{"paper_id":"2509.01743","title":"Controllable Generation of Implied Volatility Surfaces with Variational Autoencoders","authors":["Jing Wang","Shuaiqiang Liu","Cornelis Vuik"],"paper_date":"2025-09-01","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2509.01743v1","page_url":"https://thequant.space/flowcharts/controllable-generation-of-implied-volatility-surfaces-with/","scores":{"empirical_rigor":6.5,"math_complexity":8,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes a controllable variational autoencoder framework that disentangles user‑specified shape features from residual latent factors to generate implied volatility surfaces with explicit, interpretable control.","quote":"This paper presents a deep generative modeling framework for controllably synthesizing implied volatility surfaces (IVSs) using a variational autoencoder (VAE).","summary":"This paper presents a deep generative modeling framework for controllably synthesizing implied volatility surfaces (IVSs) using a variational autoencoder (VAE). Unlike conventional data-driven models, our approach provides explicit control over meaningful shape features (e.g., volatility level…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The latent variables y (controllable features) and z (residual) are independent in the generative model.","type":"model","stated":"explicit","quote":"This independence assumption encourages a disentangled latent representation,i.e., y governs specific, interpretable shape features, while z accounts for residual variation.Notably, the dimensionality of y and z is typically much smaller than that of x.","support":"full"},{"assumption":"The residual latent variable z follows a standard Gaussian distribution.","type":"statistical","stated":"explicit","quote":"For tractability, the residual latent variable is assumed to follow the standard Gaussian distribution, z ∼N(0, I), (17) and the posteriors have the following Gaussian form qϕ(z | x, y) ∼N \u0000z | µz, diag(σ2 z) \u0001 , (18a) pθ(x | y, z) ∼N \u0000µx, I \u0001 .","support":"full"},{"assumption":"The regression error when extracting shape features is negligible, making y deterministic given x.","type":"model","stated":"explicit","quote":"Regarding the controllable variable y, we assume the regression error in Equation (10) is negligible when extracting shape features (Section 3).","support":"full"},{"assumption":"The latent dimension is fixed at 5 for all experiments.","type":"model","stated":"explicit","quote":"To balance these considerations, we experimented with different dimensionalities and fixed the latent dimension at 5 in all experiments.","support":"full"},{"assumption":"Training data consist of 60,000 IVSs on a fixed grid of log‑moneyness and maturity points.","type":"data","stated":"explicit","quote":"The training dataset consists of 60,000 IVSs, each represented on a fixed grid with log-moneyness m ∈ [−0.27, 0.27] and time to maturity τ ∈[0.1, 0.6], both discretized into 28 points.","support":"full"},{"assumption":"No‑arbitrage constraints are not enforced during training; violations may arise and are handled post‑generation.","type":"model","stated":"explicit","quote":"In addition, any generated surface must respect no-arbitrage constraints, in particular the absence of butterfly and calendar spread arbitrage. While our controllable VAE enforces feature-level structure, arbi-trage violations can still arise when latent variables are sampled in the tails of the normal distribution.","support":"full"}]},"datasets":{"status":"available","items":[{"name":"IVS dataset (combined Heston–SABR)","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"Additional details are provided in Appendix A. To assess dataset independence, we trained the model on both a Heston dataset and a combined Heston–SABR dataset. The results only for the combined dataset are presented in the paper.","support":"full"},{"name":"Heston IVS dataset","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"Additional details are provided in Appendix A. To assess dataset independence, we trained the model on both a Heston dataset and a combined Heston–SABR dataset.","support":"full"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"synthetic"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R3","candidates":{"front_matter":"n/a","evidence_card":"n/a","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"unclear","findings":[{"finding":"The controllable VAE achieves precise control over three features with maximum absolute errors of ~5.5×10⁻⁴ (level), 8×10⁻⁴ (slope) and 5.4×10⁻⁴ (term‑structure).","kind":"performance","quote":"The maximum absolute errors are approximately 5.5 × 10−4 for level, 8 × 10−4 for the slope over moneyness, and 5.4 × 10−4 for the term structure slope.","support":"full"},{"finding":"When controlling four features, maximum absolute errors increase but remain below 1×10⁻² (e.g., 5.7×10⁻⁴ level, 4.6×10⁻³ slope, 9.4×10⁻³ curvature, 2.0×10⁻³ term‑structure).","kind":"performance","quote":"The maximum absolute errors are approximately 5.7 × 10−4 for level, 4.6 × 10−3 for slope, 9.4 × 10−3 for curvature, and 2.0 × 10−3 for term-structure.","support":"full"},{"finding":"No arbitrage violations are observed when controllable variables are sampled within the convex hull and z is drawn from the central Gaussian region.","kind":"statistical","quote":"Specifically, when y lies within the convex hull of the dataset label space and z is drawn from the central region of N(0, 1), all 60,000 generated IVSs satisfy both the calendar-spread and butterfly conditions.","support":"full"},{"finding":"When sampling controllable variables from an extended domain and z from the full Gaussian prior, about 9% of generated IVSs violate calendar‑spread or butterfly arbitrage conditions.","kind":"statistical","quote":"When y is drawn from the extended hyper-rectangular domain while z is sampled from the full Gaussian prior, approximately 9% of the generated IVSs (5, 412 out of 60, 000) contains data points which violate either the calendar-spread condition, the butterfly condition, or both.","support":"full"},{"finding":"The post‑generation repair algorithm corrects roughly half of the arbitrage‑violating surfaces.","kind":"performance","quote":"For the remaining small proportion that violate no-arbitrage constraints, our post-generation repair algorithm successfully corrects approximately half.","support":"full"}],"reported_performance":null,"out_of_sample":{"method":"none","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":17,"citations":{"cited_by_in_corpus":1,"by_independent_authors":1,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":1,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":2,"extracted_at":"2026-11-14T19:48:17Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}