{"paper_id":"2402.13326","title":"Deep Hedging with Market Impact","authors":["Andrei Neagu","Frédéric Godin","Clarence Simard","Leila Kosseim"],"paper_date":"2024-02-20","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2402.13326v2","page_url":"https://thequant.space/flowcharts/deep-hedging-with-market-impact/","scores":{"empirical_rigor":4,"math_complexity":8,"quadrant":"Lab Rats"},"proposes":{"status":"available","statement":"The paper proposes a novel deep reinforcement learning model for dynamic option hedging that incorporates convex market impacts and impact persistence.","quote":"In this paper, we propose a novel general market impact dynamic hedging model based on Deep Reinforcement Learning (DRL) that considers several realistic features such as convex market impacts, and impact persistence through time.","summary":"Dynamic hedging is the practice of periodically transacting financial instruments to offset the risk caused by an investment or a liability. Dynamic hedging optimization can be framed as a sequential decision problem; thus, Reinforcement Learning (RL) models were recently proposed to tackle this…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"Price dynamics follow a geometric Brownian motion with constant drift µ and volatility σ.","type":"model","stated":"explicit","quote":"we used the geometric Brownian motion (GBM) to simulate the underlying asset prices St at each time step t: St = St−1 exp (µ −σ2/2) δt + σ"},{"assumption":"Market impact is convex and can be parameterized by illiquidity parameters α and β.","type":"market","stated":"explicit","quote":"The following three sets of market impact parameters are considered: α = β = 1 (infinite liquidity); α = 1.01, β = 0.99 (high liquidity); α = 1.02, β = 0.98 (low liquidity)."},{"assumption":"Impact persistence is modeled by parameters λa and λb; the no‑persistence case sets λa = λb = ∞.","type":"market","stated":"explicit","quote":"The first set of analyses assumes no impact persistence (i.e. λa = λb = ∞), which means that market transactions have only an immediate impact."},{"assumption":"Risk‑free rate is set to zero for simplicity.","type":"model","stated":"explicit","quote":"For simplicity, the risk‑free rate is set to r = 0."},{"assumption":"The hedging portfolio value Vt and the underlying drift µ are included as state variables.","type":"execution","stated":"explicit","quote":"The hedging portfolio value Vt and the underlying asset drift (i.e. expected return) parameter µ materially impact optimal positions."},{"assumption":"Training uses a policy‑gradient algorithm with Adam‑optimized learning rates.","type":"statistical","stated":"explicit","quote":"the learning rates sequence {ηj}j≥1 contains small positive real numbers determined by the Adam optimizer [26]"}]},"datasets":{"status":"available","items":[{"name":"Synthetic GBM price series","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"we used the geometric Brownian motion (GBM) to simulate the underlying asset prices St at each time step t"}],"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":"A proportional transaction cost rate k is calibrated so that buying one share at initiation with fees k matches the DRL model revenue.","quote":"In all experiments, the parameter k is set so that the revenue of buying one share of the underlying asset at the initiation date with the proportional fees k is equal to that of the DRL model F b(0, 0, 1).","source":"front_matter","rule":"R3","candidates":{"front_matter":"yes","evidence_card":"yes","evidence_card_quoted":true},"card_detail":"A proportional transaction cost rate k is calibrated so that buying one share at initiation with fees k matches the DRL model revenue.","card_quote":"In all experiments, the parameter k is set so that the revenue of buying one share of the underlying asset at the initiation date with the proportional fees k is equal to that of the DRL model F b(0, 0, 1).","meaning":"costs are part of the reported headline results"},"results":{"status":"available","evaluation_type":"unclear","findings":[{"finding":"The DRL model outperforms delta‑hedging baselines in low‑liquidity settings by learning to dampen or delay rebalancing actions.","kind":"performance","quote":"Results show our DRL model behaves better in contexts of low liquidity by, among others: 1) learning the extent to which portfolio rebalancing actions should be dampened or delayed to avoid high costs"},{"finding":"Optimal positions depend on the hedging portfolio value Vt, unlike delta‑hedging which ignores Vt.","kind":"theoretical","quote":"our optimal DRL strategy depends on the hedging portfolio value at time t (Vt) whereas both delta hedging strategies ignore Vt in their decisions"},{"finding":"The DRL policy implicitly incorporates the underlying drift µ, whereas traditional delta hedging does not.","kind":"theoretical","quote":"our DRL model’s optimal hedging positions do implicitly take into account the drift"},{"finding":"Large illiquidity parameters α,β encourage delaying corrections, but large impact‑persistence parameters λa,λb counteract this effect.","kind":"theoretical","quote":"while large values of illiquidity parameters α, β incentivize partially delaying the correction of inadequate hedging positions"},{"finding":"When impact persistence is absent, market impacts vanish immediately after a trade (At = Bt = 0).","kind":"statistical","quote":"The absence of impact persistence is reflected by setting λa = λb = ∞ and thus At = Bt = 0 for all time"}],"reported_performance":null,"out_of_sample":{"method":"none","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":24,"citations":{"cited_by_in_corpus":7,"by_independent_authors":2,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":2,"unclassified":5},"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":"2026-11-14T18:14:28Z"},"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/"}