{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,8]],"date-time":"2026-10-08T19:21:02Z","timestamp":1791487262640,"version":"4.3.4"},"posted":{"date-parts":[[2026]]},"group-title":"SSRN","reference-count":0,"publisher":"Elsevier BV","license":[{"start":{"date-parts":[[2026,1,1]],"date-time":"2026-01-01T00:00:00Z","timestamp":1767225600000},"content-version":"unspecified","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"abstract":"<div>\n We introduce Delta-Guided Soft Actor-Critic (DG-SAC), a reward-shaping\n<\/div>\n<div>\n architecture that embeds the Black-Scholes delta as a time-decaying inductive\n<\/div>\n<div>\n bias into the Soft Actor-Critic reinforcement learning framework for derivative\n<\/div>\n<div>\n hedging. The shaped reward takes the form rDG(st, at) = r(st, at) \u2212 \u03b2(t) (at \u2212\n<\/div>\n<div>\n \u2206BS\n<\/div>\n<div>\n t )2, where \u03b2(t) = \u03b20(1 \u2212 \u03b7 t\/T ) decays from strong early guidance to weak\n<\/div>\n<div>\n late-stage regularisation. We prove that (i) DG-SAC recovers standard SAC\n<\/div>\n<div>\n performance asymptotically, with a guidance bias of O(\u03b20T ) that shrinks to\n<\/div>\n<div>\n zero as the optimal policy aligns with the Black-Scholes delta (Theorem 4),\n<\/div>\n<div>\n and (ii) a finite-time regret decomposition (Theorem 7) whose guidance-\n<\/div>\n<div>\n bias term is governed by the decay schedule: sublinear O(\u221aN log N ) regret\n<\/div>\n<div>\n requires annealing the guidance across training, whereas a fixed schedule\n<\/div>\n<div>\n contributes a small linear term that manifests as faster early convergence\n<\/div>\n<div>\n rather than an improved asymptotic rate.\n<\/div>\n<div>\n Empirically, DG-SAC reduces hedging error by 19-38% relative to SAC\n<\/div>\n<div>\n across geometric Brownian motion, Merton jump-diffusion, and rough Bergomi\n<\/div>\n<div>\n environments, with all nine single-seed comparisons significant at p &lt; 0.001.\n<\/div>\n<div>\n Multi-seed robustness analysis (five independent seeds) confirms advantages\n<\/div>\n<div>\n in all 15 seed-environment pairs (p = 0.005-0.062), with a 2\u00d7 convergence\n<\/div>\n<div>\n speedup. On real options data (SPY, QQQ, IWM), fine-tuned DG-SAC\n<\/div>\n<div>\n achieves hedging error reductions of 26-28% on liquid tickers. We transparently\n<\/div>\n<div>\n report that delta hedging outperforms all reinforcement learning agents when\n<\/div>\n<div>\n transaction costs are below 0.3% and model specification is correct, situating\n<\/div>\n<div>\n the practical value of DG-SAC in realistic-friction regimes (transaction costs\n<\/div>\n<div>\n \u2265 0.5%) where reinforcement learning agents learn endogenous rebalancing\n<\/div>\n<div>\n frequency.\n<\/div><\/jats:p>","DOI":"10.2139\/ssrn.7585078","type":"posted-content","created":{"date-parts":[[2026,10,8]],"date-time":"2026-10-08T18:45:41Z","timestamp":1791485141000},"source":"Crossref","is-referenced-by-count":0,"title":["Delta-Guided Soft Actor-Critic for Derivative Hedgingunder Model Misspecification and Transaction Costs"],"prefix":"10.2139","author":[{"ORCID":"https:\/\/orcid.org\/0009-0008-9343-2895","authenticated-orcid":true,"given":"Sparsh","family":"Patel","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"78","container-title":[],"original-title":[],"deposited":{"date-parts":[[2026,10,8]],"date-time":"2026-10-08T18:45:41Z","timestamp":1791485141000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.ssrn.com\/abstract=7585078"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":0,"URL":"https:\/\/doi.org\/10.2139\/ssrn.7585078","relation":{},"subject":[],"published":{"date-parts":[[2026]]},"subtype":"preprint"}}