{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,14]],"date-time":"2026-10-14T03:44:45Z","timestamp":1791949485695,"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:\/\/www.uspto.gov\/ip-policy\/copyright-policy\/copyright-basics"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"abstract":"<jats:p>Continuous-time reinforcement learning (RL) is a promising approach to dynamic portfolio selection when market coefficients are unknown or when standard dynamic programming becomes difficult. Its practical use, however, is limited by noisy policy-gradient estimators, especially in stochastic-volatility and path-dependent settings. This paper develops a Malliavin-informed policy-gradient (MIPG) estimator that combines a discrete Gaussian integration-by-parts identity with learned, zero-mean stochastic control variates. The control variate is parameterized by either a linear model, a multilayer perceptron (MLP), or a recurrent gated recurrent unit (GRU), with predictability enforced by restricting inputs to information available before each Brownian innovation. We establish unbiasedness of the discrete Malliavin estimator under the simulation design and show that subtracting any square-integrable predictable stochastic integral preserves the target mean. We then evaluate the method in three controlled Monte Carlo experiments: Black-Scholes, Heston stochastic volatility, and rough Bergomi volatility. Across 25,000 out-of-sample paths per model, the GRU control variate reduces raw Malliavin variance by factors of 7.18, 3.51, and 3.33, respectively, with bootstrap intervals excluding one. A paired squared-error test strongly rejects equality with the unadjusted Malliavin estimator. In a projected stochastic-gradient optimization experiment, MIPG-GRU attains the smallest mean absolute parameter error under rough Bergomi among the four gradient estimators compared. At the same time, ordinary pathwise differentiation remains more efficient in the present smooth log-wealth problem. The evidence therefore supports MIPG as a variance-reduction mechanism for Malliavin gradients, not as a universal replacement for pathwise differentiation. The results motivate future work on non-smooth objectives, downside-risk constraints, transaction costs, and partial-information control, where Malliavin integration by parts may provide a larger comparative advantage.<\/jats:p>","DOI":"10.2139\/ssrn.7598820","type":"posted-content","created":{"date-parts":[[2026,10,14]],"date-time":"2026-10-14T02:46:41Z","timestamp":1791946001000},"source":"Crossref","is-referenced-by-count":0,"title":["Malliavin-Informed Policy Gradients for Continuous-Time Portfolio Optimization&amp;nbsp;Neural Variance Reduction under Black-Scholes, Heston, and Rough Bergomi Dynamics"],"prefix":"10.2139","author":[{"ORCID":"https:\/\/orcid.org\/0009-0002-4919-9201","authenticated-orcid":true,"given":"Abdou","family":"Kaba","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"78","container-title":[],"original-title":[],"deposited":{"date-parts":[[2026,10,14]],"date-time":"2026-10-14T02:46:41Z","timestamp":1791946001000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.ssrn.com\/abstract=7598820"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":0,"URL":"https:\/\/doi.org\/10.2139\/ssrn.7598820","relation":{},"subject":[],"published":{"date-parts":[[2026]]},"subtype":"preprint"}}