In the present work we deal with the problem of etremal value of time series, especially of maxima. We study times and values of maximum by an approach of point process and we model distribution of extremal values by statistical methods. We estimate parameters of distribution using different methods, namely graphical methods of data analysis and subsequently we test the estimated distribution by tests of goodness of fit. We study the stationary case and also the cases with a trend. In connection with distribution of excesess and exceedances over a threshold we deal with generalized Pareto distribution.

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

This thesis is focused on the models based on extreme value theory and their practical applications. Specifically are described the block maxima models and the models based on threshold exceedances. Both of these methods are described in thesis theoretically. Apart from theoretical description there are also practical calculations based on simulated or real data. The applications of block maxima models are focused on choice of block size, suitability of the models for specific data and possibilities of extreme data analysis. The applications of models based on threshold exceedances are focused on choice of threshold and on suitability of the models. There is an example of the model used for calculations of reinsurance premium for extreme claims in the case of nonproportional reinsurance.

Dans les systèmes informatiques temps réel, les tâches logicielles sont contraintes par le temps. Pour garantir la sûreté du système critique contrôlé par le système temps réel, il est primordial d'estimer de manière sûre le pire temps d'exécution de chaque tâche. Les performances des processeurs actuels du commerce permettent de réduire en moyenne le temps d'exécution des tâches, mais la complexité des composants d'optimisation de la plateforme rendent difficile l'estimation du pire temps d'exécution. Il existe différentes approches d'estimation du pire temps d'exécution, souvent ségréguées et difficilement généralisables ou au prix de modèles coûteux. Les approches probabilistes basées mesures existantes sont vues comme étant rapides et simples à mettre en œuvre, mais souffrent d'un manque de systématisme et de confiance dans les estimations qu'elles fournissent. Les travaux de cette thèse étudient les conditions d'application de la théorie des valeurs extrêmes à une suite de mesures de temps d'exécution pour l'estimation du pire temps d'exécution probabiliste, et ont été implémentées dans l'outil diagxtrm. Les capacités et les limites de l'outil ont été étudiées grâce à diverses suites de mesures issues de systèmes temps réel différents. Enfin, des méthodes sont proposées pour déterminer les conditions de mesure propices à l'application de la théorie des valeurs extrêmes et donner davantage de confiance dans les estimations. / Software tasks are time constrained in real time computing systems. To ensure the safety of the critical systems that embeds the real time system, it is of paramount importance to safely estimate the worst-case execution time of each task. Modern commercial processors optimisation components enable to reduce in average the task execution time at the cost of a hard to determine task worst-case execution time. Many approaches for executing a task worst-case execution time exist but are usually segregated and hardly scalable, or by building very complex models. Measurement-based probabilistic timing analysis approaches are said to be easy and fast, but they suffer from a lack of systematism and confidence in their estimates. This thesis studies the applicability of the extreme value theory to a sequence of execution time measurements for the estimation of the probabilistic worst-case execution time, leading to the development of the diagxtrm tool. Thanks to a large panel of sequences of measurements from different real time systems, capabilities and limits of the tool are enlightened. Finally, a couple of methods are provided for determining measurements conditions that foster the application of the theory and raise more confidence in the estimates.

Une interface est une zone de l'espace qui sépare deux régions possédant des propriétés physiques différentes. La plupart des interfaces de la nature résultent d'un processus de croissance, mêlant une composante aléatoire et une dynamique déterministe régie par les symétries du problème. Le résultat du processus de croissance est un objet présentant des corrélations à longue portée. Dans cette thèse, nous nous proposons d'étudier la statistique d'extrême de différents types d'interfaces. Une première motivation est de raffiner la compréhension géométrique de tels objets, via leur maximum. Une seconde motivation s'inscrit dans la démarche plus générale de la statistique d'extrême de variables aléatoires fortement corrélées. A l'aide de méthodes analytiques d'intégrales de chemin nous analysons la distribution du maximum d'interfaces à l'équilibre, dont l'énergie es t purement élastique à courte portée. Nous attaquons ensuite le problème d'interfaces élastiques en milieu désordonné, principalement à l'aide de simulations numériques. Enfin nous étudierons une interface hors-équilibre dans son régime de croissance. L'équivalence de ce type d'interface avec le polymère dirigé en milieu aléatoire, un des paradigmes de la physique statistique des systèmes désordonnés, donne une portée étendue aux résultats concernant la statistique du maximum de l'interface. Nous exposerons les résultats que nous avons obtenus sur un modèle de mouvements browniens qui ne se croisent pas, tout en explicitant le lien entre ce modèle, l'interface en croissance et le polymère dirigé. / An interface is an area of space that separates two regions having different physical properties. Most interfaces in nature are the result of a growth process, mixing a random behavior and a deterministic dynamic derived from the symmetries of the problem. This growth process gives an object with extended correlations. In this thesis, we focus on the study of the extremum of different kinds of interfaces. A first motivation is to refine the geometric properties of such objects, looking at their maximum. A second motivation is to explore the extreme value statistics of strongly correlated random variables. Using path integral techniques we analyse the probability distribution of the maximum of equilibrium interfaces, possessing short range elastic energy. We then extend this to elastic interfaces in random media, with essentially numerical simulations. Finally we study a particular type of out-of-equilibrium interface, in its growing regime. Such interface is equivalent to the directed polymer in random media, a paradigm of the statistical mechanics of disordered systems. This equivalence reinforces the interest in the extreme value statistics of the interface. We will show the exact results we obtained for a non-intersecting Brownian motion model, explaining precisely the link with the growing interface and the directed polymer.

A teoria do valor extremo é aplicada em áreas de pesquisa tais como hidrologia, estudos de poluição, engenharia de materiais, controle de tráfego e economia. A distribuição valor extremo ou Gumbel é amplamente utilizada na modelagem de valores extremos de fenômenos da natureza e no contexto de análise de sobrevivência para modelar o logaritmo do tempo de vida. A modelagem de valores extremos de fenômenos da natureza tais como velocidade de vento, nível da água de rio ou mar, altura de onda ou umidade é importante em estatística ambiental pois o conhecimento de valores extremos de tais eventos é crucial na prevenção de catátrofes. Ultimamente esta teoria é de particular interesse pois fenômenos extremos da natureza têm sido mais comuns e intensos. A maioria dos artigos sobre teoria do valor extremo para modelagem de dados considera amostras de tamanho moderado ou grande. A distribuição Gumbel é frequentemente incluída nas análises mas a qualidade do ajuste pode ser pobre em função de presença de ouliers. Investigamos modelagem estatística de eventos extremos com base na teoria de valores extremos. Consideramos um modelo de regressão valor extremo introduzido por Barreto-Souza & Vasconcellos (2011). Os autores trataram da questão de corrigir o viés do estimador de máxima verossimilhança para pequenas amostras. Nosso primeiro objetivo é deduzir ajustes para testes de hipótese nesta classe de modelos. Derivamos a estatística da razão de verossimilhanças ajustada de Skovgaard (2001) e cinco ajustes da estatística da razão de verossimilhanças sinalizada, que foram propostos por Barndorff-Nielsen (1986, 1991), DiCiccio & Martin (1993), Skovgaard (1996), Severini (1999) e Fraser et al. (1999). As estatísticas ajustadas são aproximadamente distribuídas como uma distribuição $\\chi^2$ e normal padrão com alto grau de acurácia. Os termos dos ajustes têm formas compactas simples que podem ser facilmente implementadas em softwares disponíveis. Comparamos a performance do teste da razão de verossimilhanças, do teste da razão de verossimilanças sinalizada e dos testes ajustados obtidos neste trabalho em amostras pequenas. Ilustramos uma aplicação dos testes usuais e suas versões modificadas em conjuntos de dados reais. As distribuições das estatísticas ajustadas são mais próximas das respectivas distribuições limites comparadas com as distribuições das estatísticas usuais quando o tamanho da amostra é relativamente pequeno. Os resultados de simulação indicaram que as estatísticas ajustadas são recomendadas para inferência em modelo de regressão valor extremo quando o tamanho da amostra é moderado ou pequeno. Parcimônia é importante quando os dados são escassos, mas flexibilidade também é crucial pois um ajuste pobre pode levar a uma conclusão completamente errada. Uma revisão da literatura foi feita para listar as distribuições que são generalizações da distribuição Gumbel. Nosso segundo objetivo é avaliar a parcimônia e flexibilidade destas distribuições. Com este propósito, comparamos tais distribuições através de momentos, coeficientes de assimetria e de curtose e índice da cauda. As famílias mais amplas obtidas pela inclusão de parâmetros adicionais, que têm a distribuição Gumbel como caso particular, apresentam assimetria e curtose flexíveis enquanto a distribuição Gumbel apresenta tais características constantes. Dentre estas distribuições, a distribuição valor extremo generalizada é a única com índice da cauda que pode ser qualquer número real positivo enquanto os índices da cauda das outras distribuições são zero. Observamos que algumas generalizações da distribuição Gumbel estudadas na literatura são não identificáveis. Portanto, para estes modelos a interpretação e estimação de parâmetros individuais não é factível. Selecionamos as distribuições identificáveis e as ajustamos a um conjunto de dados simulado e a um conjunto de dados reais de velocidade de vento. Como esperado, tais distribuições se ajustaram bastante bem ao conjunto de dados simulados de uma distribuição Gumbel. A distribuição valor extremo generalizada e a mistura de duas distribuições Gumbel produziram melhores ajustes aos dados do que as outras distribuições na presença não desprezível de observações discrepantes que não podem ser acomodadas pela distribuição Gumbel e, portanto, sugerimos que tais distribuições devem ser utilizadas neste contexto. / The extreme value theory is applied in research fields such as hydrology, pollution studies, materials engineering, traffic management, economics and finance. The Gumbel distribution is widely used in statistical modeling of extreme values of a natural process such as rainfall and wind. Also, the Gumbel distribution is important in the context of survival analysis for modeling lifetime in logarithmic scale. The statistical modeling of extreme values of a natural process such as wind or humidity is important in environmental statistics; for example, understanding extreme wind speed is crucial in catastrophe/disaster protection. Lately this is of particular interest as extreme natural phenomena/episodes are more common and intense. The majority of papers on extreme value theory for modeling extreme data is supported by moderate or large sample sizes. The Gumbel distribution is often considered but the resulting fit may be poor in the presence of ouliers since its skewness and kurtosis are constant. We deal with statistical modeling of extreme events data based on extreme value theory. We consider a general extreme-value regression model family introduced by Barreto-Souza & Vasconcellos (2011). The authors addressed the issue of correcting the bias of the maximum likelihood estimators in small samples. Here, our first goal is to derive hypothesis test adjustments in this class of models. We derive Skovgaard\'s adjusted likelihood ratio statistics Skovgaard (2001) and five adjusted signed likelihood ratio statistics, which have been proposed by Barndorff-Nielsen (1986, 1991), DiCiccio & Martin (1993), Skovgaard (1996), Severini (1999) and Fraser et al. (1999). The adjusted statistics are approximately distributed as $\\chi^2$ and standard normal with high accuracy. The adjustment terms have simple compact forms which may be easily implemented by readily available software. We compare the finite sample performance of the likelihood ratio test, the signed likelihood ratio test and the adjusted tests obtained in this work. We illustrate the application of the usual tests and their modified versions in real datasets. The adjusted statistics are closer to the respective limiting distribution compared to the usual ones when the sample size is relatively small. Simulation results indicate that the adjusted statistics can be recommended for inference in extreme value regression model with small or moderate sample size. Parsimony is important when data are scarce, but flexibility is also crucial since a poor fit may lead to a completely wrong conclusion. A literature review was conducted to list distributions which nest the Gumbel distribution. Our second goal is to evaluate their parsimony and flexibility. For this purpose, we compare such distributions regarding moments, skewness, kurtosis and tail index. The larger families obtained by introducing additional parameters, which have Gumbel embedded in, present flexible skewness and kurtosis while the Gumbel distribution skewness and kurtosis are constant. Among these distributions the generalized extreme value is the only one with tail index that can be any positive real number while the tail indeces of the other distributions investigated here are zero. We notice that some generalizations of the Gumbel distribution studied in the literature are not indetifiable. Hence, for these models meaningful interpretation and estimation of individual parameters are not feasible. We select the identifiable distributions and fit them to a simulated dataset and to real wind speed data. As expected, such distributions fit the Gumbel simulated data quite well. The generalized extreme value distribution and the two-component extreme value distribution fit the data better than the others in the non-negligible presence of outliers that cannot be accommodated by the Gumbel distribution, and therefore we suggest them to be applied in this context.

Diante das exigências estipuladas pelos órgãos reguladores pelos acordos internacionais, tendo em vistas as inúmeras crises financeiras ocorridas nos últimos séculos, as instituições financeiras desenvolveram diversas ferramentas para a mensuração e controle do risco inerente aos negócios. Apesar da crescente evolução das metodologias de cálculo e mensuração do risco, o Value at Risk (VaR) se tornou referência como ferramenta de estimação do risco de mercado. Nos últimos anos novas técnicas de cálculo do Value at Risk (VaR) vêm sendo desenvolvidas. Porém, nenhuma tem sido considerada como a que melhor ajusta os riscos para diversos mercados e em diferentes momentos. Não existe na literatura um modelo conciso e coerente com as diversidades dos mercados. Assim, o presente trabalho tem por objetivo geral avaliar os estimadores de risco de mercado, gerados pela aplicação de modelos baseados no Value at Risk (VaR), aplicados aos índices das principais bolsas dos países desenvolvidos e emergentes, para os períodos normais e de crise financeira, de modo a apurar os mais efetivos nessa função. Foram considerados no estudo os modelos VaR Não condicional, pelos modelos tradicionais (Simulação Histórica, Delta-Normal e t-Student) e baseados na Teoria de Valores Extremos; o VaR Condicional, comparando os modelos da família ARCH e Riskmetrics e o VaR Multivariado, com os modelos GARCH bivariados (Vech, Bekk e CCC), funções cópulas (t-Student, Clayton, Frank e Gumbel) e por Redes Neurais Artificiais. A base de dados utilizada refere-se as amostras diárias dos retornos dos principais índices de ações dos países desenvolvidos (Alemanha, Estados Unidos, França, Reino Unido e Japão) e emergentes (Brasil, Rússia, Índia, China e África do Sul), no período de 1995 a 2013, contemplando as crises de 1997 e 2008. Os resultados do estudo foram, de certa forma, distintos das premissas iniciais estabelecidas pelas hipóteses de pesquisa. Diante de mais de mil modelagens realizadas, os modelos condicionais foram superiores aos não condicionais, na maioria dos casos. Em específico o modelo GARCH (1,1), tradicional na literatura, teve uma efetividade de ajuste de 93% dos casos. Para a análise Multivariada, não foi possível definir um modelo mais assertivo. Os modelos Vech, Bekk e Cópula - Clayton tiveram desempenho semelhantes, com bons ajustes em 100% dos testes. Diferentemente do que era esperado, não foi possível perceber diferenças significativas entre os ajustes para países desenvolvidos e emergentes e os momentos de crise e normal. O estudo contribuiu na percepção de que os modelos utilizados pelas instituições financeiras não são os que apresentam melhores resultados na estimação dos riscos de mercado, mesmo sendo recomendados pelas instituições renomadas. Cabe uma análise mais profunda sobre o desempenho dos estimadores de riscos, utilizando simulações com as carteiras de cada instituição financeira. / Given the requirements stipulated by regulatory agencies for international agreements, in considering the numerous financial crises in the last centuries, financial institutions have developed several tools to measure and control the risk of the business. Despite the growing evolution of the methodologies of calculation and measurement of Value at Risk (VaR) has become a reference tool as estimate market risk. In recent years new calculation techniques of Value at Risk (VaR) have been developed. However, none has been considered the one that best fits the risks for different markets and in different times. There is no literature in a concise and coherent model with the diversity of markets. Thus, this work has the objective to assess the market risk estimates generated by the application of models based on Value at Risk (VaR), applied to the indices of the major stock exchanges in developed and emerging countries, for normal and crisis periods financial, in order to ascertain the most effective in that role. Were considered in the study models conditional VaR, the conventional models (Historical Simulation, Delta-Normal and Student t test) and based on Extreme Value Theory; Conditional VaR by comparing the models of ARCH family and RiskMetrics and the Multivariate VaR, with bivariate GARCH (VECH, Bekk and CCC), copula functions (Student t, Clayton, Frank and Gumbel) and Artificial Neural Networks. The database used refers to the daily samples of the returns of major stock indexes of developed countries (Germany, USA, France, UK and Japan) and emerging (Brazil, Russia, India, China and South Africa) from 1995 to 2013, covering the crisis in 1997 and 2008. The results were somewhat different from the initial premises established by the research hypotheses. Before more than 1 mil modeling performed, the conditional models were superior to non-contingent, in the majority of cases. In particular the GARCH (1,1) model, traditional literature, had a 93% adjustment effectiveness of cases. For multivariate analysis, it was not possible to set a more assertive style. VECH models, and Bekk, Copula - Clayton had similar performance with good fits to 100% of the tests. Unlike what was expected, it was not possible to see significant differences between the settings for developed and emerging countries and the moments of crisis and normal. The study contributed to the perception that the models used by financial institutions are not the best performing in the estimation of market risk, even if recommended by renowned institutions. It is a deeper analysis on the performance of the estimators of risk, using simulations with the portfolios of each financial institution.

[pt] Com diversas aplicações em matemática, física e finanças, Teoria das Matrizes Aleatórias (RMT) recentemente atraiu muita atenção. Enquanto o RMT Hermitiano é de especial importância na física por causa da Hermenticidade de operadores associados a observáveis em mecânica quântica, O RMT não-Hermitiano também atraiu uma atenção considerável, em particular porque eles podem ser usados como modelos para sistemas físicos dissipativos ou abertos. No entanto, devido à ausência de uma simetria simplificada, o estudo de matrizes aleatórias não-Hermitianas é, em geral, uma tarefa difícil. Um subconjunto especial de matrizes aleat órias não-Hermitianas, as chamadas matrizes aleatórias normais, são modelos interessantes a serem considerados, uma vez que oferecem mais simetria, tornando-as mais acessíveis às investigções analíticas. Por definição, uma matriz normal M é uma matriz quadrada que troca com seu adjunto Hermitiano. Nesta tese, amplicamos a derivação de estatísticas de valores extremos (EVS) de matrizes aleatórias Hermitianas, com base na abordagem de polinômios ortogonais, em matrizes aleatórias normais e em gases Coulomb 2D em geral. A força desta abordagem a sua compreensão física e intuitiva. Em primeiro lugar, essa abordagem fornece uma derivação alternativa de resultados na literatura. Precisamente falando, mostramos a convergência do autovalor redimensionado com o maior módulo de um conjunto de Ginibre para uma distribuição de Gumbel, bem como a universalidade para um potencial arbitrário radialmente simtérico que atenda certas condições. Em segundo lugar, mostra-se que esta abordagem pode ser generalizada para obter a convergência do autovalor com menor módulo e sua universalidade no limite interno finito do suporte do autovalor. Um aspecto interessante deste trabalho é o fato de que podemos usar técnicas padrão de matrizes aleatórias Hermitianas para obter o EVS de matrizes aleatórias não Hermitianas. / [en] With diverse applications in mathematics, physics, and finance, Random Matrix Theory (RMT) has recently attracted a great deal of attention. While Hermitian RMT is of special importance in physics because of the Hermiticity of operators associated with observables in quantum mechanics, non-Hermitian RMT has also attracted a considerable attention, in particular because they can be used as models for dissipative or open physical systems. However, due to the absence of a simplifying symmetry, the study of non-Hermitian random matrices is, in general, a diffcult task. A special subset of non-Hermitian random matrices, the so-called random normal matrices, are interesting models to consider, since they offer more symmetry, thus making them more amenable to analytical investigations. By definition, a normal matrix M is a square matrix which commutes with its Hermitian adjoint, i.e., (M, M (1)). In this thesis, we present a novel derivation of extreme value statistics (EVS) of Hermitian random matrices, namely the approach of orthogonal polynomials, to normal random matrices and 2D Coulomb gases in general. The strength of this approach is its physical and intuitive understanding. Firstly, this approach provides an alternative derivation of results in the literature. Precisely speaking, we show convergence of the rescaled eigenvalue with largest modulus of a Ginibre ensemble to a Gumbel distribution, as well as universality for an arbitrary radially symmetric potential which meets certain conditions. Secondly, it is shown that this approach can be generalised to obtain convergence of the eigenvalue with smallest modulus and its universality at the finite inner edge of the eigenvalue support. One interesting aspect of this work is the fact that we can use standard techniques from Hermitian random matrices to obtain the EVS of non-Hermitian random matrices.

La protection des zones littorales contre les agressions naturelles provenant de la mer, et notamment contre le risque de submersion marine, est essentielle pour sécuriser les installations côtières. La prévention de ce risque est assurée par des protections côtières qui sont conçues et régulièrement vérifiées grâce généralement à la définition du concept de niveau de retour d’un événement extrême particulier. Le niveau de retour lié à une période de retour assez grande (de 1000 ans ou plus) est estimé par des méthodes statistiques basées sur la Théorie des Valeurs Extrêmes (TVE). Ces approches statistiques sont appliquées à des séries temporelles d’une variable extrême observée et permettent de connaître la probabilité d’occurrence de telle variable. Dans le passé, les niveaux de retour des aléas maritimes extrêmes étaient estimés le plus souvent à partir de méthodes statistiques appliquées à des séries d’observation locales. En général, les séries locales des niveaux marins sont observées sur une période limitée (pour les niveaux marins environ 50 ans) et on cherche à trouver des bonnes estimations des extrêmes associées à des périodes de retour très grandes. Pour cette raison, de nombreuses méthodologies sont utilisées pour augmenter la taille des échantillons des extrêmes et réduire les incertitudes sur les estimations. En génie côtier, une des approches actuellement assez utilisées est l’analyse régionale. L’analyse régionale est indiquée par Weiss (2014) comme une manière très performante pour réduire les incertitudes sur les estimations des événements extrêmes. Le principe de cette méthodologie est de profiter de la grande disponibilité spatiale des données observées sur différents sites pour créer des régions homogènes. Cela permet d’estimer des lois statistiques sur des échantillons régionaux plus étendus regroupant tous les événements extrêmes qui ont frappé un ou plusieurs sites de la région (...) Cela ainsi que le caractère particulier de chaque événement historique ne permet pas son utilisation dans une analyse régionale classique. Une méthodologie statistique appelée FAB qui permet de réaliser une analyse régionale tenant en compte les données historiques est développée dans ce manuscrit. Élaborée pour des données POT (Peaks Over Threshold), cette méthode est basée sur une nouvelle définition d’une durée d’observation, appelée durée crédible, locale et régionale et elle est capable de tenir en compte dans l’analyse statistique les trois types les plus classiques de données historiques (données ponctuelles, données définies par un intervalle, données au-dessus d’une borne inférieure). En plus, une approche pour déterminer un seuil d’échantillonnage optimal est définie dans cette étude. La méthode FAB est assez polyvalente et permet d’estimer des niveaux de retour soit dans un cadre fréquentiste soit dans un cadre bayésien. Une application de cette méthodologie est réalisée pour une base de données enregistrées des surcotes de pleine mer (données systématiques) et 14 surcotes de pleine mer historiques collectées pour différents sites positionnés le long des côtes françaises, anglaises, belges et espagnoles de l’Atlantique, de la Manche et de la mer du Nord. Enfin, ce manuscrit examine la problématique de la découverte et de la validation des données historiques / The protection of coastal areas against the risk of flooding is necessary to safeguard all types of waterside structures and, in particular, nuclear power plants. The prevention of flooding is guaranteed by coastal protection commonly built and verified thanks to the definition of the return level’s concept of a particular extreme event. Return levels linked to very high return periods (up to 1000 years) are estimated through statistical methods based on the Extreme Value Theory (EVT). These statistical approaches are applied to time series of a particular extreme variable observed and enables the computation of its occurrence probability. In the past, return levels of extreme coastal events were frequently estimated by applying statistical methods to time series of local observations. Local series of sea levels are typically observed in too short a period (for sea levels about 50 years) in order to compute reliable estimations linked to high return periods. For this reason, several approaches are used to enlarge the size of the extreme data samples and to reduce uncertainties of their estimations. Currently, one of the most widely used methods in coastal engineering is the Regional Analysis. Regional Analysis is denoted by Weiss (2014) as a valid means to reduce uncertainties in the estimations of extreme events. The main idea of this method is to take advantage of the wide spatial availability of observed data in different locations in order to form homogeneous regions. This enables the estimation of statistical distributions of enlarged regional data samples by clustering all extreme events occurred in one or more sites of the region. Recent investigations have highlighted the importance of using past events when estimating extreme events. When historical data are available, they cannot be neglected in order to compute reliable estimations of extreme events. Historical data are collected from different sources and they are identified as data that do not come from time series. In fact, in most cases, no information about other extreme events occurring before and after a historical observation is available. This, and the particular nature of each historical data, do not permit their use in a Regional Analysis. A statistical methodology that enables the use of historical data in a regional context is needed in order to estimate reliable return levels and to reduce their associated uncertainties. In this manuscript, a statistical method called FAB is developed enabling the performance of a Regional Analysis using historical data. This method is formulated for POT (Peaks Over Threshold) data. It is based on the new definition of duration of local and regional observation period (denominated credible duration) and it is able to take into account all the three typical kinds of historical data (exact point, range and lower limit value). In addition, an approach to identify an optimal sampling threshold is defined in this study. This allows to get better estimations through using the optimal extreme data sample in the FAB method.FAB method is a flexible approach that enables the estimation of return levels both in frequentist and Bayesian contexts. An application of this method is carried out for a database of recorded skew surges (systematic data) and for 14 historical skew surges recovered from different sites located on French, British, Belgian and Spanish coasts of the Atlantic Ocean, the English Channel and the North Sea. Frequentist and Bayesian estimations of skew surges are computed for each homogeneous region and for every site. Finally, this manuscript explores the issues surrounding the finding and validation of historical data

Cette thèse s'intéresse à l'estimation de certaines mesures de risque d'une variable aléatoire réelle Y en présence d'une covariable X. Pour cela, on va considérer que le vecteur (X,Y) suit une loi elliptique. Dans un premier temps, on va s'intéresser aux quantiles de Y sachant X=x. On va alors tester d'abord un modèle de régression quantile assez répandu dans la littérature, pour lequel on obtient des résultats théoriques que l'on discutera. Face aux limites d'un tel modèle, en particulier pour des niveaux de quantile dits extrêmes, on proposera une nouvelle approche plus adaptée. Des résultats asymptotiques sont donnés, appuyés par une étude numérique puis par un exemple sur des données réelles. Dans un second chapitre, on s'intéressera à une autre mesure de risque appelée expectile. La structure du chapitre est sensiblement la même que celle du précédent, à savoir le test d'un modèle de régression inadapté aux expectiles extrêmes, pour lesquels on propose une approche méthodologique puis statistique. De plus, en mettant en évidence le lien entre les quantiles et expectiles extrêmes, on s'aperçoit que d'autres mesures de risque extrêmes sont étroitement liées aux quantiles extrêmes. On se concentrera sur deux familles appelées Lp-quantiles et mesures d'Haezendonck-Goovaerts, pour lesquelles on propose des estimateurs extrêmes. Une étude numérique est également fournie. Enfin, le dernier chapitre propose quelques pistes pour traiter le cas où la taille de la covariable X est grande. En constatant que nos estimateurs définis précédemment étaient moins performants dans ce cas, on s'inspire alors de quelques méthodes d'estimation en grande dimension pour proposer d'autres estimateurs. Une étude numérique permet d'avoir un aperçu de leurs performances / This PhD thesis focuses on the estimation of some risk measures for a real random variable Y with a covariate vector X. For that purpose, we will consider that the random vector (X,Y) is elliptically distributed. In a first time, we will deal with the quantiles of Y given X=x. We thus firstly investigate a quantile regression model, widespread in the litterature, for which we get theoretical results that we discuss. Indeed, such a model has some limitations, especially when the quantile level is said extreme. Therefore, we propose another more adapted approach. Asymptotic results are given, illustrated by a simulation study and a real data example.In a second chapter, we focus on another risk measure called expectile. The structure of the chapter is essentially the same as that of the previous one. Indeed, we first use a regression model that is not adapted to extreme expectiles, for which a methodological and statistical approach is proposed. Furthermore, highlighting the link between extreme quantiles and expectiles, we realize that other extreme risk measures are closely related to extreme quantiles. We will focus on two families called Lp-quantiles and Haezendonck-Goovaerts risk measures, for which we propose extreme estimators. A simulation study is also provided. Finally, the last chapter is devoted to the case where the size of the covariate vector X is tall. By noticing that our previous estimators perform poorly in this case, we rely on some high dimensional estimation methods to propose other estimators. A simulation study gives a visual overview of their performances