{"paper_id":"2411.02947","title":"Time-Causal VAE: Robust Financial Time Series Generator","authors":["Beatrice Acciaio","Stephan Eckstein","Songyan Hou"],"paper_date":"2024-11-05","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2411.02947v1","page_url":"https://thequant.space/flowcharts/time-causal-vae-robust-financial-time-series-generator/","scores":{"empirical_rigor":7,"math_complexity":8.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes a time‑causal variational autoencoder (TC‑VAE) that generates financial time series with causal Wasserstein guarantees.","quote":"We build a time-causal variational autoencoder (TC-VAE) for robust generation of financial time series data.","summary":"We build a time-causal variational autoencoder (TC-VAE) for robust generation of financial time series data. Our approach imposes a causality constraint on the encoder and decoder networks, ensuring a causal transport from the real market time series to the fake generated time series. Specifically…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"All random variables are defined on a fixed probability space (Ω, F, P) and equalities hold P‑almost surely.","type":"statistical","stated":"explicit","quote":"All random variables are defined on a fixed probability space (Ω, F, P), and all equalities and inequalities are intended to hold in P-almost sure sense."},{"assumption":"The market data distribution µdata belongs to P1(R^dT) and the goal is to approximate it with a generated distribution µgen.","type":"market","stated":"explicit","quote":"The aim then is to build a generator that produces paths from a generated distribution µgen ∈P1(RdT ) which we want to be as close as possible to µdata."},{"assumption":"Causal Wasserstein distance is the appropriate metric for robustness of stochastic optimization problems.","type":"model","stated":"explicit","quote":"use a modified version of the classical Wasserstein distance, called causal Wasserstein distance, which we recall in the next subsection."},{"assumption":"The encoder and decoder are constrained to be causal maps.","type":"model","stated":"explicit","quote":"we impose causality constraints on the encoder and decoder maps, and build a generator connected to it"},{"assumption":"The prior distribution can be modeled by a RealNVP flow.","type":"model","stated":"explicit","quote":"In practice, we choose the sequence of invertible transformations from the class of real-valued non-volume-preserving (real NVP) transformations"},{"assumption":"Synthetic experiments use Black‑Scholes, Heston and Path‑Dependent‑Volatility models as ground truth data.","type":"other","stated":"explicit","quote":"On synthetic datasets like the Black‑Scholes model, Heston model and Path‑Dependent‑Volatility model, TC‑VAE learns the data distribution very well"}]},"datasets":{"status":"available","items":[{"name":"Black‑Scholes model","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"On synthetic datasets like the Black‑Scholes model, Heston model and Path‑Dependent‑Volatility model, TC‑VAE learns the data distribution very well"},{"name":"Heston model","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"On synthetic datasets like the Black‑Scholes model, Heston model and Path‑Dependent‑Volatility model, TC‑VAE learns the data distribution very well"},{"name":"Path‑Dependent‑Volatility model","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"On synthetic datasets like the Black‑Scholes model, Heston model and Path‑Dependent‑Volatility model, TC‑VAE learns the data distribution very well"},{"name":"S&P 500","provider":null,"kind":"market-prices","access":"public","period":null,"universe":null,"quote":"On real market datasets, such as S&P 500 and VIX, conditional TC‑VAE enables us to generate paths, as many as possible and as long as possible."},{"name":"VIX","provider":null,"kind":"market-prices","access":"public","period":null,"universe":null,"quote":"On real market datasets, such as S&P 500 and VIX, conditional TC‑VAE enables us to generate paths, as many as possible and as long as possible."}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"daily"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R1","candidates":{"front_matter":"n/a","evidence_card":"unclear","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"unclear","findings":[{"finding":"TC‑VAE generates synthetic paths that closely match real data in drift, volatility, marginal distribution and stylized facts such as heavy tails and volatility clustering.","kind":"statistical","quote":"... capturing key properties such as gain/loss asymmetry, skewness and kurtosis of returns, heavy-tail returns, no correlation in returns, short time correlation in square returns, long time correlation in absolute returns, and volatility clustering."},{"finding":"Mean‑variance efficient frontiers from fake paths match those from real paths, indicating similar portfolio performance.","kind":"performance","quote":"The efficient frontier of fake paths matches again extremely well to the one by real paths, see the right-hand side of Figure 6. This implies that the performance of mean‑variance portfolio optimization problem are close between real and fake paths."},{"finding":"Sliced adapted Wasserstein distances between real and fake Heston paths are low (SAW1 ≈ 0.549), showing closeness of distributions.","kind":"statistical","quote":"real-fake SAW1(µreal, µfake) 0.549 0.039"},{"finding":"TC‑VAE achieves promising results on both unconditional and conditional financial time series generation across several metrics and datasets.","kind":"performance","quote":"TC‑VAE achieves promising results on both unconditional and conditional financial time series generation on several different metrics and datasets"}],"reported_performance":null,"out_of_sample":{"method":"none","headline_out_of_sample":null}},"independent_evidence":{"status":"available","supports":[],"contradicts":[],"qualifies":[],"compared_with":19,"citations":{"cited_by_in_corpus":6,"by_independent_authors":5,"how_cited":{"used_as_baseline":0,"uses_method_or_data":2,"extends":1,"discusses_findings":0,"mentions":1,"unclassified":2},"independent_papers":[{"paper":{"paper_id":"2510.15458","title":"Robust Optimization in Causal Models and G-Causal Normalizing Flows","paper_date":"2025-10-17","authors":["Gabriele Visentin","Patrick Cheridito"],"empirical_rigor":6,"page_url":"https://thequant.space/flowcharts/robust-optimization-in-causal-models-and-g-causal-normalizin/"},"how":"extends","quote":null},{"paper":{"paper_id":"2407.19848","title":"Generative modelling of financial time series with structured noise and MMD-based signature learning","paper_date":"2024-07-29","authors":["Chung I Lu","Julian Sester"],"empirical_rigor":6,"page_url":"https://thequant.space/flowcharts/generative-modelling-of-financial-time-series-with-structured-noise-and-mmd/"},"how":"uses_method_or_data","quote":null},{"paper":{"paper_id":"2511.04469","title":"Towards Causal Market Simulators","paper_date":"2025-11-06","authors":["Dennis Thumm","Luis Ontaneda Mijares"],"empirical_rigor":6,"page_url":"https://thequant.space/flowcharts/towards-causal-market-simulators/"},"how":"uses_method_or_data","quote":null}],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":true,"code_url":"https://github.com/justinhou95/TimeCausalVAE","note":null},"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":1,"extracted_at":"2026-11-14T17:55:52Z"},"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/"}