{"id":"96732422-83e4-44b9-8011-11763e39e782","arxiv_id":"2509.01743","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conditional VAE controlled by quantified shape features can generate implied volatility surfaces matching user-specified level, slope, curvature, and term structure.","lead":"This paper builds a variational autoencoder that generates synthetic option implied volatility surfaces with user-controlled shape features such as level, slope, curvature, and term structure. It matters because controllable synthetic market scenarios are used in stress testing, risk management, and data-driven trading simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Anchor-point features are extrapolated from τ∈[0.1,0.6] to τ→0; if the extrapolation is unstable, sub-0.1% control errors only show matching of regression coefficients, not control of robust surface shape.","rationale":"The reader's weakest assumption and my concern converge: anchor-point extrapolation is the weakest load-bearing point. The paper's strongest claim is precise control and arbitrage-free generation; both hinge on the feature extractor. The closed-loop evaluation is a compounding issue: the same regression operator defines both the target y and the measured y(generated), so low error can be achieved by matching regression coefficients even if the economically relevant surface behavior is not controlled. However, I do not think this invalidates the method internally; it only limits the financial interpretation of the claimed control. The proposed test would settle whether the high precision survives independent or alternative quantification. Since the reader already assigned CONDITIONAL for essentially this reason, no verdict change is warranted.","tokens_in":34126,"tokens_out":4043,"duration_ms":50281,"concrete_test":"Re-run Experiment I (single-feature level control) using an alternative feature extractor that estimates features at the shortest available maturity τ=0.1, m=0 via local finite differences, and also refit the original polynomial with J=2 vs J=3 and alternative L_j choices. If the generation error measured by the alternative extractor rises materially (e.g., from <0.1% to >1%), or if extracted feature values for the same 1,000 surfaces shift substantially with polynomial order, then the reported precision is an artifact of the extrapolated regression operator rather than control of a robust surface shape.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the four regression coefficients at (τ=0, m=0) being stable, identifiable proxies for economically meaningful IVS shape. The data grid only covers τ∈[0.1,0.6] (Section 5, App. A), so the anchor point is an extrapolation, not an interpolation point. Polynomial regression over this grid can produce coefficients that vary strongly with the chosen polynomial orders J and L_j (App. B Eq. 33), and the term-structure slope Σ1,0 is essentially the coefficient of a linear-in-τ term fit far from τ=0. Because the same operator is used to define y(given) and to measure y(generated), the evaluation in Sections 5.1–5.2 can report low error even if the decoder only matches polynomial coefficients while the actual short-expiry/ATM shape drifts. The paper itself acknowledges the regression cannot capture singular short-end slopes (Remark after Eq. 12), which is exactly the regime where the anchor limit matters most. Thus the 'meaningful shape features' and 'precise controllability' claims are load-bearing on an unvalidated extrapolation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a controllable VAE for generating implied volatility surfaces (IVSs) with user-specified shape features. Four features (level, slope, curvature, term-structure slope) are quantified as coefficients of a bivariate polynomial regression evaluated at an anchor point (τ→0, m=0). These features serve as controllable variables y, while residual latent variables z capture remaining variability; training uses a conditional VAE objective with a β-weighted KL term. Experiments on 60,000 synthetic Heston/SABR surfaces report low generation errors in single- and multi-feature control, latent-traversal diversity, and no-arbitrage checks with a post-generation latent-space repair algorithm.","tokens_in":34430,"tokens_out":4662,"duration_ms":55993,"significance":"If the central claims hold, the framework would be a practically useful tool for stress testing and scenario generation, combining interpretable financial features with deep generative modeling. The paper contains a clean derivation of the conditional VAE ELBO (Appendix C), tests on two stochastic-volatility model families, and explicit no-arbitrage validation on generated surfaces. However, the evidence for the main claim of 'precise controllability' is weakened by two intertwined issues: the anchor-point features are extrapolated far outside the data support, and the evaluation applies the same regression operator that defines the conditioning labels to the generated surfaces. These issues make the reported sub-0.1% errors partly self-referential and not yet established as control of market-relevant surface shape.","major_comments":[{"comment":"The four controlled features are regression coefficients fitted on τ∈[0.1,0.6] and m∈[−0.27,0.27] but evaluated at the anchor point (τ=0, m=0). In particular, Σ1,0 is the τ-derivative at τ=0 estimated from a polynomial fit far from the origin, and the Remark after Eq. (12) acknowledges that the regression cannot capture singular short-end slopes. Since the data grid does not include τ near 0, the features are extrapolations, not interpolations. The sub-0.1% errors in §5.1.1 therefore establish control over polynomial coefficients, not necessarily over economically meaningful surface shape. The paper should provide a stability analysis with respect to the polynomial orders J,L(J) (which are never stated) and validate control errors using features measured on the observed grid, e.g., at the shortest maturity τ=0.1.","section":"§3.3, Eq. (33), and Appendix B"},{"comment":"The generation error is computed by applying the same differential operator P (Eqs. 9–12) that defines the conditioning labels y(given) to the generated surface to obtain y(generated). The decoder is trained with labels extracted via this same operator (Section 4.3). Low |e| thus partly reflects how well the decoder inverts P on the training manifold, rather than how well the generated IVS matches an independently defined shape target. This closed loop does not invalidate the method, but it means the headline 'control accuracy' is not yet evidence of control over the underlying surface geometry. Please add an independent evaluation, e.g., fit SVI or a different parametric form to generated surfaces and compare those parameters, or measure features by finite differences at the shortest available maturity.","section":"§5.1.1 and Eq. (10)"},{"comment":"The post-generation repair algorithm fixes only 45.5% of the violating surfaces, leaving 2,950 of 60,000 generated samples (4.9%) in violation of static no-arbitrage conditions. The text concludes that the method 'can be corrected effectively' and that violations 'remain rare and can be corrected effectively'; given that over half of the violations are unrepaired, this is an overstatement. Moreover, Algorithm 2 optimizes z while keeping y fixed, but the decoded repaired surface may still have a different measured y, and LMSE only constrains the surface deviation, not the feature deviation. Please report the distribution of y errors before and after repair, and discuss the residual 4.9% violation rate.","section":"§5.3, Table 3"},{"comment":"The paper repeatedly contrasts the framework with 'black-box generators' and claims that it 'combines interpretability, controllability, and flexibility', but no baseline comparisons are provided. Without comparisons to, e.g., a conditional VAE that conditions on raw grid points or on PCA scores, a linear decoder-plus-regression model, or an existing IVS GAN, the experimental section cannot support the relative advantages asserted in the introduction and conclusion. At minimum, report control error and arbitrage-violation rates for a simple conditional VAE or a regression-based generator on the same dataset.","section":"§5 (all experiments)"}],"minor_comments":[{"comment":"The polynomial orders J and L(J) in Eq. (33) are never specified, and no sensitivity study is reported. Since these orders define the feature values, the reader cannot reproduce the experiments or assess the stability of the anchor-point extrapolation.","section":"§3.3 / Appendix B"},{"comment":"The KL weight β is listed as 'problem-specific' but only the three-feature experiment states a value (β=5×10^-8). Please report β for the single- and four-feature experiments, and describe how it was selected.","section":"§5.2 / Table 1"},{"comment":"The thresholds ε1, ε2, maximum iterations T, and the learning rate η in Algorithm 2 are not given numerically, and no convergence statistics are reported. Please provide the implementation values or a reference to released code.","section":"§4.4 / Algorithm 2"},{"comment":"The latent traversal uses z2=±8, which is 8 standard deviations under the standard normal prior and has negligible prior probability. The claim that the model 'can extrapolate meaningfully' would be stronger if these out-of-distribution surfaces were also checked for no-arbitrage violations and compared with in-distribution latent samples.","section":"§5.1.3"},{"comment":"The legend entries in Figures 11 and 16 appear as 'uni000000/...' encoding artifacts rather than readable feature names; these should be fixed.","section":"Figures 11 and 16"},{"comment":"Typos and minor wording: 'A serial of studies' should be 'A series of studies'; in Section 5.1.1, the sentence beginning 'whered y(given)L' has a stray 'd'.","section":"§1 and §5.1.1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things before anything else. The paper does something concrete—it makes an IVS generator that takes four interpretable shape features as explicit conditioning variables and leaves a residual latent for the rest—and it writes the whole thing up clearly and honestly. But the precision numbers (0.1% level errors, etc.) are partly a closed loop: the same polynomial regression that defines the conditioning labels is used to measure the features of the output. That does not make the result vacuous, but it means the headline claim about controlling “meaningful” shape features is not yet proven.\n\nWhat is genuinely new: earlier VAE/GAN papers on IVS generate or impute; this one adds quantified anchor-point features (level, slope, curvature, term structure) as conditioning and demonstrates that a decoder can hit specified regression coefficients with small error. The latent traversal experiments are useful—z2 mapping to slope in the one-feature model, and the way supervising all four features kills residual variation—are the kind of result that makes the architecture plausible. The 60,000-surface in-distribution arbitrage-free check is a nice sanity test, and the paper explicitly flags its own limitations (the polynomial regression cannot capture exploding short-end slopes; L-BFGS repair is local; only ~45.5% of violations get repaired). Credit for that.\n\nSoft spots, in order of size. First, the anchor-point features are extrapolations. The data grid only has tau in [0.1, 0.6]; level, slope, curvature and term-structure are coefficients of a bivariate polynomial fit evaluated at tau=0. That is a long extrapolation for the term-structure slope, and the paper reports no sensitivity to the polynomial orders J and L_j. If coefficients move around with the regression order, the sub-0.1% errors just say the decoder solves the regression inverse on the training distribution, not that it controls a stable economically meaningful shape. Second, no baselines. Third, no market data; synthetic Heston/SABR surfaces are smooth, so the hard near-expiry behavior never appears in the training distribution. Fourth, the repair rate is a minor point but worth being precise about.\n\nNone of this is fatal. The framework is coherent and the writing is exact. I would send it to a serious referee; in revision I'd ask for baseline comparisons against [18, 19, 16, 17], one real-market dataset, a sensitivity analysis of the feature regression, and an independent way to measure the generated surfaces (e.g., SVI parameters or local-surface diagnostics) that does not reuse P. With those, the contribution would be solid. Without them, the “meaningful control” claim needs to be read as a bounded claim about regression coefficients.\n\nReader: quantitative-finance people building synthetic IVS generators and any applied ML person thinking about evaluation circularity in conditional generation. Not a must-read, but worth a slot if you work in this area.","headline":"Useful applied step for controllable IVS generation, but the control evidence is partly a closed loop around extrapolated polynomial features; deserves refereeing, not desk rejection.","tokens_in":34904,"tokens_out":2924,"would_cite":false,"duration_ms":35506,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","91G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A controllable VAE generates implied volatility surfaces whose level, slope, curvature, and term-structure slope match user targets within fractions of a percent, while staying free of static arbitrage for in-distribution samples.","keywords":["implied volatility surface","variational autoencoder","controllable generation","shape features","no-arbitrage conditions","latent space disentanglement","synthetic market data","stress testing"],"falsifier":"Fit the paper's bivariate polynomial to surfaces generated by a model with known short-time asymptotics — for instance a rough-volatility model whose ATM slope diverges as τ→0 — while only observing the grid τ ∈ [0.1, 0.6]. If the extrapolated coefficients systematically miss the known limits (which finite-order polynomials must, at a singularity), then the 'slope' and 'term-structure' control variables are artifacts of the fitting basis, and a user commanding those values is not commanding the market feature the paper claims to control. A direct check of the same kind is available on real dat","tokens_in":34043,"feed_emoji":"📈","tokens_out":13930,"duration_ms":131238,"temperature":0.7,"pith_summary":"The paper sets out to build a generative model of implied volatility surfaces that takes explicit orders: give it a target volatility level, skew slope, smile curvature, or term-structure slope, and it produces a synthetic surface with those characteristics, rather than just any plausible-looking surface. The mechanism is a variational autoencoder whose latent space is split into a controllable part — the four shape features, quantified by polynomial regression at a short-maturity at-the-money anchor point — and a residual part that preserves diversity. The reported results claim control errors below 0.1% for the level feature in most single-feature trials, comparable accuracy when three or four features are controlled together, and zero static-arbitrage violations among all 60,000 in-distribution generated surfaces, with a latent-space repair step correcting about half of the violations that appear under extreme sampling. If the claims hold, the payoff is a practical tool for stress testing and market simulation, where the user specifies the kind of market state the synthetic data should represent.","feed_headline":"Volatility surfaces on demand: VAE hits shape targets under 0.1%","feed_subtitle":"A split-latent VAE lets users specify level, skew, smile, and term structure in synthetic market data for stress tests.","key_machinery":"The load-bearing object is the split latent representation (y, z) of a conditional VAE: y holds the commanded shape features, z holds residual variation. The features are the coefficients of a bivariate Taylor expansion of the surface at the anchor point (τ→0+, m=0) — level Σ0,0, slope Σ0,1, curvature Σ0,2, term-structure slope Σ1,0 — estimated by polynomial regression on the observable grid. The decoder maps (y, z) to the full surface, and a β-weighted KL term keeps z near a standard Gaussian prior. The repair step minimizes LCalendar + LButterfly + LMSE over z alone with L-BFGS, which lets the pipeline remove arbitrage violations without undoing the requested features.","core_discovery":"Central discovery: quantitative control of IVS shape and financial validity can live in one VAE. Four shape features (level, slope, curvature, term structure) are anchor-point derivatives estimated by bivariate polynomial regression and fed to the decoder as controllable latent variables. A post-generation step re-optimizes only the residual latent variables against calendar-spread and butterfly penalties, leaving commanded features untouched. Results: level-generation error below 0.1% in most trials; four-feature control keeps errors mostly below 10^-2 and renders residual latents nearly inert; all 60,000 in-distribution surfaces satisfy no-arbitrage conditions; about 45.5% of rare tail vio","pith_inferences":["The same split-latent recipe should transfer wherever a high-dimensional market object is summarized by a handful of interpretable coefficients — for instance implied correlation surfaces, variance swap term structures, or CDS surfaces — since nothing in the architecture depends on equities except the choice of the four features.","A natural upgrade the authors flag but do not pursue: replace the polynomial basis with functions that can represent the exploding ATM slope of short-expiry surfaces, as in rough-volatility models. Until then, 'slope' means the polynomial's slope on the τ ≥ 0.1 grid extrapolated to τ = 0, and that extrapolation deserves validation against true short-time asymptotics before the control variable is ","The repair algorithm keeps y fixed by construction but can push z away from the Gaussian prior; checking the posterior density of repaired z's and re-testing arbitrage after a second repair cycle would clarify whether the 45.5% correction rate is a property of the decoder or an artifact of L-BFGS initialization, which the paper itself flags as a local-optimization procedure.","For deployment, the convex-hull result suggests a cheap safety wrapper: constrain each sampled y to the training feature hull (a few linear inequality checks) and z to the central region, and arbitrage violations should effectively never occur — trading a little diversity for a certificate of validity."],"forward_implications":["Stress-test scenario design becomes a matter of dialing in feature values: surfaces with elevated volatility levels, steepened skews, or inverted term structures can be generated in a single decoder pass, without re-calibrating a pricing model.","Surfaces sampled within the convex hull of the training feature space are, according to Section 5.3, observed to be free of calendar-spread and butterfly arbitrage in all 60,000 draws, so downstream pricing or hedging tasks can consume them directly.","When fewer than four features are controlled, the residual latent variables still carry meaningful structure (z2 correlates with slope at about -0.991 in the single-feature experiment), so the model doubles as an exploratory tool for finding stylized facts the user did not specify.","When all four stylized factors are supervised, the residual latent variables become nearly inert (feature fluctuations on the order of 10^-3 across an eight-standard-deviation sweep), meaning the four-feature model gives essentially complete control over the dominant modes of IVS variation.","Extreme latent draws can produce slope values beyond the training range, so the generator can extrapolate to market states — steep skews or inverted term structures — that are rare or absent in the training data."],"supporting_citations":[{"why":"Supplies the regression-based method for quantifying the four shape features at the anchor point (τ→0+, m=0), used in Sections 3.3 and Appendix B.","marker":"[27]"},{"why":"Defines the total-variance formulation and the calendar-spread and butterfly no-arbitrage conditions used to validate and repair generated surfaces.","marker":"[8]"},{"why":"Provides the conditional/semi-supervised VAE variational objective (ELBO, KL term) that the controllable VAE architecture extends.","marker":"[13]"},{"why":"Supplies the reparameterization trick and the reduction of the reconstruction term to mean squared error used in training.","marker":"[45]"},{"why":"Motivates quantitative feature control by showing PCA components leave variance unexplained and do not measure the resulting features.","marker":"[12]"},{"why":"The first VAE model of IVSs; the paper extends it by adding explicit and quantitative control of shape features.","marker":"[18]"},{"why":"Prior hybrid VAE-with-SDE work on arbitrage-free IVS generation; supplies the baseline on latent dimensionality and arbitrage enforcement.","marker":"[19]"},{"why":"Heston model used to generate half of the 60,000-surface training dataset.","marker":"[6]"},{"why":"SABR model used to generate the other half of the training dataset, including its closed-form implied volatility approximation.","marker":"[1]"},{"why":"L-BFGS optimizer used in the post-generation latent-space repair algorithm.","marker":"[40]"}],"fun_headline_variants":["Dial in volatility shape: VAE generates surfaces to spec","User-defined IVS features: VAE hits shape targets","VAE crafts no-arbitrage vol surfaces with feature control","Specify level, skew, curvature: VAE obeys IVS shape","Sub-0.1% level error: VAE for controllable IVS"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The four controlled 'shape features' are coefficients of a polynomial fitted to the surface on a grid whose shortest maturity is τ = 0.1 and then extrapolated to the anchor point τ → 0+, m = 0; if that extrapolation does not capture the economically meaningful limiting shape, hitting the coefficients precisely does not control the market-relevant surface.","fun_headline_variants_meta":{"raw":{"variants":["Dial in volatility shape: VAE generates surfaces to spec","User-defined IVS features: VAE hits shape targets","VAE crafts no-arbitrage vol surfaces with feature control","Specify level, skew, curvature: VAE obeys IVS shape","Sub-0.1% level error: VAE for controllable IVS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000337,"raw_usage":{"total_tokens":1704,"prompt_tokens":751,"completion_tokens":953,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":862}},"tokens_in":495,"tokens_out":953,"duration_ms":10779,"temperature":1.0,"reasoning_tokens":862,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:13:04.012573+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the paper's bivariate polynomial to surfaces generated by a model with known short-time asymptotics — for instance a rough-volatility model whose ATM slope diverges as τ→0 — while only observing the grid τ ∈ [0.1, 0.6]. If the extrapolated coefficients systematically miss the known limits (which finite-order polynomials must, at a singularity), then the 'slope' and 'term-structure' control variables are artifacts of the fitting basis, and a user commanding those values is not commanding the market feature the paper claims to control. A direct check of the same kind is available on real dat","supporting_citations":[{"cited_title":"Implied stochastic volatility models","cited_arxiv_id":null,"evidence_quote":"Supplies the regression-based method for quantifying the four shape features at the anchor point (τ→0+, m=0), used in Sections 3.3 and Appendix B."},{"cited_title":"Arbitrage-free svi volatility surfaces","cited_arxiv_id":null,"evidence_quote":"Defines the total-variance formulation and the calendar-spread and butterfly no-arbitrage conditions used to validate and repair generated surfaces."},{"cited_title":"Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling","cited_arxiv_id":null,"evidence_quote":"Provides the conditional/semi-supervised VAE variational objective (ELBO, KL term) that the controllable VAE architecture extends."},{"cited_title":"Simulation of arbitrage-free implied volatility surfaces","cited_arxiv_id":null,"evidence_quote":"Motivates quantitative feature control by showing PCA components leave variance unexplained and do not measure the resulting features."},{"cited_title":"Variational Autoencoders: A Hands-Off Approach to Volatility","cited_arxiv_id":"2102.03945","evidence_quote":"The first VAE model of IVSs; the paper extends it by adding explicit and quantitative control of shape features."},{"cited_title":"Arbitrage-free implied volatility surface generation with variational autoencoders","cited_arxiv_id":null,"evidence_quote":"Prior hybrid VAE-with-SDE work on arbitrage-free IVS generation; supplies the baseline on latent dimensionality and arbitrage enforcement."},{"cited_title":"A closed-form solution for options with stochastic volatility with applications to bond and currency options","cited_arxiv_id":null,"evidence_quote":"Heston model used to generate half of the 60,000-surface training dataset."},{"cited_title":"Managing smile risk","cited_arxiv_id":null,"evidence_quote":"SABR model used to generate the other half of the training dataset, including its closed-form implied volatility approximation."},{"cited_title":"Liu and Jorge Nocedal","cited_arxiv_id":null,"evidence_quote":"L-BFGS optimizer used in the post-generation latent-space repair algorithm."}],"review_version":1}