{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,30]],"date-time":"2026-10-30T08:30:44Z","timestamp":1793349044167,"version":"4.5.0"},"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>Generative models of option surfaces now construct risk-neutral measures by projecting a learned physical generator onto option prices. A premium recovered in this way can be produced by the surface representation, by the finite sample of the projection or by the generator itself. We build an audit that separates these sources. It combines a discretized, completely affine Heston oracle with a known pathwise pricing kernel, independent Monte Carlo clouds for fitting, normalizing and auditing the projection, a sequential accounting of the squared premium error across identification, representation, systematic generator, training and Monte Carlo layers, and a test of the implied oneday pricing kernel on realized returns of SPX option contracts excluded from the projection book, reported together with its power. Representation contamination is large in entropy but confined to nearly hedged directions, and the finite-sample entropy floor \ud835\udc5f\/(2\ud835\udc41) holds from 5,000 draws per cloud. In this benchmark the binding error is the generator's conditional mean of the one-day return. A conditional latent diffusion that beats an AR(1) forecast misses a true premium of about 1.6 bp by 20 to 29 bp on every training seed, while a DYSANOS-style generator that forecasts worse comes within 2 bp of it once Monte Carlo noise is removed. Re-centring the diffusion's own draws on a one-parameter return drift lowers its squared premium error by a factor of 15 to 34 and recovers the level of the premium but not its daily variation. On SPX, in exploratory analyses of a test period that an earlier run had already used, no learned kernel satisfies the Stage I joint diagnostic criteria fixed before the final exploratory run, and with all quotes included only the tests in the calm volatility regime have the power to reject a kernel at the size of the observed option premia.<\/jats:p>","DOI":"10.2139\/ssrn.7672039","type":"posted-content","created":{"date-parts":[[2026,10,30]],"date-time":"2026-10-30T07:36:03Z","timestamp":1793345763000},"source":"Crossref","is-referenced-by-count":0,"title":["Where Does the Risk Premium Come From: Auditing Generative Risk-Neutral Option Models"],"prefix":"10.2139","author":[{"given":"Yichai","family":"Huang","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Yujie","family":"Zuo","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"78","container-title":[],"original-title":[],"deposited":{"date-parts":[[2026,10,30]],"date-time":"2026-10-30T07:36:03Z","timestamp":1793345763000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.ssrn.com\/abstract=7672039"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":0,"URL":"https:\/\/doi.org\/10.2139\/ssrn.7672039","relation":{},"subject":[],"published":{"date-parts":[[2026]]},"subtype":"preprint"}}