{"paper_id":"2309.02994","title":"An Offline Learning Approach to Propagator Models","authors":["Eyal Neuman","Wolfgang Stockinger","Yufei Zhang"],"paper_date":"2023-09-06","paper_type":"theoretical","paper_url":"https://arxiv.org/pdf/2309.02994v1","page_url":"https://thequant.space/flowcharts/an-offline-learning-approach-to-propagator-models/","scores":{"empirical_rigor":3,"math_complexity":8.5,"quadrant":"Lab Rats"},"proposes":{"status":"available","statement":"The paper proposes an offline reinforcement learning approach with a pessimistic loss functional that accounts for uncertainty in the estimated price impact propagator and eliminates spurious correlation, yielding an asymptotically optimal bound on execution costs.","quote":"By adopting an offline reinforcement learning approach, we introduce a pessimistic loss functional taking the uncertainty of the estimated propagator into account, with an optimiser which eliminates the spurious correlation, and derive an asymptotically optimal bound on the execution costs even without precise information on the true propagator.","summary":"We consider an offline learning problem for an agent who first estimates an unknown price impact kernel from a static dataset, and then designs strategies to liquidate a risky asset while creating transient price impact. We propose a novel approach for a nonparametric estimation of the propagator…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The noise process ε(n) is conditionally sub‑Gaussian.","type":"statistical","stated":"explicit","quote":"In the following assumption we classify the noise in (2.11) as conditionally sub-Gaussian. Assumption 2.8. The agent has an oﬄine dataset D of size N ∈N as in Deﬁnition 2.4 and there exists a known constant R > 0 such that for all n = 1, . . . , N, EP′ h exp \u0010 ⟨v, ε(n)⟩ \u0011 | Gn−1 i ≤exp \u0012R2∥v∥2 2 \u0013 , for all v ∈RM.","support":"full"},{"assumption":"A static dataset D containing N price and signal trajectories and the corresponding metaorders exists.","type":"data","stated":"explicit","quote":"In the 2 following, we assume that there exists a dataset D with N price and signal trajectories, including the corresponding agents’ metaorders following the dynamics in (1.1), D = n (S(n), u(n), A(n)) | n = 1, . . . , N o , (1.2) see Deﬁnition 2.4 for details.","support":"full"},{"assumption":"The price impact kernel G (propagator) exists and determines execution prices via the linear relation in (1.1).","type":"model","stated":"explicit","quote":"Propagator models assume that there exists a matrix G = (Gi,j), known as the propagator or the price impact kernel, such that the asset’s execution price S is given by Sti+1 = St1 + i X j=1 Gi,jutj + Ati + εti, i = 1, . . . , M, (1.1) where the process A + ε represents the fundamental (or unaﬀected) asset price.","support":"full"},{"assumption":"The dataset may contain any price‑adaptive trading strategies u(n), not only a single deterministic strategy.","type":"data","stated":"explicit","quote":"(ii) We allow for any price adaptive strategies u(n) in the dataset D, in contrast to choosing only one deterministic strategy that will repeat itself (i.e., u(n) = u(m) for all 0 ≤m, n ≤N) as in Theorem 2.10 of [34] .","support":"full"},{"assumption":"Non‑convolution propagators are allowed; when the propagator is convolutional, faster convergence rates are obtained.","type":"model","stated":"explicit","quote":"(iii) We allow for estimation of non-convolution propagators G and we show that in the convolution case the convergence rate is substantially faster (see Remark 2.15).","support":"full"},{"assumption":"Synthetic data generated as described (fixed trading days, 5‑minute bins, specific signal and price dynamics) is representative for evaluating the estimators and strategies.","type":"other","stated":"explicit","quote":"3 Numerical Illustration In this section, we examine the performance of the propagator estimators in Section 2.2 and the pessimistic strategies presented in Section 2.4. We start by describing the construction of the synthetic dataset D for our experiments.","support":"full"}]},"datasets":{"status":"available","items":[{"name":"Synthetic dataset D","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"In the 2 following, we assume that there exists a dataset D with N price and signal trajectories, including the corresponding agents’ metaorders following the dynamics in (1.1), D = n (S(n), u(n), A(n)) | n = 1, . . . , N o , (1.2) see Deﬁnition 2.4 for details. 3 Numerical Illustration In this section, we examine the performance of the propagator estimators in Section 2.2 and the pessimistic strategies presented in","support":"full"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"synthetic"},"transaction_costs":{"status":"available","modeled":"yes","headline_net_of_costs":null,"detail":"Execution costs are modeled via the propagator kernel G and the cost functional; the greedy strategy based on the estimated propagator incurs suboptimal costs, while the pessimistic strategy reduces these costs.","quote":"We show that a trader who tries to minimise her execution costs by using a greedy strategy purely based on the estimated propagator will encounter suboptimality due to so-called spurious correlation between the trading strategy and the estimator and due to intrinsic uncertainty resulting from a biased cost functional. The pessimistic strategy takes the model uncertainty into account and achieves more stable","source":"evidence_card","rule":"R2","candidates":{"front_matter":"unclear","evidence_card":"yes","evidence_card_quoted":true},"card_detail":"Execution costs are modeled via the propagator kernel G and the cost functional; the greedy strategy based on the estimated propagator incurs suboptimal costs, while the pessimistic strategy reduces these costs.","card_quote":"We show that a trader who tries to minimise her execution costs by using a greedy strategy purely based on the estimated propagator will encounter suboptimality due to so-called spurious correlation between the trading strategy and the estimator and due to intrinsic uncertainty resulting from a biased cost functional. The pessimistic strategy takes the model uncertainty into account and achieves more stable","meaning":"costs are part of the reported headline results"},"results":{"status":"available","evaluation_type":"unclear","findings":[{"finding":"A pessimistic offline learning strategy eliminates spurious correlation and achieves an asymptotically optimal bound on execution costs even without precise knowledge of the true propagator.","kind":"theoretical","quote":"By adopting an offline reinforcement learning approach, we introduce a pessimistic loss functional taking the uncertainty of the estimated propagator into account, with an optimiser which eliminates the spurious correlation, and derive an asymptotically optimal bound on the execution costs even without precise information on the true propagator.","support":"full"},{"finding":"Numerical experiments demonstrate that the proposed non‑parametric propagator estimator and the pessimistic trading strategy are effective.","kind":"performance","quote":"Numerical experiments are included to demonstrate the effectiveness of the proposed propagator estimator and the pessimistic trading strategy. The pessimistic strategy takes the model uncertainty into account and achieves more stable performance and drastically reduces the execution costs.","support":"full"},{"finding":"Direct Volterra (non‑convolution) estimation yields large errors unless the dataset contains sufficiently noisy trading strategies.","kind":"statistical","quote":"Using a synthetic dataset, we illustrate the following characteristics of our methods: 17 • Directly estimating a Volterra propagator using (2.16) may result in large estimation errors unless the dataset contains suﬃciently noisy trading strategies. Our numerical results show that in the present setting, the Volterra estimator (2.18) yields large estimation errors, as the observed trading speeds in the dataset D are","support":"full"},{"finding":"Imposing a convolution structure and shape constraints dramatically improves estimation accuracy, achieving relative errors of order 10⁻³ with modest sample sizes.","kind":"statistical","quote":"By imposing a convolution structure and shape constraints of the estimated model, (2.25) signiﬁcantly improves the esti-mation accuracy, even with a smaller sample size. • Minimising the execution costs in (2.4) after substituting the estimated propagator instead of the true one, yields a greedy strategy that is very sensitive to the accuracy of the estimated model and also creates unfavorable transaction costs. In","support":"partial"},{"finding":"The pessimistic strategy, which accounts for model uncertainty, provides more stable performance and drastically reduces execution costs compared to a naive greedy strategy.","kind":"performance","quote":"The pessimistic strategy takes the model uncertainty into account and achieves more stable performance and drastically reduces the execution costs. Our numerical results show that the performance of a naive greedy strategy using the estimated model is very sensitive to the quality of the estimated model.","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":2,"by_independent_authors":2,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":2,"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":"2026-11-14T20:01:26Z"},"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/"}