Cette thèse est consacrée à l'étude théorique et numérique de deux problèmes non linéaires au sens de McKean en finance. Nous abordons dans la première partie le problème de calibration d'un modèle à volatilité locale et stochastique pour tenir compte des prix d'options Européennes vanilles observés sur le marché. Ce problème se traduit par l'étude d'une équation différentielle stochastique (EDS) non linéaire au sens de McKean à cause de la présence dans le coefficient de diffusion d'une espérance conditionnelle du facteur de volatilité stochastique par rapport à la solution de l'EDS. Nous obtenons l'existence du processus dans le cas particulier où le facteur de volatilité stochastique est un processus de sauts ayant un nombre fini d'états. Nous obtenons de plus la convergence faible à l'ordre 1 de la discrétisation en temps de l'EDS non linéaire au sens de McKean pour des facteurs de volatilité stochastique généraux. Dans l'industrie, la calibration est effectuée efficacement à l'aide d'une régularisation de l'espérance conditionnelle par un estimateur à noyau de type Nadaraya-Watson, comme proposé par Guyon et Henry-Labordère dans [JGPHL]. Nous proposons également un schéma numérique demi-pas de temps et étudions le système de particules associé que nous comparons à l'algorithme proposé par [JGPHL]. Dans la deuxième partie de la thèse, nous nous intéressons à un problème de valorisation de contrat avec appels de marge, une problématique apparue avec l'application de nouvelles régulations depuis la crise financière de 2008. Ce problème peut être modélisé par une équation différentielle stochastique rétrograde (EDSR) anticipative avec dépendance en la loi de la solution dans le générateur. Nous montrons que cette équation est bien posée et proposons une approximation de sa solution à l'aide d'EDSR standards linéaires lorsque la durée de liquidation de l'option en cas de défaut est petite. Enfin, nous montrons que le calcul des solutions de ces EDSR standards peut être amélioré à l'aide de la méthode de Monte-Carlo multiniveaux introduite par Giles dans [G] / This thesis is dedicated to the theoretical and numerical study of two problems which are nonlinear in the sense of McKean in finance. In the first part, we study the calibration of a local and stochastic volatility model taking into account the prices of European vanilla options observed in the market. This problem can be rewritten as a stochastic differential equation (SDE) nonlinear in the sense of McKean, due to the presence in the diffusion coefficient of a conditional expectation of the stochastic volatility factor computed w.r.t. the solution to the SDE. We obtain existence in the particular case where the stochastic volatility factor is a jump process with a finite number of states. Moreover, we obtain weak convergence at order 1 for the Euler scheme discretizing in time the SDE nonlinear in the sense of McKean for general stochastic volatility factors. In the industry, Guyon and Henry Labordere proposed in [JGPHL] an efficient calibration procedure which consists in approximating the conditional expectation using a kernel estimator such as the Nadaraya-Watson one. We also introduce a numerical half-step scheme and study the the associated particle system that we compare with the algorithm presented in [JGPHL]. In the second part of the thesis, we tackle the pricing of derivatives with initial margin requirements, a recent problem that appeared along with new regulation since the 2008 financial crisis. This problem can be modelled by an anticipative backward stochastic differential equation (BSDE) with dependence in the law of the solution in the driver. We show that the equation is well posed and propose an approximation of its solution by standard linear BSDEs when the liquidation duration in case of default is small. Finally, we show that the computation of the solutions to the standard BSDEs can be improved thanks to the multilevel Monte Carlo technique introduced by Giles in [G]

Epidemics have shaped, sometimes more than wars and natural disasters, demo- graphic aspects of human populations around the world, their health habits and their economies. Ebola and the Middle East Respiratory Syndrome (MERS) are clear and current examples of potential hazards at planetary scale. During the spread of an epidemic disease, there are phenomena, like the sudden extinction of the epidemic, that can not be captured by deterministic models. As a consequence, stochastic models have been proposed during the last decades. A typical forward problem in the stochastic setting could be the approximation of the expected number of infected individuals found in one month from now. On the other hand, a typical inverse problem could be, given a discretely observed set of epidemiological data, infer the transmission rate of the epidemic or its basic reproduction number. Markovian epidemic models are stochastic models belonging to a wide class of pure jump processes known as Stochastic Reaction Networks (SRNs), that are intended to describe the time evolution of interacting particle systems where one particle interacts with the others through a finite set of reaction channels. SRNs have been mainly developed to model biochemical reactions but they also have applications in neural networks, virus kinetics, and dynamics of social networks, among others. 4 This PhD thesis is focused on novel fast simulation algorithms and statistical inference methods for SRNs. Our novel Multi-level Monte Carlo (MLMC) hybrid simulation algorithms provide accurate estimates of expected values of a given observable of SRNs at a prescribed final time. They are designed to control the global approximation error up to a user-selected accuracy and up to a certain confidence level, and with near optimal computational work. We also present novel dual-weighted residual expansions for fast estimation of weak and strong errors arising from the MLMC methodology. Regarding the statistical inference aspect, we first mention an innovative multi- scale approach, where we introduce a deterministic systematic way of using up-scaled likelihoods for parameter estimation while the statistical fittings are done in the base model through the use of the Master Equation. In a di↵erent approach, we derive a new forward-reverse representation for simulating stochastic bridges between con- secutive observations. This allows us to use the well-known EM Algorithm to infer the reaction rates. The forward-reverse methodology is boosted by an initial phase where, using multi-scale approximation techniques, we provide initial values for the EM Algorithm.