Multilevel Monte Carlo for exponential Lévy models (Q2412390)

The authors study pricing of contingent claims under exponential Lévy processes by means of Monte Carlo simulation. A bound of \(\mathcal{O}\left(\varepsilon\right)\) on the root mean square error of an approximation can require time steps of size \(\mathcal{O}\left(\varepsilon^{-1}\right)\) and \(\mathcal{O}\left(\varepsilon^{-2}\right)\) many sample paths leading to a complexity of \(\mathcal{O}\left(\varepsilon^{-3}\right)\). For example, [\textit{E. Dia} and \textit{D. Lamberton}, Adv. Appl. Probab. 43, 1136--1165, (2011; Zbl 1235.60049)] treats the valuation of lookback options under the variance gamma process. The multilevel Monte Carlo approach, surveyed in [the author, Acta Numer. 24, 259--328, (2015; Zbl 1316.65010)], can reduce the complexity to \(\mathcal{O}\left(\varepsilon^{-2}\right)\) for Brownian motion. The approach splits the estimation into a sum of differences between estimators from successive refinements-leveraging the nearness of those refinements to limit the number of sample paths. Using the multilevel Monte Carlo method, the authors can treat exponential Lévy processes with small jumps obeying a power law and large jumps obeying an exponential law. The restriction comes from a key lemma that bounds the estimation error of the maximum of a discretely sampled process. Variance gamma, normal inverse Gaussian and \(\alpha\)-stable processes are among the class of exponential Lévy processes satisfying the restriction. The authors present numerical findings for the valuation of lookback, barrier and Asian options under these processes.