On some integer-valued autoregressive moving average models (Q1333195)

Let \(\{\varepsilon_ t\}\) be a sequence of i.i.d. integer-valued random variables. For \(\alpha \in (0,1)\) and \(\theta \in [0,1]\) define the operator \(A_{\alpha,\theta}\) by \[ A_{\alpha,\theta} \circ\varepsilon_{t - 1} = \sum_{i = 1}^{\varepsilon_{t - 1}} Y_ i W_ i, \] where \(\{Y_ i\}\) and \(\{W_ i\}\) are independent sequences of i.i.d. random variables independend of \(\varepsilon_{t - 1}\) such that \(Y_ i\) is a Bernoulli variable with parameter \(\alpha\) and \[ P(W_ i = x) = [1 - (1 - \alpha)\theta][(1-\alpha)\theta]^{x- 1}. \] The authors define an integer-valued moving average process of order 1, INMA(1), by \[ X_ t = \varepsilon_ t + A_{\alpha,\theta} \circ \varepsilon_{t - 1}. \] Time reversibility of the model is investigated. A necessary and sufficient condition is given under which \(E(X_ t \mid X_{t - 1})\) is a linear function of \(X_{t - 1}\). Special choices of the distribution of \(\varepsilon_ t\) lead to Poisson geometric and negative binomial processes \(\{X_ t\}\). The model is generalized to INMA\((q)\) processes in an obvious way. Further, an ARMA\((1,q)\) process is introduced. It consists of two components. The first one is an INAR(1) component \(Y_ t = A_{\alpha,\theta} \circ Y_{t -1} + \varepsilon_ t\) and the second one is an INMA\((q)\) component with the same innovations \(\{\varepsilon_ t\}\). The last part of the paper describes vector integer-valued AR(1) processes and briefly sketches areas of applications.