Probabilistic and statistical aspects of extremes of univariate processes have been extensively studied, and recent developments in extremes have focused on multivariate theory and its application. Multivariate extreme value theory encompasses two separate aspects: marginal features, which may be handled by standard univariate methods, and dependence features. Both will be examined in this study. First we focus on testing independence in multivariate extremes. All existing score tests of independence in multivariate extreme values have non-regular properties that arise due to violations of the usual regularity conditions of maximum likelihood. Some of these violations may be dealt with using standard techniques, for example when independence corresponds to a boundary point of the parameter space of the underlying model. However, another type of regularity violation, the infinite second moment of the score function, is more difficult to deal with and has important consequences for applications, resulting in score statistics with non-standard normalisation and poor rates of convergence. We propose a likelihood based approach that provides asymptotically normal score tests of independence with regular normalisation and rapid convergence. The resulting tests are straightforward to implement and are beneficial in practical situations with realistic amounts of data. A fundamental issue in applied multivariate extreme value (MEV) analysis is modelling dependence within joint tail regions. The primary aim of the remainder of this thesis is to develop a pseudo-polar framework for modelling extremal dependence that extends the existing classical results for multivariate extremes to encompass asymptotically independent tails. Accordingly, a constructional procedure for obtaining parametric asymptotically independent joint tail models is developed. The practical application of this framework is analysed through applications to bivariate simulated and environmental data, and joint estimation of dependence and marginal parameters via likelihood methodology is detailed. Inference under our models is examined and tests of extremal asymptotic independence and asymmetry are derived which are useful for model selection. In contrast to the classical MEV approach, which concentrates on the distribution of the normalised componentwise maxima, our framework is based on modelling joint tails and focuses directly on the tail structure of the joint survivor function. Consequently, this framework provides significant extensions of both the theoretical and applicable tools of joint tail modelling. Analogous point process theory is developed and the classical componentwise maxima result for multivariate extremes is extended to the asymptotically independent case. Finally, methods for simulating from two of our bivariate parametric models are provided.

La détection d'anomalies est tout d'abord une étape utile de pré-traitement des données pour entraîner un algorithme d'apprentissage statistique. C'est aussi une composante importante d'une grande variété d'applications concrètes, allant de la finance, de l'assurance à la biologie computationnelle en passant par la santé, les télécommunications ou les sciences environnementales. La détection d'anomalies est aussi de plus en plus utile au monde contemporain, où il est nécessaire de surveiller et de diagnostiquer un nombre croissant de systèmes autonomes. La recherche en détection d'anomalies inclut la création d'algorithmes efficaces accompagnée d'une étude théorique, mais pose aussi la question de l'évaluation de tels algorithmes, particulièrement lorsque l'on ne dispose pas de données labellisées -- comme dans une multitude de contextes industriels. En d'autres termes, l'élaboration du modèle et son étude théorique, mais aussi la sélection du modèle. Dans cette thèse, nous abordons ces deux aspects. Tout d'abord, nous introduisons un critère alternatif au critère masse-volume existant, pour mesurer les performances d'une fonction de score. Puis nous nous intéressons aux régions extrêmes, qui sont d'un intérêt particulier en détection d'anomalies, pour diminuer le taux de fausse alarme. Enfin, nous proposons deux méthodes heuristiques, l'une pour évaluer les performances d'algorithmes de détection d'anomalies en grande dimension, l'autre pour étendre l'usage des forets aléatoires à la classification à une classe. / Anomaly detection is not only a useful preprocessing step for training machine learning algorithms. It is also a crucial component of many real-world applications, from various fields like finance, insurance, telecommunication, computational biology, health or environmental sciences. Anomaly detection is also more and more relevant in the modern world, as an increasing number of autonomous systems need to be monitored and diagnosed. Important research areas in anomaly detection include the design of efficient algorithms and their theoretical study but also the evaluation of such algorithms, in particular when no labeled data is available -- as in lots of industrial setups. In other words, model design and study, and model selection. In this thesis, we focus on both of these aspects. We first propose a criterion for measuring the performance of any anomaly detection algorithm. Then we focus on extreme regions, which are of particular interest in anomaly detection, to obtain lower false alarm rates. Eventually, two heuristic methods are proposed, the first one to evaluate anomaly detection algorithms in the case of high dimensional data, the other to extend the use of random forests to the one-class setting.

In this thesis we will elaborate on multivariate extreme value modelling, re- lated practical and theoretical aspects. We will mainly focus on the dependence models, the extreme value copulas in particular. Extreme value copulas effec- tively unify the univariate extreme value theory and the copula framework itself in a single view. We familiarize ourselves with both of them in the first two chapters. Those chapters present generalized extreme value distribution, gen- eralized Pareto distribution and Archimedean copulas, that are suitable for the multivariate maxima and the threshold exceedances description. These two top- ics will be addressed in the third chapter in detail. Taking into consideration rather practical focus of this thesis, we examine the methods of data analysis extensively. Furthermore, we will employ these methods in a comprehensive case study, that will aim to reveal the importance of extreme value theory application in the Catastrophe Insurance. 1

In the present work we study multivariate extreme value theory. Our main focus is on exceedances over linear thresholds. Smaller part is devoted to exce- edances over elliptical thresholds. We consider extreme values as those which belong to remote regions and investigate convergence of their distribution to the limit distribution. The regions are either halfspaces or ellipsoids. Working with halfspaces we distinguish between two setups: we either assume that the distribution of extreme values is directionally homogeneous and we let the halfspaces diverge in any direction, or we assume that there are some irre- gularities in the sample cloud which show us the fixed direction we should let the halfspaces drift out. In the first case there are three limit laws. The domains of attraction contain unimodal and rotund-exponential distributions. In the second case there exist a lot of limit laws without general form. The domains of attraction also fail to have common structure. The similar situation occurs for the exceedances over elliptical thresholds. The task here is to investigate convergence of the random vectors living in the complements of ellipsoids. For all, the limit distributions are determined by affine transformations and distribution of spectral measure. 1

La prédiction de futurs évènements extrêmes est d’un grand intérêt dans de nombreux domaines tels que l’environnement ou la gestion des risques. Alors que la théorie des valeurs extrêmes univariées est bien connue, la complexité s’accroît lorsque l’on s’intéresse au comportement joint d’extrêmes de plusieurs variables. Un intérêt particulier est porté aux évènements de nature spatiale, définissant le cadre d’un nombre infini de dimensions. Sous l’hypothèse que ces évènements soient marginalement extrêmes, nous focalisons sur la structure de dépendance qui les lie. Dans un premier temps, nous faisons une revue des modèles paramétriques de dépendance dans le cadre multivarié et présentons différentes méthodes d’estimation. Les processus maxstables permettent l’extension au contexte spatial. Nous dérivons la loi en dimension finie du célèbre modèle de Brown- Resnick, permettant de faire de l’inférence par des méthodes de vraisemblance ou de vraisemblance composée. Nous utilisons ensuite des lois asymétriques afin de définir la représentation spectrale d’un modèle plus large : le modèle Extremal Skew-t, généralisant la plupart des modèles présents dans la littérature. Ce modèle a l’agréable propriété d’être asymétrique et non-stationnaire, deux notions présentées par les évènements environnementaux spatiaux. Ce dernier permet un large spectre de structures de dépendance. Les indicateurs de dépendance sont obtenus en utilisant la loi en dimension finie.Enfin, nous présentons une méthode d’estimation non-paramétrique par noyau pour les queues de distributions et l’appliquons à la sélection de modèles. Nous illustrons notre méthode à partir de l’exemple de modèles climatiques. / Projection of future extreme events is a major issue in a large number of areas including the environment and risk management. Although univariate extreme value theory is well understood, there is an increase in complexity when trying to understand the joint extreme behavior between two or more variables. Particular interest is given to events that are spatial by nature and which define the context of infinite dimensions. Under the assumption that events correspond marginally to univariate extremes, the main focus is then on the dependence structure that links them. First, we provide a review of parametric dependence models in the multivariate framework and illustrate different estimation strategies. The spatial extension of multivariate extremes is introduced through max-stable processes. We derive the finite-dimensional distribution of the widely used Brown-Resnick model which permits inference via full and composite likelihood methods. We then use Skew-symmetric distributions to develop a spectral representation of a wider max-stable model: the extremal Skew-t model from which most models available in the literature can be recovered. This model has the nice advantages of exhibiting skewness and nonstationarity, two properties often held by environmental spatial events. The latter enables a larger spectrum of dependence structures. Indicators of extremal dependence can be calculated using its finite-dimensional distribution. Finally, we introduce a kernel based non-parametric estimation procedure for univariate and multivariate tail density and apply it for model selection. Our method is illustrated by the example of selection of physical climate models.

This thesis analyzes market risk factors in commodity and currency markets. It focuses on the impact of extreme events on the prices of financial products traded in these markets, and on the overall market risk faced by the investors. The first chapter develops a simple two-factor jump-diffusion model for valuation of contingent claims on commodities in order to investigate the pricing implications of shocks that are exogenous to this market. The second chapter analyzes the nature and pricing implications of the abrupt changes in exchange rates, as well as the ability of these changes to explain the shapes of option-implied volatility "smiles". Finally, the third chapter employs the notion that key results of the univariate extreme value theory can be applied separately to the principal components of ARMA-GARCH residuals of a multivariate return series. The proposed approach yields more precise Value at Risk forecasts than conventional multivariate methods, while maintaining the same efficiency. / El objetivo de esta tesis es analizar los factores del riesgo del mercado de las materias primas y las divisas. Está centrada en el impacto de los eventos extremos tanto en los precios de los productos financieros como en el riesgo total de mercado al cual se enfrentan los inversores. En el primer capítulo se introduce un modelo simple de difusión y saltos (jump-diffusion) con dos factores para la valuación de activos contingentes sobre las materias primas, con el objetivo de investigar las implicaciones de shocks en los precios que son exógenos a este mercado. En el segundo capítulo se analiza la naturaleza e implicaciones para la valuación de los saltos en los tipos de cambio, así como la capacidad de éstos para explicar las formas de sonrisa en la volatilidad implicada. Por último, en el tercer capítulo se utiliza la idea de que los resultados principales de la Teoria de Valores Extremos univariada se pueden aplicar por separado a los componentes principales de los residuos de un modelo ARMA-GARCH de series multivariadas de retorno. El enfoque propuesto produce pronósticos de Value at Risk más precisos que los convencionales métodos multivariados, manteniendo la misma eficiencia.