We study in this dissertation Generalized Autoregressive Conditionally Heteroskedastic (GARCH) time series. The research focuses on squared GARCH sequences. Our main results are as follows: 1. We compare three methods of constructing confidence intervals for sample autocorrelations of squared returns modeled by models from the GARCH family. We compare the residual bootstrap, block bootstrap and subsampling methods. The residual bootstrap based on the standard GARCH(l,1) model is seen to perform best. Confidence intervals for cross-correlations of a bivariate GARCH model are also studied. 2. We study a test to discriminate between long memory and volatility changes in financial returns data. Finite sample performance of the test is examined and compared using various variance estimators. The Bartlett kernel estimator with truncation lag determined by a calibrated bandwidth selection procedure is seen to perform best. The testing procedure is robust to various GARCH-type models. 3. We propose several methods of on-line detection of a change in unconditional variance in a conditionally heteroskedastic time series. We follow a paradigm in which the first m observations are assumed to follow a stationary process and the monitoring scheme has asymptotically controlled probability of falsely rejecting the null hypothesis of no change. Our theory is applicable to broad classes of GARCH-type time series and relies on a strong invariance principle which holds for the squares of observations generated by such models. Practical implementation of the procedures is proposed and the performance of the methods is investigated by a simulation study.

Dans cette thèse, nous nous intéressons à l'estimation de modèles conditionnellement hétéroscédastiques (CH) sous différentes hypothèses. Dans une première partie, en modifiant l'hypothèse d'identification usuelle du modèle, nous définissions un estimateur de quasi-maximum de vraisemblance (QMV) non gaussien et nous montrons que, sous certaines conditions, cet estimateur est plus efficace que l'estimateur du quasi maximum de vraisemblance gaussien. Nous étudions dans une deuxième partie l'inférence d'un modèle CH dans le cas où le processus des innovations est distribué selon une loi alpha stable. Nous établissons la consistance et la normalité asymptotique de l'estimateur du maximum de vraisemblance. La loi alpha stable n'apparaissant que comme loi limite, nous étudions ensuite le comportement de ce même estimateur dans le cas où la loi du processus des innovations n'est plus une loi alpha stable mais est dans le domaine d'attraction d'une telle loi. Dans la dernière partie, nous étudions l'estimation d'un modèle GARCH lorsque le processus générateur de données est un modèle CH dont les coefficients sont sujets à des changements de régimes markoviens. Nous montrons que cet estimateur, dans un cadre mal spécifié, converge vers une pseudo vraie valeur et nous établissons sa loi asymptotique. Nous étudions cet estimateur lorsque le processus observé est stationnaire mais nous détaillons également ses propriétés asymptotiques lorsque ce processus est non stationnaire et explosif. Par des simulations, nous étudions les capacités prédictives du modèle GARCH mal spécifié. Nous déterminons ainsi la robustesse de ce modèle et de l'estimateur du QMV à une erreur de spécification de la volatilité. / In this thesis, we focus on the inference of conditionally heteroskedastic models under different assumptions. This thesis consists of three parts and an introductory chapter. In the first part, we use an alternate identification assumption of the model and we define a non Gaussian quasi maximum likelihood estimator. We show that, under certain conditions, this estimator is more efficient than the Gaussian quasi maximum likelihood estimator. In a second part, we study the inference of a conditionally heteroskedastic model when the process of the innovations is distributed as an alpha stable law. We establish the consistency and the asymptotic normality of the maximum likelihood estimator. Since the alpha stable laws appear in general as a limit, we then focus of the behavior of this same estimator when the law of the innovation process is not stable but in the domain of attraction of a stable law. In the last part of this thesis, we study the estimation of a GARCH model when the data generating process is a conditionally heteroskedastic model whose coefficients are subject to Markov switching regimes. We show that, in a missspecified framework, this estimator converges toward a pseudo true value and we establish its asymptotic properties when this process is non stationary and explosive. Through simulations, we investigate the predictive ability of the missspecified GARCH model. Thus we determinate the robustness of the model and of the estimator of the quasi maximum likelihood to the missspecification of the volatility

No description available.

In an increasingly competitive economic environment, as in the current global context, risk management becomes essential for the survival of companies and investment portfolio managers. Both companies and managers need to have a model that can be able to quantify the risks inherent in their investments in the best possible way in order to guide them in making decisions to get the highest expected return on their investments. Currently, there are several heterogeneous models which seek to quantify risk, making the choice of a particular model very complex. In order to confront and find models that can serve, efficiently, to the quantification of risk, the objective of this research is to compare the predictive ability of five models of conditional heteroskedasticity by estimating the Value-at-Risk, assuming eight different statistical probability distributions, for the series of financial ratios of the capital market of the five largest emerging countries: Brazil, Russia, India, China and South Africa, in the period between February 26, 2001 and December 31, 2015. For this goal was achieved, were held predictions of Value-at-Risk for 50 steps ahead, for all competing models in the study, with adjustment of parameters at every step. Since all the forecasts have been computed for every steps forward, it was possible to compare predictive ability of competing models studied by means of some loss functions. The evidences suggests that heterocedastic Component GARCH is preferable, to make predictions of Value-at-Risk, to all other competing models, however the distribution of statistical probability that this model uses interferes too much in the results of forecasts obtained by it. The data for each financial index studied showed to adapt themselves to a particular different type of probability density function, not reflecting a distribution which can be considered superior to all other. Thus, the results do not provide a single and ideal tool for use in the risk measurement, of generalized form, for all capital markets of emerging countries studied, only provide specific tools to be used in each financial index individually. The results found can be used for the purposes previously described or to elaborate statistical formulas that combine different models estimated in order to get better volatilities forecast measures so that it can measure, more precisely, the market risks. / Em um ambiente econômico cada vez mais competitivo, como é no atual contexto mundial, a gestão de risco torna-se indispensável para a sobrevivência de empresas e de gestores de carteiras de investimento. Tanto as empresas quanto os gestores precisam de um modelo que seja capaz de quantificar os riscos inerentes aos seus investimentos financeiros da melhor maneira possível, de forma a orientá-los na tomada de decisões para que obtenham o maior retorno esperado de seus investimentos. Atualmente, existem inúmeros modelos heterogêneos que buscam quantificar riscos, tornando a escolha de um determinado modelo bastante complexa. Com o intuito de confrontar e encontrar modelos que possam servir, de forma eficiente, à quantificação de riscos, o objetivo desta pesquisa é o de comparar a capacidade preditiva de cinco modelos de heterocedasticidade condicional através da estimação do Value-at-Risk, levando em consideração oito distribuições de probabilidade estatística diferentes, para as séries de índices financeiros do mercado de capitais dos cinco maiores países emergentes: Brasil, Rússia, Índia, China e África do Sul, no período compreendido entre 26 de fevereiro de 2001 e 31 de dezembro de 2015. Para alcançar tal objetivo, realizaram-se previsões do Value-at-Risk para 50 passos à frente, em todos os modelos concorrentes em estudo, com reajuste dos parâmetros a cada passo. Uma vez que todas as previsões foram computadas para todos os passos à frente, foi possível realizar a comparação da capacidade preditiva dos modelos concorrentes estudados por meio de determinadas funções de perda específicas. As evidências encontradas sugerem que o modelo heterocedástico Component GARCH é preferível, para realizar previsões do Value-at-Risk, a todos os outros modelos concorrentes, porém a distribuição de probabilidade estatística que este modelo utiliza interfere demasiadamente nos resultados das previsões obtidas por ele. Os dados de cada índice financeiro estudado mostraram-se adequar-se a um determinado tipo de função de densidade de probabilidade diferente, não refletindo uma distribuição que possa ser considerada superior a todas as outras. Deste modo, os resultados encontrados não oferecem uma ferramenta única e ideal para ser utilizada na mensuração de risco, de forma generalizada, para todos os mercados de capitais dos países emergentes estudados, apenas fornecem ferramentas pontuais para serem utilizadas em cada índice financeiro de forma individual. Os resultados obtidos podem servir para os fins descritos anteriormente ou para elaborar fórmulas estatísticas que combinem diferentes modelos estimados com a finalidade de obter melhores medidas de previsão de volatilidades para que se possa mensurar, de forma mais precisa, os riscos de mercado.

Given the wide range of macroeconomic, financial and econometric frameworks commonly used to accommodate uncomfortable empirical evidence associated with the Forex market, this article aims to model and predict the monthly variation in American Dollar-Brazilian Real exchange rate, from January 2000 to December 2009, based on asset pricing theory. Wang (2008) and Engel and West (2005) are closer to ours, in terms of fundamentals of finance, while methodologically, we are close to Chong, Chung and Ahmad (2002) and da Costa et al. (2010). Our work is relevant to the empirical literature, since the prediction results are better than the random walk approach ones. The prediction error is about 5% and 14% for the exchange rate variation and in level, respectively. In 57.5% of the changes, our model predicts the correct change direction. The main contribution based on this framework, already used to understand the Forward Premium Puzzle for advancedeconomies, consists in the derivation and the implications of a system of linear relationships characterized by a Bivariate Generalized Autoregressive Conditional Heteroskedasticity-in-Mean (GARCH-M), useful empirically, once we have extracted a time series for a Stochastic Discount Factor (SDF) able to price the covered and the uncovered trading with U.S. Government bonds. The results suggest to the theoretical literature that, at least for monthly frequency, one should not omit the temporal variation of conditional moments of the second order. The hypothesis about the lognormal distribution of discounted returns and a parsimonious specification for conditional Heteroskedastic models can influence the predictive power of SDF, as well as the effects of the inclusion of risk premium. / Diante da vasta gama de arcabouÃos macroeconÃmicos, economÃtricos e financeiros que visam acomodar evidÃncias empÃricas desconfortÃveis associadas ao mercado cambial, este artigo visa modelar e prever a variaÃÃo mensal entre as moedas real brasileiro e dÃlar americano, de janeiro de 2000 a dezembro de 2009, baseado na teoria de apreÃamento de ativos. Este estudo agrega-se à literatura empÃrica, ao obter resultados preditivos superiores a um modelo de passeio aleatÃrio, com erros de previsÃo da ordem de grandeza de 5% e 14% para depreciaÃÃo e para o cÃmbio em nÃvel, respectivamente, e um acerto em 57,5% das vezes com relaÃÃo à direÃÃo da variaÃÃo cambial. Alinhado em fundamentos a Wang (2008) e Engel e West (2005) e metodologicamente a Chong, Chung e Ahmad (2002) e da Costa et al. (2010), a principal contribuiÃÃo no uso deste arcabouÃo, jà utilizado no entendimento do Forward Premium Puzzle para economias avanÃadas, consiste na derivaÃÃo e nas implicaÃÃes de um sistema de relaÃÃes lineares caracterizado por um Generalized Autoregressive Conditional Heteroskedasticity-in- Mean (GARCH-M) bivariado, o qual pode ser testÃvel, a partir da extraÃÃo via componentes principais da sÃrie temporal para um Fator EstocÃstico de Desconto capaz de apreÃar operaÃÃes coberta e descoberta de aquisiÃÃo de tÃtulos do governo americano. Os resultados sugerem, ainda, à literatura teÃrica que, ao menos para frequÃncia mensal, nÃo se deve desprezar a variaÃÃo temporal dos momentos condicionais de segunda ordem. A hipÃtese sobre a distribuiÃÃo lognormal dos retornos descontados e uma especificaÃÃo parcimoniosa para modelos de heterocedasticidade condicional podem prejudicar a capacidade preditiva associada do Fator EstocÃstico de Desconto, assim como os efeitos da incorporaÃÃo do prÃmio de risco.

Nesta tese, desenvolvemos refinamentos assintóticos em modelos lineares generalizados heteroscedásticos (Smyth, 1989). Inicialmente, obtemos a matriz de covariâncias de segunda ordem dos estimadores de máxima verossimilhança corrigidos pelos viés de primeira ordem. Com base na matriz obtida, sugerimos modificações na estatística de Wald. Posteriormente, derivamos os coeficientes do fator de correção tipo-Bartlett para a estatística do teste gradiente. Em seguida, obtemos o coeficiente de assimetria assintótico da distribuição dos estimadores de máxima verossimilhança dos parâmetros do modelo. Finalmente, exibimos o coeficiente de curtose assintótico da distribuição dos estimadores de máxima verossimilhança dos parâmetros do modelo. Analisamos os resultados obtidos através de estudos de simulação de Monte Carlo. / In this thesis, we have developed asymptotic refinements in heteroskedastic generalized linear models (Smyth, 1989). Initially, we obtain the second-order covariance matrix for the maximum likelihood estimators corrected by the bias of first-order. Based on the obtained matrix, we suggest changes in Wald statistics. In addition, we derive the coeficients of the Bartlett-type correction factor for the statistical gradient test. After, we get asymptotic skewness of the distribution of the maximum likelihood estimators of the model parameters. Finally, we show the asymptotic kurtosis coeficient of the distribution of the maximum likelihood estimators of the model parameters. Monte Carlo simulation studies are developed to evaluate the results obtained.

Nesta tese, desenvolvemos refinamentos assintóticos em modelos lineares generalizados heteroscedásticos (Smyth, 1989). Inicialmente, obtemos a matriz de covariâncias de segunda ordem dos estimadores de máxima verossimilhança corrigidos pelos viés de primeira ordem. Com base na matriz obtida, sugerimos modificações na estatística de Wald. Posteriormente, derivamos os coeficientes do fator de correção tipo-Bartlett para a estatística do teste gradiente. Em seguida, obtemos o coeficiente de assimetria assintótico da distribuição dos estimadores de máxima verossimilhança dos parâmetros do modelo. Finalmente, exibimos o coeficiente de curtose assintótico da distribuição dos estimadores de máxima verossimilhança dos parâmetros do modelo. Analisamos os resultados obtidos através de estudos de simulação de Monte Carlo. / In this thesis, we have developed asymptotic refinements in heteroskedastic generalized linear models (Smyth, 1989). Initially, we obtain the second-order covariance matrix for the maximum likelihood estimators corrected by the bias of first-order. Based on the obtained matrix, we suggest changes in Wald statistics. In addition, we derive the coeficients of the Bartlett-type correction factor for the statistical gradient test. After, we get asymptotic skewness of the distribution of the maximum likelihood estimators of the model parameters. Finally, we show the asymptotic kurtosis coeficient of the distribution of the maximum likelihood estimators of the model parameters. Monte Carlo simulation studies are developed to evaluate the results obtained.

Regression is a central problem in statistics and machine learning with applications everywhere in science and technology. In probabilistic regression the relationship between a set of features and a real-valued target variable is modelled as a conditional probability distribution. There are cases where this distribution is very complex and not properly captured by simple approximations, such as assuming a normal distribution. This thesis investigates how conditional Generative Adversarial Networks (GANs) can be used to properly capture more complex conditional distributions. GANs have seen great success in generating complex high-dimensional data, but less work has been done on their use for regression problems. This thesis presents experiments to better understand how conditional GANs can be used in probabilistic regression. Different versions of GANs are extended to the conditional case and evaluated on synthetic and real datasets. It is shown that conditional GANs can learn to estimate a wide range of different distributions and be competitive with existing probabilistic regression models.