static no-arbitrage conditions
and, as we verify experimentally, can yield negative PnL in downstream trading tasks. We show that a nearest
arbitrage-free correction of an unconstrained predicted surface need not exist, implying that noarbitrage cannot, in general, be restored through post-processing. Further, we show that standard
NOs such as the FNO and certain encoder-decoder NOs admit parameter configurations whose
outputs violate static no-arbitrage.Where the paper says it
Where the paper says it
Where the paper says it
The main approximation result is Proposition C. p. 6 1] and first part of a QAS space [AKP24,
Definition 3. 4], required for that paper’s main “static” approximation theorem in non-linear spaces. = CAF ∩ X p
λ
is called uniformly controlled if there exists a constant B > 0 and a modulus
of continuity ω such that
3By contrast, it is closed in the (global) uniform norm. Indeed, if Cn ∈ CAF and ∥Cn − C∥∞ → 0, then continuity,
non-negativity, the boundary values, strike convexity, maturity monotonicity, and the strike Lipschitz bound pass to C. While standard
NO architectures resolve this computation trouble they are inadequate for predicting European
call-price surfaces from market or model features cannot guarantee static no-arbitrage conditions
and, as we verify experimentally, can yield negative PnL in downstream trading tasks. We introduce a certifiably arbitrage-free (AF) and universal encoder-decoder NO architectural
framework—continuously mapping inputs from a Reproducing Kernel Hilbert Space (RKHS) H over
any fixed subset of [0,∞)
2
to a convex combination of arbitrage-free European call-price surfaces;
by continuously resolving an approximate nearest neighbour problem.Where the paper says it
Where the paper says it
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
1
Improve AI systems to predict European call-price surfaces by implementing an Arbitrage-Free Neural Operator (AFNO) framework that structurally enforces no-arbitrage conditions at every parameter configuration. This allows downstream trading agents to focus on genuinely profitable strategies rather than spurious arbitrage opportunities created by defects in the predicted surface.
2
Implement an encoder-processor-decoder architecture where the decoder maps directly into the nonlinear space of arbitrage-free European call-price surfaces, as described in Definition 3.5 and Proposition 4.1. This ensures that every output from the model is certified to be arbitrage-free by construction, satisfying Definition 3.3(ii).
3
Enable universal approximation capabilities by utilizing the AFNO framework to approximate any Lipschitz map from an input Reproducing Kernel Hilbert Space H into the space of arbitrage-free call surfaces CAF. This means the system can approximate parameter-to-call surface maps generated by broad parametric families of risk-neutral martingale models (Corollary 4.5) without violating static no-arbitrage conditions.
4
Develop a learning mechanism for input features that adapts to the geometry of the input Reproducing Kernel Hilbert Space H using a trainable encoder EK, as detailed in Section 3.1 and Proposition 3.1. This allows the system to efficiently encode market data or model features into a representation suitable for deep neural networks, even when dealing with irregular observation grids.
5
Allow the AI system to learn complex non-parametric mappings between input features and arbitrage-free call surfaces using an AF decoder that combines outputs from multiple arbitrage-free surface-decoders via a convex combination (Definition 3.4). This structure, utilizing attention mechanisms for alignment scores, enables the model to approximate target operators uniformly on compact input sets (Proposition C.10).
6
Create a robust interpolation mechanism for approximating call prices on irregular observation grids using structured ReLU-MLP decoders (Definition 3.7) and the Universal Arbitrage-Free Interpolation theorem (Proposition 3.8). This allows the system to exactly interpolate any arbitrage-free call surface on a prescribed grid while remaining arbitrage-free globally, even when dealing with non-uniform strike and maturity data.
7
Ensure that predicted surfaces maintain uniform control over their right-wing decay rate by utilizing smooth Black-Scholes based decoders (Proposition 3.11) which exhibit no static arbitrage, as shown in Proposition 3.11 and the proof of Theorem 4.2 for uniformly controlled call surface classes (Definition 2.9).
Abstract
Neural operators (NOs) have been introduced as deep learning-powered solutions to the computational bottleneck of solving families of problems. Once trained, an NO can rapidly output approximate solutions, avoiding the need to rerun classical solvers. However, standard NO architectures cannot guarantee static no-arbitrage when predicting European call-price surfaces from market or model features and, as we verify experimentally, can yield negative PnL in downstream trading tasks. We introduce a certifiably arbitrage-free (AF) and universal encoder-decoder NO framework, continuously mapping inputs from a reproducing kernel Hilbert space (RKHS) H over any fixed subset of [0, infinity) squared to a convex combination of arbitrage-free European call-price surfaces by continuously solving an approximate nearest-neighbour problem. We show that a nearest arbitrage-free correction of an unconstrained predicted surface need not exist, implying that no-arbitrage cannot, in general, be restored through post-processing. Further, we show that standard NOs, such as the FNO and certain encoder-decoder NOs, admit parameter configurations whose outputs violate static no-arbitrage. Thus, out-of-the-box NOs generally cannot be certifiably restricted to arbitrage-free configurations, but may only be guided towards them through PINN-type regularization during training.
Sources
- Deep Neural Operator Learning for Probabilistic Models
- SANOS Smooth strictly Arbitrage-free Non-parametric Option Surfaces
- Moments in Rough Bergomi and Boundary Attainment in Rough Heston
- Polynomial Scaling is Possible For Neural Operator Approximations of Structured Families of BSDEs
- One model to solve them all: 2BSDE families via neural operators
- Simultaneously Solving Infinitely Many LQ Mean Field Games In Hilbert Spaces: The Power of Neural Operators
- Arbitrage-free SVI volatility surfaces
- Designing Universal Causal Deep Learning Models: The Case of Infinite-Dimensional Dynamical Systems from Stochastic Analysis
- Consistent Recalibration Models and Deep Calibration
- When is Fine-Tuning Scalable? Certified Recovery from Noisy Data and Linear-Cost Fine-Tuning on Structured Task Shifts
- Is In-Context Universality Enough? MLPs are Also Universal In-Context
- An Approximation Theory for Metric Space-Valued Functions With A View Towards Deep Learning
- Adaptivity Under Realizability Constraints: Comparing In-Context and Agentic Learning
- Neural Operators Can Discover Functional Clusters
- The Parametric Complexity of Operator Learning
- Solving Optimal Execution Problems via In-Context Operator Networks
- Mathematical theory of deep learning
- Weighted universal approximation of differentiable maps on infinite-dimensional manifolds
Related papers
- Diffusion models for dynamic volatility surface generation and data-driven hedging
- 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