{"paper_id":"2505.15602","title":"Deep Learning for Continuous-time Stochastic Control with Jumps","authors":["Patrick Cheridito","Jean-Loup Dupret","Donatien Hainaut"],"paper_date":"2025-05-21","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2505.15602v2","page_url":"https://thequant.space/flowcharts/deep-learning-for-continuous-time-stochastic-control-with-ju/","scores":{"empirical_rigor":6.5,"math_complexity":9.2,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper introduces two iterative deep‑learning algorithms, GPI‑PINN and GPI‑CBU, for solving finite‑horizon continuous‑time stochastic control problems with jumps.","quote":"In this paper, we introduce two neural network-based numerical algorithms for such problems in high dimensions with finite time horizon and jumps. 6 Conclusions, limitations and future work In this paper, we have introduced two iterative deep learning algorithms for solving finite-horizon stochastic control problems with jumps.","summary":"In this paper, we introduce a model-based deep-learning approach to solve finite-horizon continuous-time stochastic control problems with jumps. We iteratively train two neural networks: one to represent the optimal policy and the other to approximate the value function. Leveraging a…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The controlled random measure is given by Nα(B×[0,t]) = ∑_{j=1}^{Mα} 1{Zj∈B} with a Cox process intensity λ(t,Xα_{t-},αt).","type":"model","stated":"explicit","quote":"We assume the controlled random measure is given by N α(B × [0, t]) = PM α t j=1 1{Zj∈B} for measurable subsets B ⊆E, where M α is a Cox process with a stochastic intensity of the form λ(t, Xα t−, αt) and Z1, Z2, . . . are i.i.d.","support":"full"},{"assumption":"The underlying state dynamics (drift, diffusion and jump coefficients) are known and can be used directly in the training objectives.","type":"model","stated":"implied","quote":"We iteratively train two neural networks: one to represent the optimal policy and the other to approximate the value function. In this paper, we introduce a deep model-based approach for stochastic control problems with jumps that takes the system dynamics (2) into account by leveraging the HJB equation (3).","support":"full"}]},"datasets":{"status":"available","items":[{"name":"10‑dimensional LQR problem without jumps","provider":null,"kind":"synthetic","access":"public","period":null,"universe":"10‑dimensional state space","quote":"Figure 1: Comparison of MAEV (blue) and runtime in seconds (green) of GPI-PINN (solid lines) and GPI-CBU (dashed lines) for a 10-dimensional LQR problem (18) without jumps (left) and with jumps (right).","support":"full"},{"name":"10‑dimensional LQR problem with jumps","provider":null,"kind":"synthetic","access":"public","period":null,"universe":"10‑dimensional state space","quote":"Figure 1: Comparison of MAEV (blue) and runtime in seconds (green) of GPI-PINN (solid lines) and GPI-CBU (dashed lines) for a 10-dimensional LQR problem (18) without jumps (left) and with jumps (right). Figure 3: log MAEV of different deep-learning methods for a 10-dimensional LQR problem with jumps.","support":"full"},{"name":"50‑dimensional LQR problem with jumps","provider":null,"kind":"synthetic","access":"public","period":null,"universe":"50‑dimensional state space","quote":"Figure 2 shows results of GPI-CBU for the same LQR problem with jumps in d = 50 dimensions. While the high dimensionality renders GPI-PINN infeasible, GPI-CBU achieves high accuracy in the approximation of the value function as well as the optimal policy.","support":"full"},{"name":"Consumption‑investment problem with 25 assets","provider":null,"kind":"synthetic","access":"public","period":null,"universe":"25 financial assets","quote":"Figure 4: Training losses c L (k) 1 (θ(k+1)) (left) and c L (k+1) 2 (ϕ(k+1)) (right) of GPI-CBU as functions of the epoch k for a consumption-investment problem of the form (20)–(21) with n = 25 financial assets.","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":"GPI‑CBU achieves numerical solutions virtually indistinguishable from high‑accuracy Runge–Kutta results.","kind":"performance","quote":"GPI-CBU yields numerical solutions that are virtually indistinguishable from the Runge–Kutta results.","support":"full"},{"finding":"GPI‑CBU outperforms model‑free RL methods PPO and SAC on the 10‑dimensional LQR problem with jumps.","kind":"performance","quote":"It can be seen that in this setup, PPO and SAC cannot compete with the two model-based approaches since they do not explicitly use the dynamics (19) but instead only rely on sampling from the environment.","support":"partial"},{"finding":"GPI‑CBU attains higher accuracy than the model‑based discrete‑time approach of Han & E (2016) due to avoiding discretization errors and better generalisation.","kind":"performance","quote":"The method of Han & E (2016) outperforms PPO and SAC but does not achieve the same accuracy as GPI-CBU due to discretization er-rors and since it does not generalize well to unseen points in the test set.","support":"full"},{"finding":"GPI‑PINN works well for high‑dimensional problems without jumps but becomes computationally infeasible when jumps are present.","kind":"negative","quote":"Our first algorithm, GPI-PINN, works well for high-dimensional problems without jumps but becomes computationally infeasible in the presence of jumps.","support":"full"},{"finding":"For the 50‑dimensional LQR problem with jumps, GPI‑CBU provides high‑accuracy approximations of both the value function and the optimal policy, whereas GPI‑PINN is infeasible.","kind":"performance","quote":"While the high dimensionality renders GPI-PINN infeasible, GPI-CBU achieves high accuracy in the approximation of the value function as well as the optimal policy.","support":"partial"}],"reported_performance":null,"out_of_sample":{"method":"none","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":27,"citations":{"cited_by_in_corpus":5,"by_independent_authors":3,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":1,"unclassified":4},"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/jdupret97/Deep-Learning-for-CT-Stochastic-Control-with-Jumps","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":"2026-11-14T19:50:10Z"},"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/"}