{"paper_id":"2511.04469","title":"Towards Causal Market Simulators","authors":["Dennis Thumm","Luis Ontaneda Mijares"],"paper_date":"2025-11-06","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2511.04469v3","page_url":"https://thequant.space/flowcharts/towards-causal-market-simulators/","scores":{"empirical_rigor":6,"math_complexity":7.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes the Time-series Neural Causal Model VAE (TNCM-VAE), a framework that combines variational autoencoders with structural causal models to generate counterfactual financial time series while preserving temporal dependencies and causal relationships.","quote":"We propose a Time-series Neural Causal Model VAE (TNCM-VAE) that combines variational autoencoders with structural causal models to generate counterfactual financial time series while preserving both temporal dependencies and causal relationships.","summary":"Market generators using deep generative models have shown promise for synthetic financial data generation, but existing approaches lack causal reasoning capabilities essential for counterfactual analysis and risk assessment. We propose a Time-series Neural Causal Model VAE (TNCM-VAE) that combines…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The synthetic experiments use two autoregressive AR models inspired by the Ornstein–Uhlenbeck process, assuming stationary, mean-reverting dynamics.","type":"model","stated":"explicit","quote":"3 Experiments For the experiments, we used two autoregressive AR models, in-spired by the Ornstein–Uhlenbeck process [10]. This formulation Towards Causal Market Simulators ICAIF ’25, November 15–18, 2025, Singapore was chosen due to its well-known stationary properties, its mean-reverting behavior, and its common use in modeling financial time series [4].","support":"full"},{"assumption":"The decoder network enforces a directed acyclic graph (DAG) structure to encode causal relationships among variables.","type":"model","stated":"explicit","quote":"In the decoder network 𝑃𝜃(𝑉| 𝑈), we enforce a DAG structure that encodes the causal relationships among the variables.","support":"full"},{"assumption":"A separate latent space is encoded for each marginal stochastic process (e.g., X and Y).","type":"model","stated":"explicit","quote":"Following our construc-tion, a separate latent space is encoded for each marginal stochastic process (e.g., 𝑋and 𝑌).","support":"full"},{"assumption":"RealNVP transformations are used to model complex conditional distributions for each time step.","type":"model","stated":"explicit","quote":"When sampling from the inferred probability distribution 𝑃𝜃(𝑋| 𝑢), we employ the RealNVP [9] transformation to model complex conditional distributions.","support":"full"},{"assumption":"The prior distribution is time‑dependent, approximated separately at each time step.","type":"model","stated":"explicit","quote":"(2) For the prior, we approximate a distinct distribution at each time step, since each step corresponds to a different marginal in 𝑃(𝑋𝑡| 𝑋𝑡−1). As previously noted, each marginal conditional probability density function dif-fers across time steps, motivating the need for a time-dependent prior.","support":"full"}]},"datasets":{"status":"available","items":[{"name":"Synthetic AR data","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"In addition, working in a controlled setting grants us access to the ground truth. We generate synthetic data for   𝑋𝑡= 0.8𝑋𝑡−1 + 0.5𝜂𝑡 𝑌𝑡= 0.7𝑌𝑡−1 + 0.5𝑋𝑡−1 + 0.6𝜖𝑡 (3) where: 𝜂𝑡∼N (0, 1), 𝜖𝑡∼N (0, 1).","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":"For Experiment 1, the TNCM‑VAE achieves L1 distances between 0.04 and 0.09 across five time steps, with an average distance of 0.064.","kind":"predictive-accuracy","quote":"For Experiment 1, evaluating 𝑃(𝑌𝑡+1 > 0 | do(𝑋𝑡= 0)), our model achieves L1 distances ranging from 0.04 to 0.09 across the five time steps, with an average distance of 0.064.","support":"full"},{"finding":"For Experiment 2, the model attains L1 distances between 0.03 and 0.10, averaging 0.058, with accuracy improving over longer horizons.","kind":"predictive-accuracy","quote":"In Experiment 2, assessing 𝑃(𝑌𝑡+1 > 2 | do(𝑋𝑡= −2)), our ap-proach demonstrates even better performance with L1 distances between 0.03 and 0.10, achieving an average distance of 0.058.","support":"full"},{"finding":"The causal constraints enforced via DAG‑structured decoders and causal Wasserstein training lead to counterfactual generations that closely align with analytical ground truth.","kind":"theoretical","quote":"By enforcing causal constraints through DAG-structured decoders and leveraging causal Wasser-stein distances for training, our approach achieves superior coun-terfactual generation quality compared to existing methods.","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":22,"citations":{"cited_by_in_corpus":1,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":1},"independent_papers":[],"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/thummd/tncm","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":2,"extracted_at":"2026-11-14T19:46:26Z"},"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/"}