| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.92 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
A Teoria de Valores Extremos (TVE) surge naturalmente na presença de observações muito elevadas ou muito reduzidas. A TVE tem sido usada em diversas áreas, tais como seguros, hidrologia e ambiente. A distribuição generalizada de valores extremos e a distribuição de Pareto generalizada (GPD) destacam-se como sendo as distribuições usadas na modelação de observações extremas. A GPD foi introduzida na literatura por Pickands (1975) como a distribuição limite da amostra de excessos ou excedências acima de um limiar suficientemente elevado. A estimação dos parâmetros da GPD é um assunto abordado frequentemente na literatura. Métodos clássicos como a Máxima Verosimilhanc¸a, Momentos ou Momentos Ponderados de Probabilidade apresentam certas limitações, pelo que têm vindo a ser desenvolvidos outros métodos de estimação ao longo dos tempos, tais como o Método dos Percentis Elementares, proposto por Castillo & Hadi (1997), e os estimadores empirical bayes desenvolvidos por Zhang & Stephens (2009). Certos estimadores apresentam boas propriedades quando a GPD tem cauda pesada, enquanto que outros são preferíveis quando a distribuição tem cauda leve. Este facto sugere a possibilidade de desenvolver procedimentos que permitam escolher, de forma adaptativa, os melhores estimadores para os parâmetros da GPD, consoante o peso de cauda da distribuição subjacente. Com este trabalho, pretendeu-se mostrar que a combinação de métodos de estimação pode resultar na obtenção de procedimentos mais eficazes. Para este efeito, recorreu-se a algoritmos de otimização não linear, técnicas de ajustamento polinomial e metodologias de classificação, como o perceptrão. A avaliação destes procedimentos foi feita com recurso a um estudo de simulação, que produziu resultados muito promissores em termos de viés e raiz quadrada do erro quadrático médio. Os procedimentos foram aplicados a dados reais das áreas do desporto (triplo salto) e da atividade seguradora (danos corporais associados ao ramo automóvel).
Extreme value theory (EVT) comes naturally whenever we are faced with very large or very small observations. EVT has been applied to several areas such as insurance, hydrology and environment. The generalized extreme value distribution and the generalized Pareto distribution (GPD) stand out as the distributions used for modeling extreme observations. The GPD was introduced in the literature by Pickands (1975) as the limit distribution of the excesses or exceedances above a sufficiently high threshold. The estimation of the parameters of the GPD is a topic that is frequently addressed in the literature. Classical methods like maximum likelihood, moments or probability weighted moments have several limitations and consequently other estimation methods have been developed such as the Elemental Percentile Method, proposed by Castillo & Hadi (1997) and the empirical bayes estimators developed by Zhang & Stephens (2009). Some estimators have good properties when the GPD is heavy-tailed, whereas others perform better for light tails. These features suggest the possibility of developing procedures that enable the choice, in an adaptive manner, of the best estimation method for the parameters of the GPD according to the tail weight of the underlying distribution. The purpose of this thesis is to show that the combination of estimation methods may result in better procedures. For that purpose, non-linear optimization algorithms were used, as well as polynomial fits and classification techniques, such as the perceptron. The performance of these techniques was assessed by means of a simulations study, which produced very good results both in terms of bias and root mean squared error. The procedures were applied to sports and insurance datasets.
Extreme value theory (EVT) comes naturally whenever we are faced with very large or very small observations. EVT has been applied to several areas such as insurance, hydrology and environment. The generalized extreme value distribution and the generalized Pareto distribution (GPD) stand out as the distributions used for modeling extreme observations. The GPD was introduced in the literature by Pickands (1975) as the limit distribution of the excesses or exceedances above a sufficiently high threshold. The estimation of the parameters of the GPD is a topic that is frequently addressed in the literature. Classical methods like maximum likelihood, moments or probability weighted moments have several limitations and consequently other estimation methods have been developed such as the Elemental Percentile Method, proposed by Castillo & Hadi (1997) and the empirical bayes estimators developed by Zhang & Stephens (2009). Some estimators have good properties when the GPD is heavy-tailed, whereas others perform better for light tails. These features suggest the possibility of developing procedures that enable the choice, in an adaptive manner, of the best estimation method for the parameters of the GPD according to the tail weight of the underlying distribution. The purpose of this thesis is to show that the combination of estimation methods may result in better procedures. For that purpose, non-linear optimization algorithms were used, as well as polynomial fits and classification techniques, such as the perceptron. The performance of these techniques was assessed by means of a simulations study, which produced very good results both in terms of bias and root mean squared error. The procedures were applied to sports and insurance datasets.
Descrição
Tese de mestrado, Estatística e Investigação Operacional (Estatística e Investigação Operacional), Universidade de Lisboa, Faculdade de Ciências, 2017
Palavras-chave
Distribuição de Pareto generalizada Teoria de valores extremos Métodos para estimação de parâmetros Métodos de decisão Teses de mestrado - 2017
