Dynamic Chain Graph Models for Financial Time Series Networks: a Bayesian approach

• Ano de defesa: 2024
• Autor(a) principal: Miranda Neto, Milton
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Dynamic models; Hierarchical models; Modelos dinâmicos; Modelos gráficos probabilisticos; Modelos hierárquicos; Probabilistic graphical models; Séries temporais; Time series
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/104/104131/tde-23092024-104613/

Time series vectors with high dimensionality are increasingly frequent in several research fields, such as neuroscience and vehicle flow. Although this amount of data has been increasing in recent years, traditional statistical methodology for time series analysis is not capable of adequately modeling all this volume of information, as, in general, the models have several parameters in the order of O(d2), d being the dimension of the data set. This difficulty makes the joint modeling of the series computationally costly and, thus, infactible, leading many researchers to univariate modeling of the series, discarding the vectors dependence structure. In this work, we propose the Linear Dynamic Hierarchical Model of Chain Graphs, which can model a vector of time series together without excessive parameters and with a parallelization structure, making it computationally very efficient. This model generalizes the Dynamic Linear Model of Chain Graphs, as it allows the introduction of hierarchical levels of dependence between the state space vector while maintaining the modularization of the problem.