{"paper_id":"2505.22836","title":"Model-Free Deep Hedging with Transaction Costs and Light Data Requirements","authors":["Pierre Brugière","Gabriel Turinici"],"paper_date":"2025-05-28","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2505.22836v1","page_url":"https://thequant.space/flowcharts/model-free-deep-hedging-with-transaction-costs-and-light-dat/","scores":{"empirical_rigor":7,"math_complexity":8.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes training a neural network on as few as 256 simulated trajectories to obtain a hedging strategy that outperforms the Black‑Scholes delta hedge and the Leland model under transaction costs.","quote":"In this work, we show that using as few as 256 trajectories is sufficient to train a neural network that significantly outperforms, in the Geometric Brownian Motion framework, both the classical Black & Scholes formula and the Leland model","summary":"Option pricing theory, such as the Black and Scholes (1973) model, provides an explicit solution to construct a strategy that perfectly hedges an option in a continuous-time setting. In practice, however, trading occurs in discrete time and often involves transaction costs, making the direct…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The underlying asset follows a geometric Brownian motion (GBM) process.","type":"model","stated":"explicit","quote":"In the first case, the asset follows a geometric Brownian motion (GBM) process, and the instantaneous interest rate rs is stochastic."},{"assumption":"Transaction costs are proportional to the dollar amount traded.","type":"market","stated":"explicit","quote":"We assume that the trading costs are proportional to the value traded."},{"assumption":"Hedging is performed at discrete time points with no costs at inception or maturity.","type":"execution","stated":"explicit","quote":"Trading costs are assumed to apply, neither at inception, when the trader exchanges the delta with the option buyer, nor at maturity for the stocks which are delivered against the strike"},{"assumption":"The risk‑free rate is taken as constant (4.5% in the S&P 500 experiment).","type":"market","stated":"explicit","quote":"We assume a constant instantaneous interest rate of 4.5%, which is the average Secured Overnight"},{"assumption":"Neural network is trained on data simulated under the real‑world probability measure (drift not zero).","type":"data","stated":"explicit","quote":"In our experiments, it was not necessary to simulate trajectories under a risk‑neutral probability; instead, we simulated under a probability measure where the drift does not ne…"},{"assumption":"The hedging objective is to minimize the standard deviation of discounted terminal wealth.","type":"statistical","stated":"explicit","quote":"we choose to minimize the standard deviation of discounted terminal wealth ZT"}]},"datasets":{"status":"available","items":[{"name":"Simulated GBM trajectories","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"In our sample of 256 trajectories simulated with σ = 20%, the estimated values of σ computed with 30 observation points vary between 12.51% and 26.89%."},{"name":"Overlapping sequences from a single simulated series","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"a single sequence of length 285 is simulated, and 256 overlapping sequences of length 30 derived from it are used to train the neural network for hedging"},{"name":"S&P 500 Total Return Index","provider":"S&P Dow Jones Indices","kind":"market-prices","access":"public","period":"01/01/2024 to 25/02/2025","universe":"S&P 500 Total Return Index","quote":"Here, a single sequence of length 286 is constructed from the closing prices of the S&P 500 Total Return Index over the period from 01/01/2024 to 25/02/2025."}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"daily"},"transaction_costs":{"status":"available","modeled":"yes","headline_net_of_costs":null,"detail":"Proportional cost α per dollar traded, applied only when deltas are readjusted at discrete times.","quote":"the transaction costs of α per dollar traded arises only when the deltas are readjusted at times ti+1","source":"evidence_card","rule":"R2","candidates":{"front_matter":"unclear","evidence_card":"yes","evidence_card_quoted":true},"card_detail":"Proportional cost α per dollar traded, applied only when deltas are readjusted at discrete times.","card_quote":"the transaction costs of α per dollar traded arises only when the deltas are readjusted at times ti+1","meaning":"costs are part of the reported headline results"},"results":{"status":"available","evaluation_type":"pricing-calibration","findings":[{"finding":"A neural network trained on 256 GBM trajectories outperforms both the Black‑Scholes delta hedge and the Leland model under proportional transaction costs.","kind":"performance","quote":"using as few as 256 trajectories is sufficient to train a neural network that significantly outperforms, in the Geometric Brownian Motion framework, both the classical Black & Scholes formula and the Leland model"},{"finding":"The network further improves on the Leland approach for transaction costs at 0.5% and above despite using only 256×90 observation points.","kind":"performance","quote":"network further improves the Leland approach for transaction costs at 0.5% and above, despite the small number of observation points (256 × 90) used to train it."},{"finding":"Training on overlapping sequences from a single simulated series yields hedging accuracy comparable to training on independent sequences.","kind":"performance","quote":"The calibration speed of the neural network is comparable to that achieved when training on independent sequences of length 30"},{"finding":"Overlapping sequences extracted from real S&P 500 data can be used to train the network, suggesting feasibility on real‑world datasets.","kind":"performance","quote":"256 sequences of 31 data points are extracted from this dataset and used to train the neural network to price options with a maturity of T ="},{"finding":"The method does not require simulation under a risk‑neutral measure, simplifying data requirements.","kind":"theoretical","quote":"In our experiments, it was not necessary to simulate trajectories under a risk‑neutral probability; instead, we simulated under a probability measure where the drift does not ne…"}],"reported_performance":null,"out_of_sample":{"method":"holdout","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":45,"citations":{"cited_by_in_corpus":0,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"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":1,"extracted_at":"2026-11-14T17:36:59Z"},"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/"}