Keno Seefeldt: Multilevel Monte Carlo simulation for local stochastic volatility models via rough PDEs
Master thesis
Time: Fri 2026-11-13 15.00
Video link: https://tuwien.zoom.us/j/67904419111?pwd=uPg6e86bSJswgIlFSbSzcekGamklQm.1
Respondent: Keno Seefeldt
Abstract: We study the pricing methodology for local stochastic volatility (LSV) models that was introduced by [P. Bank, C. Bayer, P. K. Friz, L. Pelizzari, Mathematical Finance 35.3 (2025), pp. 661-681].
There, by a conditioning of the dynamics on the Brownian motion driving the volatility, a Markovian structure is restored and the conditional price can be expressed as the solution to a linear rough partial differential equation. Our main question is whether that methodology is suitable for multilevel Monte Carlo (MLMC) estimators.
Therefore, we study a numerical scheme introduced by [P. Bank, C. Bayer, P. K. Friz, L. Pelizzari, Mathematical Finance 35.3 (2025), pp. 661-681] that uses piecewise linear drivers to approximate the rough driver as well as two alternatives to it, that are inspired by different methods of adding Itô--Stratonovich corrections.
We measure weak errors, level variances and costs of the three schemes in the Bachelier and SABR LSV models with a rough Bergomi volatility backbone.
For each scheme the MLMC estimator reaches a given error faster than the corresponding plain Monte Carlo estimator, and the original scheme introduced by [P. Bank, C. Bayer, P. K. Friz, L. Pelizzari, Mathematical Finance 35.3 (2025), pp. 661-681] shows the strongest variance decay.
An MLMC estimator built from Euler steps of the unconditioned dynamics shows no comparable speed-up, which indicates that the novel methodology of [P. Bank, C. Bayer, P. K. Friz, L. Pelizzari, Mathematical Finance 35.3 (2025), pp. 661-681] via conditioning is what makes the multilevel approach effective.
