Time-varying auto-regressive models for count time-series (Q2044402)

This paper is concerned with the analysis of time series with values in $\mathbb{N}^*=\{0,1,2,\dots\}$. It is noteworthy that the proposed method was applied to a data series consisting of the spread of COVID 19 cases over three months in New York City. It is also to be noted that this paper considers a hitherto unworked time series model to the best of my knowledge, briefly, a model in which the coefficients are considered as random variables with an a priori distribution in a Bayesian setting. See the definition of the model used in this paper. Let $(\Omega,F,P)$ be a probability space, $T$ is a positive integer, for each $t=0,1,\dots,T: X_t:\Omega\to\mathbb{N}^*=\{0,1,2,\dots\}$ is a random variable, $p$ and $q$ are non-negative integers, $\mu:\Omega\times\mathbb{R}\to\mathbb{R}$, $a_{i}:\Omega\times\mathbb{R}\to\mathbb{R}$, $b_{j}:\Omega\times\mathbb{R}\to\mathbb{R}$ are random stochastic processes, $i=1,\dots,p$, $j=1,\dots,q$, $\lambda_{t}:\Omega\to\mathbb{R}$ is recursively defined as \[ \lambda_{t}(\omega)=\mu(\omega,(t/T))+\sum_{i=1}^{p}a_{i}(\omega,(t/T))X_{t-i}(\omega)+\sum_{j=1}^{q}b_{j}(\omega,(t/T))\lambda_{t-j}(\omega). \] $\mu$, $a_{i}$, and $b_{j}$ are assumed to satisfy the following equations \[ \mu(\omega,x) = \sum_{j=1}^{K_1}\alpha_{j}(\omega)B_{j}(x),\quad a_{i}(\omega,x) = \sum_{j=1}^{K_2}\theta_{i,j}(\omega)M_{i}(\omega)B_{j}(x),\;\; i=1,\dots,p, \] \[ b_{k}(\omega,x) = \sum_{j=1}^{K_3}\eta_{k,j}(\omega)M_{p+k}(\omega)B_{j}(x),\;\; k=1,\dots,q, \] where $\{B_{j} \mid j=1,\dots,\max(K_1,K_2,K_3)\}$ is a $B$-spline functions base and the coefficients are random variables such that the following restrictions are satisfied to ensure the existence and uniqueness of the series $(X_{t})$, \[ 0<\mu(\omega,x)<\infty, \;\;0<a_{i}(\omega,x)<\infty,\;\; 0<b_{k}(\omega,x)<\infty,\;\; i=1,\dots,p,\;\; k=1,\dots,q, \] \[ \sup_{y\in\mathbb{R}}\sum_{i=1}^{p}\sum_{k=1}^{q}(a_{i}(\omega,y)+b_{k}(\omega,y))<1,\;\; \omega\in\Omega, \; x\in\mathbb{R}. \] From the Bayesian point of view, it is also assumed that $\alpha_{j}$, $j=1,\dots,K_1$, has a Gaussian distribution with mean 0 and variance $c_1^2$ truncated at $[0,\infty)$, i.e., its density is \[ f(z)=1/\left(\int_0^\infty\exp(-u^2/(2c_1^2)) \cdot du\right)\cdot\exp(-z^2/(2c_1^2)) \cdot \1_{[0,\infty)}(z). \] \[ M_{i} = \frac{\tau_{i}}{\sum_{k=0}^{p}\tau_{i}},\;\; \tau_{i}\sim U(0,1),\;\; i=0,\dots,p+q,\;\; \theta_{i,j}\sim U(0,1),\;\; i=1,\dots,p,\;\; j=1,\dots,K_2, \] \[ \eta_{k,j}\sim U(0,1), \;\;k=1,\dots,q, \;\;j=1,\dots,K_3. \] With respect to the observed process, $(X_{t})$, it is assumed that for each $t\geq 1$ the conditional distribution of $X_{t}$ given $F_{t-1}=\{X_{i} \mid 0\leq i\leq t-1\}$, and $G_{t-1}=\{\lambda_i \mid 0\leq i\leq t-1\}$ is Poisson$(\lambda_{t})$, $t\geq 1$. $\lambda_0$, parameter of $X_0$, is assumed to be inverse-gamma$(d_1,d_1)$ distributed. The authors explain at some length a justification for the proposed model. Section 3 details how to calculate the maximum likelihood estimator for the model parameters. Section 4 provides an estimation of the posterior contraction rates of the analyzed models. TVBINGARCH(1,1) \[ \lambda_{t}(\omega)=\mu(\omega,(t/T))+a_1(\omega,(t/T))X_{t-i}(\omega)+b_2(\omega,(t/T))\lambda_{t-1}(\omega), \] and TVBARC($p$) \[ \lambda_{t}(\omega)=\mu(\omega,(t/T))+\sum_{i=1}^{p}a_{i}(\omega,(t/T))X_{t-i}(\omega). \] That is, how to increase the sample size to estimate the true parameters from the posterior distribution. The mathematical definition of this concept according to the authors is ``The posterior contraction rate at the true parameter $\theta_0\in A$ with respect to the semi-metric $d$ on $A$ is a sequence $\varepsilon_{T}\to 0$ such that \[ P_{\theta_0}^{(T)}\left(\prod_{T}(\theta:d(\theta,\theta_0)>M_{T}\varepsilon_{T}|X^{(T)})\right)\to 0 \text{ for every } M_{T}\to\infty, \] where A denotes the parameter space of $\theta_0$. Here $X^{(T)}$ stands for the complete dataset. $\prod_{T}(\theta\cdot|X^{(T)})$ is the posterior distribution with respect to a prior $\prod_{T}$. An analysis of the performance of the proposed method for estimating the parameters of particular cases of TVBINGARCH(1,1) and TVBARC($p$) models, in comparison with the results based on certain models with constant coefficients, is discussed in Section 5. It is noted that in the cases studied, the techniques defined in this paper have always shown the best results. In Section 6, the authors make a detailed analysis of the results obtained by applying their models and the others previously compared to data on the daily number of COVID 19 infections in NYC over six months (January 2020 to July 2020). This analysis is very relevant to see the effectiveness of different strategies to control the spread of the epidemic. The authors suggest that in future works on the subject it would be interesting to include in the analysis demographic data, geographic characteristics of the region studied, among others. The last section is concerned with extensive and detailed proofs of the main results of this paper.