Iranian Journal of Numerical Analysis and Optimization

Iranian Journal of Numerical Analysis and Optimization

Functional neural approximation and randomised quasi-Monte Carlo for path-dependent option pricing under Heston stochastic-local volatility

Document Type : Research Article

Authors
1 Department of Mathematics, Faculty of Sciences, Arak University, Arak, Iran
2 Faculty of Management and Accounting, Islamic Azad University, Qazvin Branch, Qazvin, Iran
10.22067/ijnao.2026.100768.1968
Abstract
We study a two-stage approximation for path-dependent option pricing under Heston stochastic-local volatility from irregular implied-volatility quotes. Monte Carlo and scrambled-Sobol randomised quasi-Monte Carlo (PCA-RQMC) with a discrete Brownian principal component construction generate labels; spline coefficients for the surface and average Heston variance-factor curve feed a neural surrogate. On 2,880 contracts from 240 held-out synthetic states, mean absolute ATM-vanilla-vega-scaled error is 0.21 volatility-point equivalent, 91% of errors are within 0.5 point, and RMSE is 11% below an interpolated-grid network. A matched dense network using the same spline coefficients performs similarly, so the supported contribution is the combined spline-input/PCA-RQMC pipeline rather than a distinct functional class. H2 is partially supported: PCA-RQMC reduces label variance by factors of 3.62–4.31 and supports a fourfold, but not the pre-specified eightfold, path reduction. The mean absolute 120-versus-480-step lookback difference is \(5.32\times10^{-3}\); errors therefore concern the discretised engine, not continuous-time SLV prices. Barrier results remain conditional on the approximate crossing correction. H1 is rejected: a single-date S&P 500/Apple illustration gives mean scaled errors of 0.37–0.42, but only 0–6% meet the near-ATM vanilla half-spread proxy. H3 is rejected as formulated because global weights and input-derivative attributions do not localise by contract family or barrier level. H4 is not supported as formulated: quote deletion is not a regime-change test, and severe failures occur for NVDA and semi-synthetic states far outside the training prior.
Keywords
Subjects

Send comment about this article
Enter Name.
Enter a valid email address.
Enter a vaid affiliation.
Enter comments (At leaset 10 words)
CAPTCHA Image
Enter Security Code Correctly.

Articles in Press, Accepted Manuscript
Available Online from 16 November 2026