Une classe générale de statistiques de rangs basées sur la fonction caractéristique est introduite afin de tester l'hypothèse composite d'appartenance à une famille de copules multidimensionnelles. Ces statistiques d'adéquation sont définies comme des distances fonctionnelles de type L_2 pondérées entre une version non paramétrique et une version semi-paramétrique de la fonction caractéristique que l'on peut associer à une copule. Il est démontré que ces statistiques de test se comportent asymptotiquement comme des V-statistiques dégénérées d'ordre quatre et que leurs lois limites s'expriment en termes de sommes pondérées de variables khi-deux indépendantes. La convergence des tests sous des alternatives générales est établie, de même que la validité du bootstrap paramétrique pour le calcul de valeurs critiques. Le comportement des nouveaux tests sous des tailles d'échantillons faibles et modérées est étudié à l'aide de simulations et est comparé à celui d'un test concurrent fondé sur la copule empirique. La méthodologie est finalement illustrée sur un jeu de données à plusieurs dimensions.

This thesis deals with univariate and multivariate rank methods in making statistical inference. It is assumed that the underlying distributions belong to the class of elliptical distributions. The class of elliptical distributions is an extension of the normal distribution and includes distributions with both lighter and heavier tails than the normal distribution. In the first part of the thesis the rank covariance matrices defined via the Oja median are considered. The Oja rank covariance matrix has two important properties: it is affine equivariant and it is proportional to the inverse of the regular covariance matrix. We employ these two properties to study the problem of estimating the rank covariance matrices when they have a certain structure. The second part, which is the main part of the thesis, is devoted to rank estimation in linear regression models with symmetric heteroscedastic errors. We are interested in asymptotic properties of rank estimates. Asymptotic uniform linearity of a linear rank statistic in the case of heteroscedastic variables is proved. The asymptotic uniform linearity property enables to study asymptotic behaviour of rank regression estimates and rank tests. Existing results are generalized and it is shown that the Jaeckel estimate is consistent and asymptotically normally distributed also for heteroscedastic symmetric errors.

Ensembles are today routinely applied to estimate uncertainty in numerical predictions of complex systems such as the weather. Instead of initializing a single numerical forecast, using only the best guess of the present state as initial conditions, a collection (an ensemble) of forecasts whose members start from slightly different initial conditions is calculated. By varying the initial conditions within their error bars, the sensitivity of the resulting forecasts to these measurement errors can be accounted for. The ensemble approach can also be applied to estimate forecast errors that are due to insufficiently known model parameters by varying these parameters between ensemble members. An important (and difficult) question in ensemble weather forecasting is how well does an ensemble of forecasts reproduce the actual forecast uncertainty. A widely used criterion to assess the quality of forecast ensembles is statistical consistency which demands that the ensemble members and the corresponding measurement (the ``verification\'\') behave like random independent draws from the same underlying probability distribution. Since this forecast distribution is generally unknown, such an analysis is nontrivial. An established criterion to assess statistical consistency of a historical archive of scalar ensembles and verifications is uniformity of the verification rank: If the verification falls between the (k-1)-st and k-th largest ensemble member it is said to have rank k. Statistical consistency implies that the average frequency of occurrence should be the same for each rank. A central result of the present thesis is that, in a statistically consistent K-member ensemble, the (K+1)-dimensional vector of rank probabilities is a random vector that is uniformly distributed on the K-dimensional probability simplex. This behavior is universal for all possible forecast distributions. It thus provides a way to describe forecast ensembles in a nonparametric way, without making any assumptions about the statistical behavior of the ensemble data. The physical details of the forecast model are eliminated, and the notion of statistical consistency is captured in an elementary way. Two applications of this result to ensemble analysis are presented. Ensemble stratification, the partitioning of an archive of ensemble forecasts into subsets using a discriminating criterion, is considered in the light of the above result. It is shown that certain stratification criteria can make the individual subsets of ensembles appear statistically inconsistent, even though the unstratified ensemble is statistically consistent. This effect is explained by considering statistical fluctuations of rank probabilities. A new hypothesis test is developed to assess statistical consistency of stratified ensembles while taking these potentially misleading stratification effects into account. The distribution of rank probabilities is further used to study the predictability of outliers, which are defined as events where the verification falls outside the range of the ensemble, being either smaller than the smallest, or larger than the largest ensemble member. It is shown that these events are better predictable than by a naive benchmark prediction, which unconditionally issues the average outlier frequency of 2/(K+1) as a forecast. Predictability of outlier events, quantified in terms of probabilistic skill scores and receiver operating characteristics (ROC), is shown to be universal in a hypothetical forecast ensemble. An empirical study shows that in an operational temperature forecast ensemble, outliers are likewise predictable, and that the corresponding predictability measures agree with the analytically calculated ones.

Ensembles are today routinely applied to estimate uncertainty in numerical predictions of complex systems such as the weather. Instead of initializing a single numerical forecast, using only the best guess of the present state as initial conditions, a collection (an ensemble) of forecasts whose members start from slightly different initial conditions is calculated. By varying the initial conditions within their error bars, the sensitivity of the resulting forecasts to these measurement errors can be accounted for. The ensemble approach can also be applied to estimate forecast errors that are due to insufficiently known model parameters by varying these parameters between ensemble members. An important (and difficult) question in ensemble weather forecasting is how well does an ensemble of forecasts reproduce the actual forecast uncertainty. A widely used criterion to assess the quality of forecast ensembles is statistical consistency which demands that the ensemble members and the corresponding measurement (the ``verification\'\') behave like random independent draws from the same underlying probability distribution. Since this forecast distribution is generally unknown, such an analysis is nontrivial. An established criterion to assess statistical consistency of a historical archive of scalar ensembles and verifications is uniformity of the verification rank: If the verification falls between the (k-1)-st and k-th largest ensemble member it is said to have rank k. Statistical consistency implies that the average frequency of occurrence should be the same for each rank. A central result of the present thesis is that, in a statistically consistent K-member ensemble, the (K+1)-dimensional vector of rank probabilities is a random vector that is uniformly distributed on the K-dimensional probability simplex. This behavior is universal for all possible forecast distributions. It thus provides a way to describe forecast ensembles in a nonparametric way, without making any assumptions about the statistical behavior of the ensemble data. The physical details of the forecast model are eliminated, and the notion of statistical consistency is captured in an elementary way. Two applications of this result to ensemble analysis are presented. Ensemble stratification, the partitioning of an archive of ensemble forecasts into subsets using a discriminating criterion, is considered in the light of the above result. It is shown that certain stratification criteria can make the individual subsets of ensembles appear statistically inconsistent, even though the unstratified ensemble is statistically consistent. This effect is explained by considering statistical fluctuations of rank probabilities. A new hypothesis test is developed to assess statistical consistency of stratified ensembles while taking these potentially misleading stratification effects into account. The distribution of rank probabilities is further used to study the predictability of outliers, which are defined as events where the verification falls outside the range of the ensemble, being either smaller than the smallest, or larger than the largest ensemble member. It is shown that these events are better predictable than by a naive benchmark prediction, which unconditionally issues the average outlier frequency of 2/(K+1) as a forecast. Predictability of outlier events, quantified in terms of probabilistic skill scores and receiver operating characteristics (ROC), is shown to be universal in a hypothetical forecast ensemble. An empirical study shows that in an operational temperature forecast ensemble, outliers are likewise predictable, and that the corresponding predictability measures agree with the analytically calculated ones.

Cette thèse présente une méthode pour la détection hors-ligne de multiples ruptures dans des séries temporelles multivariées, et propose d'en exploiter les résultats pour estimer les relations de dépendance entre les variables du système. L'originalité du modèle, dit du Bernoulli Detector, réside dans la combinaison de statistiques locales issues d'un test robuste, comparant les rangs des observations, avec une approche bayésienne. Ce modèle non paramétrique ne requiert pas d'hypothèse forte sur les distributions des données. Il est applicable sans ajustement à la loi gaussienne comme sur des données corrompues par des valeurs aberrantes. Le contrôle de la détection d'une rupture est prouvé y compris pour de petits échantillons. Pour traiter des séries temporelles multivariées, un terme est introduit afin de modéliser les dépendances entre les ruptures, en supposant que si deux entités du système étudié sont connectées, les événements affectant l'une s'observent instantanément sur l'autre avec une forte probabilité. Ainsi, le modèle s'adapte aux données et la segmentation tient compte des événements communs à plusieurs signaux comme des événements isolés. La méthode est comparée avec d'autres solutions de l'état de l'art, notamment sur des données réelles de consommation électrique et génomiques. Ces expériences mettent en valeur l'intérêt du modèle pour la détection de ruptures entre des signaux indépendants, conditionnellement indépendants ou complètement connectés. Enfin, l'idée d'exploiter les synchronisations entre les ruptures pour l'estimation des relations régissant les entités du système est développée, grâce au formalisme des réseaux bayésiens. En adaptant la fonction de score d'une méthode d'apprentissage de la structure, il est vérifié que le modèle d'indépendance du système peut être en partie retrouvé grâce à l'information apportée par les ruptures, estimées par le modèle du Bernoulli Detector. / This thesis presents a method for the multiple change-points detection in multivariate time series, and exploits the results to estimate the relationships between the components of the system. The originality of the model, called the Bernoulli Detector, relies on the combination of a local statistics from a robust test, based on the computation of ranks, with a global Bayesian framework. This non parametric model does not require strong hypothesis on the distribution of the observations. It is applicable without modification on gaussian data as well as data corrupted by outliers. The detection of a single change-point is controlled even for small samples. In a multivariate context, a term is introduced to model the dependencies between the changes, assuming that if two components are connected, the events occurring in the first one tend to affect the second one instantaneously. Thanks to this flexible model, the segmentation is sensitive to common changes shared by several signals but also to isolated changes occurring in a single signal. The method is compared with other solutions of the literature, especially on real datasets of electrical household consumption and genomic measurements. These experiments enhance the interest of the model for the detection of change-points in independent, conditionally independent or fully connected signals. The synchronization of the change-points within the time series is finally exploited in order to estimate the relationships between the variables, with the Bayesian network formalism. By adapting the score function of a structure learning method, it is checked that the independency model that describes the system can be partly retrieved through the information given by the change-points, estimated by the Bernoulli Detector.

No description available.