{"id":"e572ed84-e4f8-4030-9417-980aef92d308","arxiv_id":"2608.12587","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"DYSANOS generates smooth, statically arbitrage-free option surfaces from a low-dimensional hidden state, but the generated paths still exhibit dynamic arbitrage in period-to-period trading.","lead":"This paper introduces a generative model that creates years of daily option price surfaces that are smooth and free of static arbitrage by construction. It also shows, honestly, that such surfaces still admit dynamic arbitrage when options are traded before expiry.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The static no-arbitrage construction is credible, but the realism claim rests on the untested Appendix A.4/A.5 assumption that every dropped option/expiry is a data-alignment artifact; if genuine market effects are deleted, the fitted hidden states and simulated surfaces are biased.","rationale":"The reader's weakest-assumption and the concern identified here coincide: the mathematical core of the paper, namely that every decoded surface is statically arbitrage-free by construction, is sound and independently supported by the structural argument and the to-expiry arbitrage tests. The fragile step is the empirical bridge from cleaned data to a 'generative market model.' Appendix A.4 explicitly makes an untested assumption that fit failures are alignment artifacts rather than genuine market information, and Appendix A.5 acknowledges that liquidity filters may remove real long-dated observations. Since the hidden states h_t, the PCA factors, and the AR(1) dynamics are all learned from the post-filter arbitrage-free surfaces, any systematic deletion of genuine market effects propagates directly into the simulated surface distribution. This does not invalidate the static no-arbitrage construction, which is why I do not propose moving the verdict to REJECT; it does, however, keep the paper at CONDITIONAL, exactly as the reader concluded. The dynamic-arbitrage evidence in Section 3.4.2 is real but explicitly outside the central static claim, so I do not treat it as the most load-bearing concern. The proposed concrete test would settle whether the exclusion assumption is innocuous or materially biases the generative model.","tokens_in":21403,"tokens_out":10994,"duration_ms":124053,"concrete_test":"Choose a set of dates with high drop rates (e.g., early 2020 and dates with many filtered long-dated expiries). Re-run Appendix A.4 with a slack LP that does not require exact feasibility, and for each dropped expiry/strike compute the minimal bid-ask-consistent arbitrage violation cost. Then build two target grids, the current one and a permissive one that keeps every expiry/strike whose minimal violation is below a one-basis-point threshold; refit ML-SANOS and the AR(1)/PCA dynamics on both. Compare the fitted h_t series, the implied-volatility eigensurfaces, and the generated path diagnostics. If the permissive-set h_t or simulated IV distributions differ materially, or the minimal violations concentrate in tradable strikes with wide spreads, the Appendix A.4 assumption fails and the realism claim needs to be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DYSANOS has two separable claims: (i) every decoded surface is smooth and statically arbitrage-free by construction, and (ii) the generated paths are realistic enough to serve as a generative market model. Claim (i) is well supported by Theorem 2.5, the positive-volatility/sigmoid reparametrization in Definitions 2.7 and 2.8, and the zero to-expiry arbitrage results in Tables 2 and 3. Claim (ii) is where the paper is least secure. Appendix A.4 steps 1-5 fit a linear SANOS surface, then 'drop expiries for which a valid fit cannot be obtained' and 'drop any strikes where such a fit cannot be achieved, assuming that fit issues are due to data alignment challenges rather than genuine trading opportunities.' Appendix A.5 then removes options by moneyness, volume, and open-interest filters, with the authors themselves noting that the volume filter 'removes many options post 1Y' and that 'we suspect that this information in IvyDB is not reliable in the past.' The ML-SANOS states h_t, the PCA loadings, and the AR(1) generator are all estimated on the surviving, already arbitrage-free smoothed surfaces. If the excluded contracts contain real cross-sectional information, the target distribution and the fitted dynamics are biased, and the simulated paths, although still statically arbitrage-free by construction, are not a faithful model of the real S&P option market. The bias is not a mathematical flaw in the no-arbitrage construction, but it is the load-bearing empirical assumption connecting the construction to the paper's 'market model' claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DYSANOS, a generative model for entire paths of daily spot and option prices. Each simulated date's call price surface is decoded using a neural embedding NQ into the ML-SANOS parameterization of the SANOS construction, so that every decoded surface is smooth and statically arbitrage-free by construction. The authors fit the hidden states to S&P 500 Index options from OptionMetrics IvyDB for 2020-2025, fit a simple PCA-AR(1) baseline for surface dynamics, and evaluate the generated paths with distributional diagnostics, static no-arbitrage tests, and period-return arbitrage tests. The static no-arbitrage property is structural, inherited from Theorem 2.5 and the reparameterizations in Definition 2.8. The empirical results show zero to-expiry static arbitrage certificates for DYSANOS, while the period-return tests indicate possible dynamic arbitrage, which the authors acknowledge. The paper is transparent about the data pipeline and its limitations, but the realism of the generative model rests on assumptions in the preprocessing that are not yet validated.","tokens_in":21779,"tokens_out":6024,"duration_ms":59041,"significance":"If the results hold, DYSANOS is a meaningful advance over soft-constraint generative models: it provides smooth option surfaces for all strikes and expiries that are statically arbitrage-free by construction, not by penalty terms, and it allows any option to be priced at any simulated date. This is directly useful for training hedging algorithms such as Deep Hedging. The paper also provides careful and honest empirical diagnostics, including a positive control (PCA-IV) and an exact-state resampling test that rejects one of the two strongest dynamic-arbitrage candidates. The main weaknesses are the reliance on the data-preprocessing assumption in Appendices A.4 and A.5, the absence of code or data for reproducibility, and the unresolved dynamic-arbitrage candidates that temper the model's usefulness as a full market generator. The static no-arbitrage derivation is sound; the empirical realism claim needs further support.","major_comments":[{"comment":"The realism of the generative model rests on the data-preprocessing assumption stated in A.4: after fitting a linear SANOS surface, the authors 'drop expiries for which a valid fit cannot be obtained' and 'drop any strikes where such a fit cannot be achieved, assuming that fit issues are due to data alignment challenges rather than genuine trading opportunities.' If any of the excluded strikes or expiries contain genuine market information, the target surfaces, the ML-SANOS hidden states, and the fitted AR(1) dynamics are biased. The paper should report the fraction of strikes and expiries dropped per day, compare surfaces before and after the dropping, and run a sensitivity analysis that relaxes or removes the A.4/A.5 filters on a sub-period where the volume and open-interest filters are less problematic. The authors themselves note in A.5 that the volume filter 'removes many options post 1Y' for early years and suspect the IvyDB volume data 'is not reliable in the past'; this makes the long-expiry training data particularly dependent on the exclusion assumption and should be addressed directly.","section":"Appendix A.4/A.5"},{"comment":"The period-return tests find that DYSANOS has a higher validated occurrence of dynamic-arbitrage candidates than the PCA-IV benchmark in every row of Table 5; for example, at daily trading and zero cost the occurrence is 12.500% for DYSANOS versus 1.584% for PCA-IV. The exact-state replay in Table 6 rejects the second-ranked candidate because of observed losses, but leaves the first-ranked candidate unresolved after one million continuations, with a one-sided 95% Clopper-Pearson upper bound on loss probability of 0.000300%. This is a substantive limitation for the intended application of training hedging agents on generated paths. The authors should either extend the generator to reduce these candidates or clearly state in the abstract and introduction that the model is only certified to be statically arbitrage-free, and that dynamic arbitrage has been detected in the baseline version.","section":"Section 3.4.2, Tables 4 and 5"},{"comment":"The empirical claims in Sections 3.3 and 3.4, including the fitting error of 0.3% implied volatility and all arbitrage-test tables, cannot be verified without the code or the preprocessed data. The paper describes the pipeline in detail, but no code or data repository is provided. For a paper whose central contribution is an empirical generative model, a reproducibility statement with code and, if possible, the cleaned data tensors should be included. This is not a mathematical objection, but it is necessary for the claims to be independently checked.","section":"General / Reproducibility"}],"minor_comments":[{"comment":"The abstract contains a typo: 'investigate numerical resence of dynamic arbitrage' should read 'investigate numerical presence of dynamic arbitrage'.","section":"Abstract"},{"comment":"The sentence 'We use the tilde to distinguish real observed data from siumlated data' has a typo: 'siumlated' should be 'simulated'.","section":"Section 3.3"},{"comment":"The statement that 'The previous steps, in the end, simply cleaned market data. All remaining options are market tradable instruments' is slightly misleading because A.4 explicitly drops strikes and expiries; consider rewording to note that the remaining subset is assumed tradable and representative.","section":"Appendix A.7"},{"comment":"The relationship between the log-spot state h0 and the separately fitted spot dynamics in equation (16) is not fully explained. Figure 6 shows h0 as the first state, but Section 3 fits log-spot only afterwards; clarify how h0 is used in the simulation and how it relates to the spot process in equations (23)-(28).","section":"Section 3.2 / Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a quantitative finance journal. The static no-arbitrage construction is credible and the authors are unusually transparent about the dynamic-arbitrage findings. The main risk is the data preprocessing assumption in A.4/A.5, which directly affects the training distribution. If the authors provide sensitivity analyses showing that the dropped contracts do not materially change the learned surfaces and hidden states, and if they make code/data available, the paper would be acceptable. The 'first generative market model' claim is defensible when restricted to SANOS-type surfaces, but the wording should be checked against the broader literature on local-volatility-based generative models."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful paper, and it is honest about its limits. The static no-arbitrage construction is structural, not fitted, and the authors openly report that their baseline AR(1) dynamics admit dynamic arbitrage. The real soft spot is the data pipeline, which assumes away any strike or expiry that cannot be fitted arbitrage-free, so the fitted hidden states and surfaces may inherit selection bias.\n\nWhat is new is the ML-SANOS decoder. It reparametrizes SANOS into a smooth map from a low-dimensional state to an arbitrage-free surface, and that is a real enabler for generative simulation. The paper also ships a careful test suite for arbitrage, with a PCA-IV positive control, and it reports the dynamic arbitrage results instead of sweeping them under the rug. The importance-sampling construction for spot tails is thoughtful and correct.\n\nWhere I would be cautious: Appendix A.4/A.5 drops data that do not fit the linear SANOS preprocessing, on the assumption that fit failures are alignment artifacts. That is plausible but untested, and if some exclusions are genuine market effects—or simply illiquidity that still carries information—the target surfaces are biased. The volume filter also removes many options post 1Y in early years, which could distort the term structure. No code or data is shipped, so the empirical results are not directly reproducible. And the period-return tests find dynamic arbitrage candidates; the exact-state replay rejects one candidate and leaves one unresolved. That is not a fatal flaw, since the paper presents itself as a baseline, but it does temper the 'generative market model' claim.\n\nThe math is sound: Theorem 2.5 and the reparameterization genuinely guarantee static no-arbitrage for each decoded surface. The PCA-AR(1) dynamics are standard, but the authors say so plainly. The diagnostics are detailed and honest, with clear comparisons against historical moments and eigensurfaces.\n\nBottom line: this deserves peer review. It is a solid contribution for anyone working on simulation of option markets, deep hedging, or arbitrage-free parametrizations. I would send it to a serious referee. It is not ready as is—it needs code/data availability and a less cavalier treatment of the data-filtering assumptions—but the core construction is worth engaging with.","headline":"A structurally sound and honest baseline for arbitrage-free option-surface simulation; the static no-arbitrage claim holds, but the data-filtering assumptions and admitted dynamic arbitrage temper the broader 'market model' claim.","tokens_in":22346,"tokens_out":2136,"would_cite":true,"duration_ms":20499,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91G20","91G80"],"pacs":[],"model":"deepseek-v4-flash","headline":"DYSANOS is the first generative market model whose simulated option surfaces are smooth and statically arbitrage-free for every strike and expiry.","keywords":["generative market model","option surface generation","static arbitrage","dynamic arbitrage","SANOS","implied volatility surface","state space model","PCA"],"falsifier":"Run the exact-state replay of the strongest DYSANOS weekly period-return candidate with ten million independent physical continuations: if any loss beyond the numerical band occurs, dynamic arbitrage is established. Alternatively, on a sample of IvyDB days, re-run the Appendix A.4 arbitrage-free fit without dropping failing strikes and check whether the excluded quotes consistently admit positive butterfly or calendar profits after transaction costs; persistent profits would mean the preprocessing bias is real.","tokens_in":21130,"feed_emoji":"📈","tokens_out":6001,"duration_ms":52316,"temperature":0.7,"pith_summary":"The paper introduces DYSANOS, a generative model of daily spot and option prices designed to produce entire multi-year paths of S&P 500 option surfaces. Its central claim is that every decoded surface is smooth and free of static arbitrage by construction: for each simulated date, call prices for all strikes and expiries satisfy the no-butterfly, no-calendar-spread, and boundary conditions, so any option can be priced off the surface. The authors build a baseline auto-regressive hidden-state model driven by five PCA factors, train the decoder on IvyDB SPX data from 2020 to 2025, and benchmark it against a pure implied-vol PCA simulator. A sympathetic reader would care because hedging and reinforcement-learning agents trained on simulated markets need surfaces that are realistic and free of artificial arbitrage, and earlier simulators were limited to discrete floating grids. The paper also reports that although no static arbitrage appears when options are held to expiry, numerical tests find dynamic arbitrage when the same contracts are marked to market between trading periods.","feed_headline":"Simulator generates arbitrage-free option surfaces","feed_subtitle":"Every simulated SPX surface is smooth and static-arbitrage-free; period-to-period trading still reveals dynamic arbitrage.","key_machinery":"The load-bearing object is the ML-SANOS decoder: a reparameterization of the SANOS arbitrage-free call-price family in which all surface parameters live in an unrestricted real space $\\mathbb{R}^{N+NM}$. A feed-forward network $NQ_\\theta$ maps the hidden state $h$ to these parameters; sigmoid activations keep forward total volatilities and squared discrete local volatilities positive, the discrete local volatility operator $\\Xi_j$ (computed with a tridiagonal Thomas solve) enforces the martingale density constraints, and Black–Scholes interpolation in total volatility produces a $C^\\infty$ surface that is statically arbitrage-free whenever $\\mu>0$. This decoder is what lets any generative time-series model on $h$ produce option surfaces that are arbitrage-free by construction.","core_discovery":"The paper claims that DYSANOS is the first generative market model to output smooth, strictly statically arbitrage-free option surfaces for all strikes and expiries along simulated paths. The construction is a two-stage pipeline: a decoder network maps a low-dimensional hidden state $h_t \\in \\mathbb{R}^{20}$ to the unrestricted coordinates $x$ of an ML-SANOS surface, and a deterministic reparameterization converts $x$ into a martingale density $q$, total volatilities $W$, and a smooth call price function $C(T,K)$ that satisfies the conditions of Theorem 2.2. A surface-first AR(1) state-space model then drives $h_t$ with mean reversion and PCA-factor noise, with spot simulated conditionally and a Gaussian-bridge importance-sampling scheme to reach far strikes. Trained on five years of S&P 500 options, the baseline captures the first-order level, skew, term, and leverage effects, and the paper's arbitrage tests confirm zero static-arbitrage certificates for DYSANOS while the PCA-IV canary exhibits many; the same tests, however, find dynamic arbitrage in DYSANOS's period-to-period returns.","pith_inferences":["If the decoder really guarantees static no-arbitrage for any hidden state, the generative bottleneck is purely dynamical: replacing the AR(1) driver with a more expressive time-series model would inherit the arbitrage-free property, so dynamic-arbitrage removal becomes the main target.","The dynamic arbitrage detected in period returns suggests that pricing the same cash contract across dates needs an additional conditional no-arbitrage constraint; a natural testable extension is to add a penalty on conditional expected returns and rerun the exact-state replay to see whether the remaining candidate disappears.","The preprocessing rule that drops strikes or expiries failing the arbitrage-free fit biases the target surfaces if some of those exclusions are genuine market effects; this could be tested by comparing cleaned versus unfiltered surfaces on days with heavy quoting.","A simulator with static arbitrage excluded but dynamic arbitrage present will give learning agents a spurious profit source; the paper's diagnostic tests could serve as a standard acceptance check for any generative option-market model."],"forward_implications":["Every simulated surface prices any option at any strike and expiry, so simulation-based hedging algorithms can choose instruments daily without restricting expiries.","Held-to-expiry portfolios on generated paths show no static arbitrage: the statewise LP tests found zero candidates for DYSANOS in 100,000 and 1,000,000 paths, while the PCA-IV baseline produced certificates at all cost levels.","Dynamic arbitrage is not excluded: fixed-contract period-return tests with one million paths detected arbitrage candidates in DYSANOS at daily and weekly trading intervals, with occurrence rates above those of PCA-IV.","The AR(1) baseline reproduces the dominant level, skew, term, and leverage directions and about 99.3% of surface-innovation variance with five factors, but understates volatility clustering, crisis tails, and the higher eigensurfaces.","Because the decoder is smooth and GPU-friendly, simulated surfaces can be repriced in batch gradient-descent training, linking the generative model to learning-based hedging."],"supporting_citations":[{"why":"Supplies the SANOS family of smooth, statically arbitrage-free call price surfaces that ML-SANOS reparameterizes and that the generative model decodes.","marker":"[BHK+26]"},{"why":"Establishes the discrete local volatility representation and the static-arbitrage characterization (Theorem 2.2) that the decoder enforces.","marker":"[BR15]"},{"why":"Provides the PCA view of implied-volatility surface dynamics used to choose the five driving factors of the baseline AR(1) model.","marker":"[CFD02]"},{"why":"Shows linear interpolation of option prices is the most expensive arbitrage-free interpolation, motivating the smooth SANOS surfaces used here.","marker":"[Bue06]"},{"why":"Contributes the volatility-interpolation ideas behind the in-homogeneous strike transition operator $\\Xi_j$ used in the DLV construction.","marker":"[AH11]"}],"fun_headline_variants":["First generative model outputs static-arbitrage-free option surfaces","Simulates smooth SPX option surfaces with zero static arbitrage","DYSANOS generates smooth surfaces, but paths carry dynamic arbitrage","Generative option surfaces: no static arb, but dynamic arb persists"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fidelity of the generated market depends on the data-processing assumption that any strike or expiry which cannot be fitted arbitrage-free is a data alignment error rather than a genuine trading opportunity, so dropping it does not bias the learned target surfaces.","fun_headline_variants_meta":{"raw":{"variants":["First generative model outputs static-arbitrage-free option surfaces","Simulates smooth SPX option surfaces with zero static arbitrage","DYSANOS generates smooth surfaces, but paths carry dynamic arbitrage","Generative option surfaces: no static arb, but dynamic arb persists"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1421,"prompt_tokens":896,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":451}},"tokens_in":512,"tokens_out":525,"duration_ms":4589,"temperature":1.0,"reasoning_tokens":451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:05:07.894626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact-state replay of the strongest DYSANOS weekly period-return candidate with ten million independent physical continuations: if any loss beyond the numerical band occurs, dynamic arbitrage is established. Alternatively, on a sample of IvyDB days, re-run the Appendix A.4 arbitrage-free fit without dropping failing strikes and check whether the excluded quotes consistently admit positive butterfly or calendar profits after transaction costs; persistent profits would mean the preprocessing bias is real.","supporting_citations":[],"review_version":1}