Statistical analysis of new multivariate risk measures

  1. Palacios Rodríguez, Fátima
Dirigée par:
  1. Elena Di Bernardino Directeur/trice
  2. José María Fernández Ponce Directeur/trice
  3. Rosario Rodríguez Griñolo Directrice

Université de défendre: Universidad de Sevilla

Fecha de defensa: 20 mars 2017

Jury:
  1. Miguel Angel Sordo Díaz President
  2. Rosa Elvira Lillo Rodríguez Secrétaire
  3. Gianfausto Salvadori Rapporteur
  4. Fabrizio Durante Rapporteur
  5. Rafael Infante Macías Rapporteur

Type: Thèses

Teseo: 453989 DIALNET lock_openIdus editor

Résumé

As a consequence of the need for regulators to manage risk in various sectors, a riskbased methodology is undergoing a fast expansion. Over recent decades, this problem has been mostly addressed via a univariate approach. However, risks usually involve several random variables that are often non-independent. Therefore, it is crucial to work in a multivariate setting. On the other hand, phenomena are frequently characterized by extreme events. This thesis is fundamentally concerned with two problems: the de_nition of risk measures in a multivariate setting, and the estimation of multivariate risk measures by taking extreme events into account. Chapter 1 is an introductory chapter. We present the state-of-art of the notion of multivariate risk measures. The main results in Copula Theory, Extreme Value Theory, and Stochastic Orders, which are useful in this work, are also provided. Two new multivariate risk measures are introduced in Chapter 2. Several interesting properties and, characterizations under Archimedean copulas, are studied for the proposed risk measures. Furthermore, semi-parametric estimators for the new measures are obtained and are then exempli_ed considering simulated data and a real insurance data-set. Chapter 3 deals with the non-parametric extreme estimation procedure of the multivariate measures proposed in Chapter 2. For this purpose, we _rst analyse the tail behaviour of the conditional distributions that de_ne the aforementioned measures. The main result is given by the Central limit Theorem of the extreme estimators. The performance of the extreme estimators is evaluated in simulated data and for a real rainfall data-set. The multivariate risk measure associated with the Component-wise Excess (C.-E.) design realization given by Salvadori et al. (2011) is outlined in Chapter 4. The explicit expression of the measure for Archimedean copulas is obtained. In addition, an extreme estimation procedure for the C.-E. design realization is provided and the asymptotic behaviour of the proposed estimators is studied. Finally, the estimators for the C.-E. design realization are applied to simulated data and a real dam data-set.