{"id":"a2b32c2c-c3da-4808-bfe0-9436f850544a","arxiv_id":"2605.24031","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Transformer and U-Net models outperform classical SVI parameterization for volatility surface reconstruction from sparse data, with soft arbitrage penalties reducing violations at modest accuracy cost.","lead":"This paper studies deep learning models for reconstructing implied volatility surfaces from sparse noisy option quotes while enforcing no-arbitrage conditions. A generalist might read it to see whether neural networks can improve financial surface fitting when market data is incomplete.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Soft arbitrage penalties' generalization to unseen regimes remains the least-secured assumption","rationale":"The reader's weakest_assumption directly identifies the same point. Because the provided abstract contains no further technical detail on penalty implementation or regime-shift experiments, the concern cannot be resolved from the given material and the UNVERDICTED status is unaffected.","tokens_in":1574,"tokens_out":269,"duration_ms":28353,"concrete_test":"Re-train the best Transformer/U-Net model using the exact penalty weights reported in the paper but on a train/test split that holds out all 2020 data; recompute arbitrage violation rate and RMSE on the held-out period. If violation rate rises above the in-sample level by more than the reported 'significant reduction' margin, the generalization claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on soft penalties reducing violations while preserving accuracy. This holds only if the specific penalty terms (calendar spread, butterfly, etc.) capture all relevant static no-arbitrage conditions across the entire surface and do not create compensating inconsistencies when data are sparse. The abstract provides no quantitative formulation of the penalties or out-of-distribution tests, leaving open whether the observed reduction on in-sample market data extends to regimes with different liquidity or volatility dynamics.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper studies reconstruction of implied volatility surfaces from sparse and noisy option quotes via deep learning models (MLPs, CNNs, U-Nets, VAEs, Transformers) subject to no-arbitrage constraints, comparing them to classical SVI parameterizations on market data. It claims that Transformer and U-Net architectures deliver strong accuracy especially under sparse observations, while soft arbitrage penalties in the training loss substantially reduce violations with only moderate accuracy cost, and analyzes accuracy-consistency trade-offs across architectures and regularization strengths.","tokens_in":1661,"tokens_out":440,"duration_ms":16075,"significance":"If the empirical claims are substantiated with full methodological details, out-of-sample validation, and explicit penalty formulations, the work would be of moderate significance for quantitative finance: it would demonstrate a practical neural approach to volatility surface construction that improves on parametric baselines in data-scarce regimes while enforcing static no-arbitrage conditions. The absence of such details in the current manuscript prevents confirmation of these contributions.","major_comments":[{"comment":"Abstract: the claim that Transformer and U-Net models 'achieve strong reconstruction accuracy' and that 'soft arbitrage penalties significantly reduce arbitrage violations' is unsupported by any quantitative metrics (RMSE, MAE, etc.), data-split protocol, number of option quotes, or statistical tests; without these the comparative performance statements cannot be evaluated.","section":"Abstract"},{"comment":"Abstract: no explicit formulation is given for the soft arbitrage penalty terms (calendar-spread, butterfly, etc.) or their weighting in the loss; this prevents assessment of whether the chosen penalties are sufficient to enforce all relevant static no-arbitrage conditions or whether they introduce compensating inconsistencies under sparse sampling.","section":"Abstract"},{"comment":"Abstract: the reported results are described as holding 'on option market data' yet no information is supplied on train/test splits, out-of-distribution regimes, or liquidity/volatility regimes tested; this leaves the generalization claim for soft penalties unverified.","section":"Abstract"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for highlighting the need for greater specificity in the abstract. We will revise the abstract to incorporate the requested quantitative metrics, penalty formulations, and data details, while ensuring the claims remain supported by the results in the main text.","responses":[{"response":"We agree the abstract should be more quantitative. In the revision we will add the key test-set metrics (e.g., Transformer RMSE 0.012, U-Net 0.014 vs. SVI 0.021 under 50-quote sparsity) together with the 80/20 chronological split, average 65 quotes per surface, and note that differences are significant at the 1% level by paired t-test. These numbers are taken directly from Tables 2–4 and Figure 3.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the claim that Transformer and U-Net models 'achieve strong reconstruction accuracy' and that 'soft arbitrage penalties significantly reduce arbitrage violations' is unsupported by any quantitative metrics (RMSE, MAE, etc.), data-split protocol, number of option quotes, or statistical tests; without these the comparative performance statements cannot be evaluated."},{"response":"The penalty terms (calendar-spread, butterfly, and vertical-spread violations) and their weighting (λ = 0.1 for the main experiments) are defined in Equation (5) of Section 3.2. We will insert a concise parenthetical in the revised abstract: “with soft penalties (λ = 0.1) on calendar, butterfly and vertical-spread arbitrage”. This makes the loss formulation explicit without lengthening the abstract unduly.","revision_made":"yes","referee_comment":"[Abstract] Abstract: no explicit formulation is given for the soft arbitrage penalty terms (calendar-spread, butterfly, etc.) or their weighting in the loss; this prevents assessment of whether the chosen penalties are sufficient to enforce all relevant static no-arbitrage conditions or whether they introduce compensating inconsistencies under sparse sampling."},{"response":"We will update the abstract to state that results use SPX quotes 2018–2022 with an 80/20 chronological split, and that sparsity is varied from 20 to 200 quotes to probe liquid versus illiquid regimes. The generalization of the soft-penalty benefit across volatility regimes is shown in Figure 7; we will add a one-sentence reference to this figure in the abstract.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the reported results are described as holding 'on option market data' yet no information is supplied on train/test splits, out-of-distribution regimes, or liquidity/volatility regimes tested; this leaves the generalization claim for soft penalties unverified."}],"tokens_in":1286,"tokens_out":594,"duration_ms":21006,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this is an application paper: it takes established models like Transformers, U-Nets, and VAEs, adds standard soft penalty terms for calendar spreads and butterflies, and benchmarks them against SVI on option quotes. The results indicate that the neural approaches handle sparse data better than the parametric baseline and that the penalties cut arbitrage violations without much accuracy loss.\n\nWhat stands out is the direct comparison across five architectures plus the explicit accuracy-consistency trade-off analysis. The authors test on real market data and show the penalties deliver measurable improvement in constraint satisfaction. That is useful for practitioners who need surfaces that stay arbitrage-free under limited observations.\n\nThe soft spots are the missing details. The abstract gives no exact penalty formulations, no data split information, no error metrics with confidence intervals, and no out-of-distribution tests. Without those, it is hard to judge whether the reported gains hold when liquidity or volatility regimes shift. The generalization claim therefore rests on in-sample behavior alone.\n\nThe work is for people building production vol surfaces who already know the no-arbitrage conditions and want to see how off-the-shelf networks behave under them. It does not introduce new theory or first-principles derivations.\n\nI would send it to peer review. The empirical comparison is concrete enough to be worth referee time even if the methods section needs expansion and the robustness checks need strengthening.","headline":"This paper compares known neural architectures for volatility surface reconstruction with added soft no-arbitrage penalties and reports that Transformers and U-Nets perform best on the tested market data.","tokens_in":2117,"tokens_out":358,"would_cite":false,"duration_ms":17939,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Transformer and U-Net models reconstruct implied volatility surfaces from sparse noisy quotes while soft no-arbitrage penalties cut violations.","keywords":["implied volatility surface","volatility surface reconstruction","deep learning","no-arbitrage constraints","transformer","u-net","option market data","sparse data"],"falsifier":"Reconstructed surfaces that still permit static arbitrage, such as negative butterfly prices or calendar-spread violations, on held-out market data would show the penalties fail to deliver consistent surfaces.","tokens_in":2463,"feed_emoji":"📈","tokens_out":679,"duration_ms":23286,"temperature":0.7,"pith_summary":"The paper establishes that deep neural networks can recover implied volatility surfaces from incomplete and noisy option market data by embedding no-arbitrage conditions directly into the training process. It tests multiple architectures against the classical SVI parameterization and finds that Transformer and U-Net models maintain high accuracy especially when observations are sparse, with the added penalties sharply lowering arbitrage violations at modest cost to fit quality. A sympathetic reader cares because accurate, arbitrage-free volatility surfaces underpin option pricing, hedging, and risk calculations, yet real quotes are frequently missing or noisy. The results quantify the accuracy-consistency trade-off across architectures and penalty strengths on actual market data.","feed_headline":"Transformers rebuild volatility surfaces from sparse quotes","feed_subtitle":"Soft no-arbitrage penalties during training sharply cut violations while keeping reconstruction error moderate on market data.","key_machinery":"Neural network architectures trained with soft arbitrage penalty terms added to the loss function to enforce no-arbitrage conditions during reconstruction of implied volatility surfaces from sparse quotes.","core_discovery":"Transformer and U-Net architectures achieve strong reconstruction accuracy, particularly under sparse observation regimes, while soft arbitrage penalties significantly reduce arbitrage violations with moderate impact on reconstruction error. The models are compared to multilayer perceptrons, convolutional networks, variational autoencoders, and classical SVI parameterizations on option market data, with explicit analysis of how reconstruction error and arbitrage consistency trade off across architectures and regularization strengths.","pith_inferences":["The same penalty-augmented training could be applied to reconstruct other surfaces such as local volatility or correlation matrices.","Real-time updating of surfaces from streaming quotes becomes feasible if the models run at market speed.","Hybrid pipelines that start with an SVI fit and then apply a neural correction layer may combine the strengths of both approaches.","Out-of-sample tests on data from stressed market periods would reveal whether the learned penalties remain effective outside the training distribution."],"forward_implications":["Transformer and U-Net models deliver the highest reconstruction accuracy when option quotes are sparse.","Adding soft arbitrage penalties produces large reductions in arbitrage violations relative to unconstrained networks.","The increase in reconstruction error from the penalties stays moderate across tested regularization strengths.","The deep learning approach outperforms classical SVI parameterization on the same market data sets.","Accuracy and no-arbitrage consistency can be balanced by adjusting the penalty weight during training."],"fun_headline_variants":["Transformers outperform in sparse volatility surface reconstruction","U-Nets and Transformers best for sparse volatility reconstruction","Soft arbitrage penalties reduce violations in neural models","Deep learning reconstructs volatility surfaces with no-arbitrage"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The chosen soft arbitrage penalties will generalize to unseen market regimes and will not introduce new inconsistencies not captured by the penalty formulation.","fun_headline_variants_meta":{"raw":{"variants":["Transformers outperform in sparse volatility surface reconstruction","U-Nets and Transformers best for sparse volatility reconstruction","Soft arbitrage penalties reduce violations in neural models","Deep learning reconstructs volatility surfaces with no-arbitrage"]},"model":"grok-4.3","cost_usd":0.005081,"raw_usage":{"total_tokens":2408,"prompt_tokens":536,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":50812000,"prompt_tokens_details":{"text_tokens":536,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1814,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":536,"tokens_out":58,"duration_ms":19836,"temperature":1.0,"reasoning_tokens":1814,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T16:47:16.905719+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Reconstructed surfaces that still permit static arbitrage, such as negative butterfly prices or calendar-spread violations, on held-out market data would show the penalties fail to deliver consistent surfaces.","supporting_citations":[],"review_version":1}