This thesis addresses two key areas in multivariate time series analysis: modeling multivariate count data and developing theoretical conditions for neural network-based autoregressive moving average processes. First, we introduce the multivariate int...
This thesis addresses two key areas in multivariate time series analysis: modeling multivariate count data and developing theoretical conditions for neural network-based autoregressive moving average processes. First, we introduce the multivariate integer-valued generalized autoregressive conditional heteroscedastic (MINGARCH) model to handle multivariate time series of counts with possible negative autocorrelations and cross-correlations. In addition, we explore an inferential procedure utilizing the quasi-maximum likelihood estimator (QMLE). We also introduce the change point test based on QMLE due to its practical significance. Furthermore, inferences based on the minimum density power divergence estimator (MDPDE), a robust estimator, are studied along with change point detection methods using MDPDE. We validate these methods through a Monte Carlo simulation and a case study on weekly syphilis cases in the United States.
Secondly, we establish the stationary condition for multivariate neural network autoregressive moving average (NN-ARMA) processes, showing that these conditions ensure ergodicity through absolute regularity.