Flexible bivariate Poisson integer-valued GARCH model (Q2027225)

The current paper, motivated by \textit{J. Lakshminarayana} et al. [Commun. Stat., Theory Methods 28, No. 2, 267--276 (1999; Zbl 0919.62046)] and in order to provide enough flexibility for dependence modeling, introduces initially three types of Bivariate Poisson (BP) distributions. The probability mass function of each one of them is a product of Poisson marginal with different multiplicative factor, namely the Gaussian, the Frank and the Farlie-Gumbel-Morgenstern factor, respectively. Afterward, the flexible bivariate Poisson integer-valued generalized autoregressive conditional heteroscedastic (INGARCH) models are introduced and some of the properties of the process (stationarity, ergodicity) are discussed. Since the log-likelihood function of each model is very complicated, the maximization by parts (MBP) algorithm, proposed by \textit{P. X. K. Song} et al. [J. Am. Stat. Assoc. 100, No. 472, 1145--1158 (2005; Zbl 1117.62429)], the modified maximization by parts (MMBP) algorithm, studied by \textit{Y. Liu} and \textit{R. Luger} [Comput. Stat. Data Anal. 53, No. 6, 2284--2297 (2009; Zbl 1453.62140)], together with the method based on the R package Template Model Builder (TMB) are employed. The consistency and the asymptotic normality of estimates are also established, while a simulation study evaluates the performance of the estimators with MBP, MMBP and TMB methods. Finally, the flexibility of the new models is illustrated with a real time series data set. At the same time, their fit is compared with the respective of previous known models.