Title: Some possibilities of heteroskedasticity modeling with applications to non-life insurance Author:Petra Pavlačková Department: Department of Probability and Mathematical Statistics Supervisor: Ing. Zimmermann Pavel, Ph.d. Abstract: This thesis deals with the possibilities of modeling heteroskedasticity using generalized linear models. It summarizes the assumption for these models and their application in practice. It shows the practical need for these models. Furthermore, the thesis deals with the modeling of variance using other methods than generalized lienar models - such as generalized additive models or local regression. Comparison of methods is graphically demonstrated. Keywords: Dispersion parameter, variance function, Joint modelling of mean and dispersion

Doctor of Philosophy / Department of Statistics / Weixing Song / The regression model has been given a considerable amount of attention and played a significant role in data analysis. The usual assumption in regression analysis is that the variances of the error terms are constant across the data. Occasionally, this assumption of homoscedasticity on the variance is violated; and the data generated from real world applications exhibit heteroscedasticity. The practical importance of detecting heteroscedasticity in regression analysis is widely recognized in many applications because efficient inference for the regression function requires unequal variance to be taken into account. The goal of this thesis is to propose new testing procedures to assess the adequacy of fitting parametric variance function in heteroscedastic regression models. The proposed tests are established in Chapter 2 using certain minimized L[subscript]2 distance between a nonparametric and a parametric variance function estimators. The asymptotic distribution of the test statistics corresponding to the minimum distance estimator under the fixed model and that of the corresponding minimum distance estimators are shown to be normal. These estimators turn out to be [sqrt]n consistent. The asymptotic power of the proposed test against some local nonparametric alternatives is also investigated. Numerical simulation studies are employed to evaluate the nite sample performance of the test in one dimensional and two dimensional cases. The minimum distance method in Chapter 2 requires the calculation of the integrals in the test statistics. These integrals usually do not have a tractable form. Therefore, some numerical integration methods are needed to approximate the integrations. Chapter 3 discusses a nonparametric empirical smoothing lack-of-fit test for the functional form of the variance in regression models that do not involve evaluation of integrals. empirical smoothing lack-of-fit test can be treated as a nontrivial modification of Zheng (1996)'s nonparametric smoothing test and Koul and Ni (2004)'s minimum distance test for the mean function in the classic regression models. The asymptotic normality of the proposed test under the null hypothesis is established. Consistency at some fixed alternatives and asymptotic power under some local alternatives are also discussed. Simulation studies are conducted to assess the nite sample performance of the test. The simulation studies show that the proposed empirical smoothing test is more powerful and computationally more efficient than the minimum distance test and Wang and Zhou (2006)'s test.

In this work, a method to choose the best run order for a given experimental design is proposed, for the simple linear regression model with MA errors. More specifically we investigate the best run order of an uniform design when errors follow a MA(1) or a subset MA(k) process where k is a positive integer. The correlation matrix P resulting from the errors is usually difficult to obtain a good estimate. Using the change of variance function(CVF) to see the relation of the uncorrelated and the serially correlated errors. Criterion proposed by Zhou (2001), we find the best run order of the uniform design on [-1,1] to minimize the robust criterion, |CVF|. We will display the permutation of a run order after rearrangement by our method and show how the structure is decomposed into three categories to solve the problem.

Abstract This thesis presents methods for modelling conditional variance and uncertainty of prediction at a query point on the basis of industrial process data. The introductory part of the thesis provides an extensive background of the examined methods and a summary of the results. The results are presented in detail in the original papers. The application presented in the thesis is modelling of the mean and variance of the mechanical properties of steel plates. Both the mean and variance of the mechanical properties depend on many process variables. A method for predicting the probability of rejection in a quali?cation test is presented and implemented in a tool developed for the planning of strength margins. The developed tool has been successfully utilised in the planning of mechanical properties in a steel plate mill. The methods for modelling the dependence of conditional variance on input variables are reviewed and their suitability for large industrial data sets are examined. In a comparative study, neural network modelling of the mean and dispersion narrowly performed the best. A method is presented for evaluating the uncertainty of regression-type prediction at a query point on the basis of predicted conditional variance, model variance and the effect of uncertainty about explanatory variables at early process stages. A method for measuring the uncertainty of prediction on the basis of the density of the data around the query point is proposed. The proposed distance measure is utilised in comparing the generalisation ability of models. The generalisation properties of the most important regression learning methods are studied and the results indicate that local methods and quadratic regression have a poor interpolation capability compared with multi-layer perceptron and Gaussian kernel support vector regression. The possibility of adaptively modelling a time-varying conditional variance function is disclosed. Two methods for adaptive modelling of the variance function are proposed. The background of the developed adaptive variance modelling methods is presented.

Cette thèse examine la dispersion multivariée des modelés normales stables Tweedie. Trois estimateurs de fonction variance généralisée sont discutés. Ensuite dans le cadre de la famille exponentielle naturelle deux caractérisations du modèle normal-Poisson, qui est un cas particulier de modèles normales stables Tweedie avec composante discrète, sont indiquées : d'abord par fonction variance et ensuite par fonction variance généralisée. Le dernier fournit la solution à un problème particulier d'équation de Monge-Ampère. Enfin, pour illustrer l'application de la variance généralisée des modèles Tweedie stables normales, des exemples à partir des données réelles sont fournis. / This thesis examines the multivariate dispersion of normal stable Tweedie (NST) models. Three generalize variance estimators of some NST models are discussed. Then within the framework of natural exponential family, two characterizations of normal Poisson model, which is a special case of NST models with discrete component, are shown : first by variance function and then by generalized variance function. The latter provides a solution to a particular Monge-Ampere equation problem. Finally, to illustrate the application of generalized variance of normal stable Tweedie models, examples from real data are provided.

Après avoir fait un rappel sur les généralités concernant les familles exponentielles naturelles et  les lois Tweedie univariées qui en sont un exemple particulier, nous montrerons comment étendre ces lois au cas multivarié. Une première construction permettra de définir des vecteurs aléatoires Tweedie paramétrés pas un vecteur de moyenne et une matrice de dispersion. Nous montrerons que les corrélations entre les lois marginales peuvent être contrôlées et varient entre -1 et 1. Nous verrons aussi que ces vecteurs ont quelques propriétés communes avec les vecteurs gaussiens. Nous en donnerons une représentation matricielle qui permettra d'en simuler des observations. La seconde construction permettra d'introduire les modèles Tweedie multiples constitués d'une variable Tweedie dont l'observation sera la dispersion des autres marges, toutes de lois Tweedie elles aussi. Nous donnerons la variance généralisée de ces lois et montrerons que cette dernière peut-être estimée efficacement. Enfin, nous verrons que, modulo certaines restrictions, nous pourrons donner une caractérisation par la fonction de variance généralisée des familles exponentielles naturelles générées par ces lois. / After a reminder of the natural exponential families framework and the univariate Tweedie distributions, we build two multivariate extension of the latter. A first construction, called Tweedie random vector, gives a multivariate Tweedie distribution parametrized by a mean vector and a dispersion matrix. We show that the correlations between the margins can be controlled and vary between -1 and 1. Some properties shared with the well-known Gaussian vector are given. By giving a matrix representation, we can simulate observations of Tweedie random vectors. The second construction establishes the multiple stable Tweedie models. They are vectors of which the first component is Tweedie and the others are independant Tweedie, given the first one, and with dispersion parameter given by an observation of the first component. We give the generalized variance and show that it is a product of powered component of the mean and give an efficient estimator of this parameter. Finally, we can show, with some restrictions, that the generalized variance is a tool which can be used for characterizing the natural exponential families generated by multiple stable Tweedie models.

Improving the estimation efficiency has always been one of the important aspects in statistical modelling. Our goal is to develop new statistical methodologies yielding more efficient estimators in the analysis of longitudinal data. In this thesis, we consider two different approaches, empirical likelihood and jointly modelling the mean and variance, to improve the estimation efficiency. In part I of this thesis, empirical likelihood-based inference for longitudinal data within the framework of generalized linear model is investigated. The proposed procedure takes into account the within-subject correlation without involving direct estimation of nuisance parameters in the correlation matrix and retains optimality even if the working correlation structure is misspecified. The proposed approach yields more efficient estimators than conventional generalized estimating equations and achieves the same asymptotic variance as quadratic inference functions based methods. The second part of this thesis focus on the joint mean-variance models. We proposed a data-driven approach to modelling the mean and variance simultaneously, yielding more efficient estimates of the mean regression parameters than the conventional generalized estimating equations approach even if the within-subject correlation structure is misspecified in our joint mean-variance models. The joint mean-variances in parametric form as well as semi-parametric form has been investigated. Extensive simulation studies are conducted to assess the performance of our proposed approaches. Three longitudinal data sets, Ohio Children’s wheeze status data (Ware et al., 1984), Cattle data (Kenward, 1987) and CD4+ data (Kaslowet al., 1987), are used to demonstrate our models and approaches.

Ce travail de thèse porte sur différentes caractérisations des modèles multivariés de stables-Tweedie multiples dans le cadre des familles exponentielles naturelles sous la propriété de "steepness". Ces modèles parus en 2014 dans la littérature ont été d’abord introduits et décrits sous une forme restreinte des stables-Tweedie normaux avant les extensions aux cas multiples. Ils sont composés d’un mélange d’une loi unidimensionnelle stable-Tweedie de variable réelle positive fixée, et des lois stables-Tweedie de variables réelles indépendantes conditionnées par la première fixée, de même variance égale à la valeur de la variable fixée. Les modèles stables-Tweedie normaux correspondants sont ceux du mélange d’une loi unidimensionnelle stable-Tweedie positive fixé et les autres toutes gaussiennes indépendantes. A travers des cas particuliers tels que normal, Poisson, gamma, inverse gaussienne, les modèles stables-Tweedie multiples sont très fréquents dans les études de statistique et probabilités appliquées. D’abord, nous avons caractérisé les modèles stables-Tweedie normaux à travers leurs fonctions variances ou matrices de covariance exprimées en fonction de leurs vecteurs moyens. La nature des polynômes associés à ces modèles est déduite selon les valeurs de la puissance variance à l’aide des propriétés de quasi orthogonalité, des systèmes de Lévy-Sheffer, et des relations de récurrence polynomiale. Ensuite, ces premiers résultats nous ont permis de caractériser à l’aide de la fonction variance la plus grande classe des stables-Tweedie multiples. Ce qui a conduit à une nouvelle classification laquelle rend la famille beaucoup plus compréhensible. Enfin, une extension de caractérisation des stables-Tweedie normaux par fonction variance généralisée ou déterminant de la fonction variance a été établie via leur propriété d’indéfinie divisibilité et en passant par les équations de Monge-Ampère correspondantes. Exprimées sous la forme de produit des composantes du vecteur moyen aux puissances multiples, la caractérisationde tous les modèles multivariés stables-Tweedie multiples par fonction variance généralisée reste un problème ouvert. / In the framework of natural exponential families, this thesis proposes differents characterizations of multivariate multiple stables-Tweedie under "steepness" property. These models appeared in 2014 in the literature were first introduced and described in a restricted form of the normal stables-Tweedie models before extensions to multiple cases. They are composed by a fixed univariate stable-Tweedie variable having a positive domain, and the remaining random variables given the fixed one are reals independent stables-Tweedie variables, possibly different, with the same dispersion parameter equal to the fixed component. The corresponding normal stables-Tweedie models have a fixed univariate stable-Tweedie and all the others are reals Gaussian variables. Through special cases such that normal, Poisson, gamma, inverse Gaussian, multiple stables-Tweedie models are very common in applied probability and statistical studies. We first characterized the normal stable-Tweedie through their variances function or covariance matrices expressed in terms of their means vector. According to the power variance parameter values, the nature of polynomials associated with these models is deduced with the properties of the quasi orthogonal, Levy-Sheffer systems, and polynomial recurrence relations. Then, these results allowed us to characterize by function variance the largest class of multiple stables-Tweedie. Which led to a new classification, which makes more understandable the family. Finally, a extension characterization of normal stable-Tweedie by generalized variance function or determinant of variance function have been established via their infinite divisibility property and through the corresponding Monge-Ampere equations. Expressed as product of the components of the mean vector with multiple powers parameters reals, the characterization of all multivariate multiple stable- Tweedie models by generalized variance function remains an open problem.

Cette thèse contient deux parties différentes. Dans la première partie, nous nous sommes intéressés aux familles exponentielles naturelles cubiques dont la fonction variance est un polynôme de degré inférieur ou égal à 3. Nous donnons trois caractérisations de ces familles en se basant sur une approche Bayesienne. L’une de ces caractérisations repose sur le fait que la fonction cumulante vérifie une équation différentielle. La deuxième partie de notre travail est consacrée aux conséquences de la propriété d’indépendance de type « Matsumoto-Yor » qui a été développée par Koudou et Vallois. Cette propriété fait intervenir la famille de lois de Kummer de type 2 et les lois Beta généralisées. En se basant sur la méthode de conditionnement et sur la méthode de rejet, nous donnons des réalisations presque sûre de ces distributions de probabilités. D’autre part, nous caractérisons la famille de lois de Kummer de type 2 (resp. les lois Beta généralisées) par une équation algébrique impliquant des lois gamma (resp. les lois Beta) / This thesis has two different parts. In the first part we are interested in the real cubic natural exponential families such that their variance function is a polynomial of degree less than or equal to 3. We give three characterizations of such families using a Bayesian approach. One of these characterizations is based on a differential equation verified by the cumulant function. In a second part we study in depth the independence property of the type “Matsumoto-Yor” that was developed by Koudou and Vallois. This property involves the Kummer distribution of type 2 and the generalized beta ones. Using the conditioning and the rejection method, we give almost sure realization of these distributions. We characterize the family of Kummer distribution of type 2 with an algebraic equation involving the gamma ones. We proceed similarly with the generalized beta distributions

Cette thèse a pour objet d'étude les méthodes numériques d'approximation de la densité de probabilité associée à des variables aléatoires admettant des distributions composées. Ces variables aléatoires sont couramment utilisées en actuariat pour modéliser le risque supporté par un portefeuille de contrats. En théorie de la ruine, la probabilité de ruine ultime dans le modèle de Poisson composé est égale à la fonction de survie d'une distribution géométrique composée. La méthode numérique proposée consiste en une projection orthogonale de la densité sur une base de polynômes orthogonaux. Ces polynômes sont orthogonaux par rapport à une mesure de probabilité de référence appartenant aux Familles Exponentielles Naturelles Quadratiques. La méthode d'approximation polynomiale est comparée à d'autres méthodes d'approximation de la densité basées sur les moments et la transformée de Laplace de la distribution. L'extension de la méthode en dimension supérieure à $1$ est présentée, ainsi que l'obtention d'un estimateur de la densité à partir de la formule d'approximation. Cette thèse comprend aussi la description d'une méthode d'agrégation adaptée aux portefeuilles de contrats d'assurance vie de type épargne individuelle. La procédure d'agrégation conduit à la construction de model points pour permettre l'évaluation des provisions best estimate dans des temps raisonnables et conformément à la directive européenne Solvabilité II. / This PhD thesis studies numerical methods to approximate the probability density function of random variables governed by compound distributions. These random variables are useful in actuarial science to model the risk of a portfolio of contracts. In ruin theory, the probability of ultimate ruin within the compound Poisson ruin model is the survival function of a geometric compound distribution. The proposed method consists in a projection of the probability density function onto an orthogonal polynomial system. These polynomials are orthogonal with respect to a probability measure that belongs to Natural Exponential Families with Quadratic Variance Function. The polynomiam approximation is compared to other numerical methods that recover the probability density function from the knowledge of the moments or the Laplace transform of the distribution. The polynomial method is then extended in a multidimensional setting, along with the probability density estimator derived from the approximation formula. An aggregation procedure adapted to life insurance portfolios is also described. The method aims at building a portfolio of model points in order to compute the best estimate liabilities in a timely manner and in a way that is compliant with the European directive Solvency II.