acceptodds
Under review as a conference paper at ICLR 2027

Learned Functional Projection: Learning for Recovery in Implied Volatility Smoothing

25 views

Abstract

Implied volatility smoothing requires recovering continuous surfaces from sparse, irregular market quotes, yet approximation quality alone does not determine recovery accuracy. We prove that representations with zero approximation error can have arbitrarily large recovery-risk ratios under the same observations and coefficient-fitting rule. This motivates selecting representations by their recovery risk under the available observations. We introduce Learned Functional Projection (LFP), which learns a low-dimensional continuous function space by optimizing recovery from partial quotes. Training differentiates through a weighted ridge solve and evaluates predictions on held-out quotes; each new surface requires only one explicit coefficient solve. Financial regularization penalizes shape violations. The explicit recovery rule admits exact error identities that separate approximation, regularization and input error, together with a deterministic pointwise upper bound and a representation-error lower bound on recovery risk. Experiments use 2012–2021 data from four equity-index option markets. On SPX under standard inputs, LFP reduces held-out vega-weighted relative RMSE by 61.8% against the graph neural operator (GNO). With the same nine reference quotes, LFP-9 reduces this error by 58.4% against HyperIV and achieves 3.9 times its bid–ask coverage. The SPX-trained representation also leads all evaluated baselines in quote accuracy on NDX and RUT without retraining. Surface construction takes milliseconds, with approximately CPU and GPU speedups over the public GNO implementation. Controlled ablations show that learning for recovery outperforms learning for reconstruction, supporting the method's design and the practical relevance of the theoretical distinction.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Implied volatility surface, financial engineering, functional representation learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kSLFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Learned Functional Projection (LFP) to recover implied volatility surfaces from sparse, irregular market quotes (function regression problem). The method builds on Function Encoders and differentiable ridge solvers and a shared set of continuous basis functions and uses weighted ridge regression to fit each new surface. The empirical performance is strong on real market data from four equity-index option markets, while surface construction takes only milliseconds.

Critical strengths

None

Critical weaknesses

W1. The methodological contribution seems relatively incremental. It looks more like a direct combination of Function Encoders and ridge-based meta-learning. I acknowledge the contribution of the theoretical analysis, but it remains unclear to me what substantial methodological advance LFP provides beyond adapting these ideas to volatility smoothing.

W2. The baseline comparison is not fully convincing. The comparison with Neural Process methods is missing. Moreover, I have some questions about the GNO setting; see the questions section below.

Questions for authors

Q1 SPX, NDX, RUT are used in the abstract but not defined. These abbreviations should be explained at first use .

Q2. From my understanding, GNO relies on kernel integral to approximate a continuous operator, and the radius of GNO usually has physical meaning. Is GNO appropriately configured for this finance task? I hope the authors to explain its coordinate choice and neighborhood radius. The aggregation of the GNO uses at most 35 neighbors node, in this case, why not use GNN?

Q3. For the experiment part, GNO uses released SPX weights without retraining on the daily dataset, whereas LFP and HyperIV are trained on that dataset. Another minor question for the authors is why does the abstract emphasize the relative improvement over GNO, which is not the strongest baseline as shown in Table 4.1

Q4. I think the task looks closely related to Neural Process-style conditional function regression, so a comparison with state-of-the-art Neural Process baselines would also be informative. (eg. Neural Diffusion/Flow Matching Process)

3 comments

Thank you for your detailed feedback. We have added comparisons with FlowNP, GNO-daily, and VAE, rewritten the Introduction, and added Figure 1.1 and Algorithm 1. The revised PDF is available for download at the top of this page. All revisions are highlighted in blue for easy reference. We address your comments below.

W1: Methodological contribution.

LFP’s central contribution is to select a continuous function space according to its recovery performance under limited observations. We embed coefficient recovery into space learning: coefficients are fitted from input quotes, and prediction error on query quotes jointly optimizes the basis functions and ridge parameter. The learned representation therefore accounts for both surface expressivity and coefficient recoverability.

Our theory shows that even when two spaces represent a target surface exactly, their recovery-risk ratio can be arbitrarily large. Thus, the best approximation space need not be the best recovery space. Controlled experiments with the same network architecture and space dimension further validate this design: recovery-oriented training reduces RW by 29.9% and 11.2% relative to the two alternative training objectives. Our contribution is to establish and validate this recovery-oriented space-learning principle and realize it through a simple shared-basis design that supports financial shape control and fast surface updates. The revised Introduction and Figure 1.1 clarify this motivation and its connection to the method.

W2 and Q4: Comparison with Neural Process methods.

We have added FlowNP, a flow-matching Neural Process method of the type you suggested. The following results are for SPX. RW is reported as a decimal, violation rates as percentages, and latency as median end-to-end inference time.

MetricFlowNPLFP / LFP-9
Standard-input RW ↓0.01220.0055
Nine-quote RW ↓0.08300.0444
Standard-input butterfly violation rate ↓22.55%0.023%
Standard-input CPU latency ↓4870.61 ms4.79 ms
Standard-input GPU latency ↓69.11 ms2.79 ms

FlowNP achieves competitive quote prediction, but substantial butterfly violations remain even after incorporating and tuning financial regularization during training. A possible explanation is that butterfly conditions involve second-order spatial derivatives: small pointwise prediction errors do not ensure financially admissible local curvature. For FlowNP, shape regularization must act through the multistep sampling process to constrain the final surface. LFP’s explicit continuous basis representation facilitates direct computation and regularization of surface derivatives.

Furthermore, even with a reduced sampling budget of three samples and eight steps, FlowNP’s GPU inference is approximately 25 times slower than LFP’s, and its CPU inference approximately 1,000 times slower. This computational cost poses a practical challenge when surfaces must be updated frequently as market quotes change, particularly across multiple assets. Full results on accuracy, financial shape, cross-asset transfer, and computational cost have been added to the main paper and appendices.

Q2: GNO coordinates, neighborhoods, and its relationship to GNNs.

We use the GNO developed by Wiedemann et al. specifically for implied volatility smoothing. Its coordinates are r=τr=\sqrt{\tau} and z=log⁡(K/F)/τz=\log(K/F)/\sqrt{\tau}, where τ\tau is time to maturity, KK is the strike price, and FF is the forward price. Following the publicly released code configuration, the neighborhood radius is 0.30.3, applied only along the rr-axis to control information exchange between maturity slices. At most 35 neighbors per node are retained for the discrete approximation of the kernel integral.

An important motivation for using GNO is its discretization-invariant operator formulation. Option quotes are highly irregular across strikes and maturities, and their locations and number change over time. The model must therefore accommodate different input discretizations and predict surfaces at query coordinates. GNO addresses this through shared coordinate-dependent kernels. It is itself implemented through graph message passing; the 35-neighbor cap is a computational budget and does not change its continuous-operator modeling objective. The rationale for using GNO is therefore its suitability for changing quote sets.

[Continued in the next comment]

[Continuing previous comment]

Q3: GNO training data and the abstract comparison.

We now report both GNO-20min and GNO-daily. The former uses the officially released SPX weights; the latter is retrained using the same daily SPX training split as the other learned models, directly addressing the comparison under a common daily-data setting.

The original abstract emphasized GNO and HyperIV because they were recent advanced neural methods for implied volatility smoothing, and we aimed to demonstrate the advantages of learning low-dimensional function spaces relative to neural operators and hypernetworks. The revised abstract reports improvements over the strongest error baseline under each protocol: a 39.3% reduction in RW relative to PCA ridge under standard inputs and a 46.5% reduction relative to FlowNP under nine-quote inputs. It also summarizes cross-asset performance, financial regularization, and computational efficiency.

Q1: Abbreviations in the abstract.

We have provided the full index names corresponding to SPX, NDX, RUT, and DJX at their first occurrence.

Dear Reviewer,

Thank you for your valuable feedback. We have carefully addressed all the concerns raised in the reviews.

As the discussion deadline is approaching, we would greatly appreciate it if you could take a moment to review our rebuttal responses and revised PDF, and provide any follow-up feedback, particularly on whether their concerns have been adequately addressed. Thank you for your time!

Best regards,

The Authors

Reviewer_UKe7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

For implied volatility smoothing, this work first decompose total recovery error into representation error and coefficient recovery error. Then, it shows there is no universal upper bound on the recovery-risk ratio based only on both representations having zero approximation error. Above theoretical analysis motivates the design of LFP which combines neural function-space learning, explicit coefficient recovery and financial shape regularization. The reported results demonstrate substantial improvements over SSVI, PCA-ridge, GNO and HyperIV, along with millisecond-scale inference and promising cross-asset transfer. A controlled comparison of training objectives also shows that optimizing recovery can outperform optimizing reconstruction, providing empirical support for the theoretical motivation.

Critical strengths
  1. The research question, approximation-versus-recovery distinction, is clearly presented, and the paper sticks to it throughout.
Critical weaknesses
  1. The ablation study of financial penalties is missing. This makes the contribution of financial penalties remain unknown.

  2. The presentation of this work is not clear. The algorithm box or pseudocode of LFP is missing which makes the proposed method hard to follow. The abstract part does not fully summarize the content of this work.

  3. More baselines are needed to show the real performance of the proposed method, e.g. Variational autoencoders: A hands-off approach to volatility.

Questions for authors
  1. Why the abstract part only reports the comparison with GNO baseline? This did not fully conclude the empirical results of this work.

  2. Are “LFP” and “LFP-Arb” the same method?

2 comments

Thank you for recognizing the paper's consistent focus on approximation versus recovery. We have added experiments and revised the manuscript following your suggestions. The revised PDF is available for download at the top of this page. All revisions are highlighted in blue for easy reference.

1. Ablation of financial penalties

We have added experiments removing the arbitrage penalties, the smoothness penalties, and both, with results included in Table 4.3(e) of the main paper. The following results use standard inputs on SPX; all values are percentages:

VariantRW ↓Bfly ↓Cal ↓
Full LFP0.554360.025920.00081
Without arbitrage penalties0.499710.292170.00960
Without smoothness penalties0.570600.023270.05521
Without both0.597545.501160.43700

Removing the arbitrage penalties reduces quote error but raises butterfly and calendar violation rates to approximately 11 and 12 times those of the full model, respectively. Removing smoothness increases both quote error and calendar violations. Removing both further increases shape violations, showing that the two types of penalties jointly support LFP's financial shape control.

2. Additional baselines

We have added the VAE you suggested, and further included GNO-daily and FlowNP. GNO-20min uses the officially released weights; GNO-daily is retrained on the same daily SPX data as the other learned models. The table below reports the main metrics and median end-to-end inference times under standard inputs on SPX. The first three metrics are decimals, and times are in ms:

MethodRW ↓MAPE ↓BA ↑CPU ↓GPU ↓
PCA ridge0.00910.00540.56504.16—
GNO-20min0.01450.00900.3671200.67314.99
HyperIV0.10760.07870.03852.492.27
GNO-daily (added)0.03700.02470.1555196.03240.56
VAE (added)0.14760.09120.0407100.65210.13
FlowNP (added)0.01220.00750.41964870.6169.11
LFP0.00550.00340.74284.792.79

LFP leads on all three quote metrics while maintaining millisecond-scale inference. Results for nine-quote inputs, cross-asset transfer, financial shape, and paired comparisons have also been updated in the main paper and appendices.

3. Readability, pseudocode, and abstract

We have rewritten the Introduction, added Figure 1.1 to illustrate training, inference, and the theoretical intuition, and included Algorithm 1 in the appendix with a reference in the main method section.

The original abstract emphasized GNO and HyperIV because they were recent state-of-the-art methods for implied volatility smoothing at the time, and we aimed to demonstrate the advantages of learning low-dimensional function spaces relative to neural operators and hypernetworks. PCA ridge was included as a linear-representation control motivated by our formulation, to compare historical approximation with recovery-oriented space learning. The revised abstract instead reports improvements over the strongest error baseline under each protocol: a 39.3% reduction in RW relative to PCA ridge under standard inputs and a 46.5% reduction relative to FlowNP under nine-quote inputs. It also summarizes cross-asset performance, financial-regularization ablations, and absolute inference times to provide a more comprehensive account of the empirical results.

4. Naming of LFP and LFP-Arb

Both names refer to the same full method used in our experiments, including financial regularization. We have standardized the name to LFP and removed the term LFP-Arb. LFP-9 denotes the version trained under the nine-quote input protocol.

Dear Reviewer,

Thank you for your valuable feedback. We have carefully addressed all the concerns raised in the reviews.

As the discussion deadline is approaching, we would greatly appreciate it if you could take a moment to review our rebuttal responses and revised PDF, and provide any follow-up feedback, particularly on whether their concerns have been adequately addressed. Thank you for your time!

Best regards,

The Authors

Reviewer_EFBhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces the idea of learned functional projections for implied volatility smoothing. In the domain of finance, volatility smoothing constructs a continuous 3d surface based on available discrete strike prices and maturities. The paper argues and theoretically shows that, unlike in typical scenarios, in volatility smoothing a good approximation to the surface does not necessarily mean good recovery. So, the paper proposes to learn representations by prediction after quote fitting. Technically it uses a neural network with 33 per surface coefficients to represent the function, and it is trained using a ridge penalty. The paper theoretically proves the $\textbf{recovery error} $ guarantees of this formulation. Experiments on real world indices shows the accuracy and computational efficiency of the method in most cases.

Critical strengths
  1. The distinction between recovery and approximation error is important and interesting and would be of interest to community. The formulation shows reasonable empirical performance.
Critical weaknesses
  1. The paper is not so easy to understand for a reader not belonging to the domain. The paper tries to explain basic concepts, but the writing felt to me too rushed, especially in the introduction section.
  2. Many of the ideas follow directly from Wiedemann et al. and the network architecture from Ingebrand et al. In that sense the novelty seems restricted to the distinction between approximation and recovery only.
Questions for authors

More of a suggestion: It would be good to improve readability by restructuring or expanding the introduction section by a bit and also highlight novel contributions and challenges to the formulation better.

5 comments

The authors have not provided answers or clarifications to the reviewers yet. In this case, I will keep my score and inclination towards rejection.

Thank you for carefully reviewing our paper. We also appreciate your comments on readability and the presentation of our novel contributions. We apologize for the delayed response. We have recently completed several additional experiments and rewritten the relevant parts of the manuscript to provide more comprehensive empirical results and clearer explanations. We address your comments below.

1. Readability, organization of the Introduction, and presentation of the method

We have rewritten and reorganized the Introduction to follow the progression from the financial task, through the low-dimensional structural prior and the challenges of recovery from finite quotes, to the design and contributions of LFP. We have added Figure 1.1 to provide an intuitive overview of the paper. We have also added Algorithm 1 to the LFP Implementation Details appendix and referenced it in the main method section. It summarizes coefficient fitting, the training objective, and inference for new surfaces. The Introduction now ends with explicit statements of the methodological, theoretical, and empirical contributions.

2. Relationship to prior work and novelty

The central contribution of this work is to turn recoverability from finite quotes into a criterion for learning function spaces, and to establish through theoretical analysis and controlled experiments why this design matters and how it produces practical gains. The distinction between approximation and recovery therefore directly guides LFP's optimization objective, representation learning, and inference procedure.

We adopt some of the financial modeling and evaluation settings of Wiedemann et al. and build on the learned-basis framework of Ingebrand et al. Whereas Wiedemann et al. use a graph neural operator to map quotes to surfaces, LFP explicitly represents surfaces as linear combinations of shared continuous basis functions, using low-dimensional structure to connect learning from historical data with adaptation to current quotes. The coefficients of each new surface are determined by one weighted ridge solve, and financial shape penalties act on the resulting continuous surface.

Compared with the training objective formulated around Hilbert-space approximation error in Ingebrand et al., LFP explicitly separates the input quotes S used for coefficient recovery from the query quotes Q used to evaluate recovery quality: coefficients are first fitted on S, and prediction error on Q then jointly optimizes the basis functions and ridge parameter. This directs function-space selection toward recovery performance under actual observation conditions, accounting for representational capacity, coefficient recoverability, and the role of regularization. For financial data, where historical surfaces are only partially observed and quote locations continually change, this design incorporates the recovery task encountered at deployment into representation learning.

3. Theoretical and controlled empirical support

Theoretically, we prove that even when two spaces of the same dimension both have zero representation error, the ratio of their recovery risks can be arbitrarily large under the same quotes, weights, and ridge parameter. This shows that optimal approximation is insufficient to determine a representation suitable for recovery from finite quotes. Our further analysis of weighted ridge recovery characterizes how representation residuals, regularization, and input errors jointly affect recovery, providing an analytical basis for jointly learning the basis functions and ridge parameter.

Controlled ablations directly validate the role of the space-learning criterion. With the same architecture and representation dimension, recovery objective C reduces RW by 29.9% and 11.2% relative to full-pool reconstruction objective A and input reconstruction objective B, respectively, with significant improvements in MAPE and bid–ask coverage as well. In particular, A achieves the lowest approximation error but produces the largest recovery error, while C reduces coefficient recovery error by 46.6% relative to A. These results show that changing the representation-learning objective can significantly improve recovery without increasing model capacity, and that the distinction between approximation and recovery identified by the theory also appears in real market data.

LFP's contributions thus span the problem formulation, learning design, theoretical analysis, and empirical validation. Its simple structure brings shared low-dimensional representations, financial shape control, and rapid adaptation to quotes into a single framework. The revised manuscript also includes additional comparisons with GNO-daily, VAE, and FlowNP, together with financial-regularization ablations and computational cost results, further demonstrating the practical value of this design in recovery accuracy, shape control, and millisecond-scale surface updates.

Dear Reviewer,

Thank you for your valuable feedback. We have carefully addressed all the concerns raised in the reviews.

As the discussion deadline is approaching, we would greatly appreciate it if you could take a moment to review our rebuttal responses and revised PDF, and provide any follow-up feedback, particularly on whether their concerns have been adequately addressed. Thank you for your time!

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.