Sequential point estimation of parameters in a threshold AR(1) model (Q1613666)

This paper investigates the sequential estimator \(\hat \theta _n\), \(n=1,2,\dots \), of the vector of parameters \(\theta =(\theta _1,\theta _2)\) of a stationary ergodic threshold autoregressive process \(\{X_i\}_{i=0}^\infty , \) defined by \[ X_i=\theta _1X_{i-1}^++\theta _2X_{i-1}^-,\quad i=1,2,\dots , \] where \(x^+=\max (x,0)\), \(x^-=\min (x,0)\) and \( \{ \varepsilon _i\}_{i=1}^\infty \) is a sequence of i.i.d. random variables with zero mean and finite variance \(\sigma ^2>0\), which is unknown and estimated with the usual sequential estimator \(\hat \sigma _n^2\). The stopping rule \(T\) used to determine the sample size \(n\) is assumed to depend on the constant \(q\) from the definition of the loss function \[ L_n=qn^{-1}(\hat \theta _n-\theta)'\text{diag}(\sum _{i=1}^n{X_{i-1}^+}^2, \sum _{i=1}^n{X_{i-1}^-}^2)(\hat \theta _n-\theta)+n, \] capturing the relative importance of the quadratic error w.r.t. the sampling cost: \[ T (q)=\max (n_q,\lceil \hat \sigma _n\sqrt {2q}\rceil), \] where \(\lceil \rceil \) denotes the ceiling integer and \(n_q\) is an initial sample size possibly depending on \(q\). Several fundamental properties of \(T\) and \(\hat \theta _n\) are established and formulated in two theorems. Theorem~1 states conditions sufficient for the stopping rule \(T\) to be asymptotically efficient, and for the sequential estimator \(\hat \theta _n\) to be asymptotically risk efficient. Its proof relies on optimal stopping theory, mixing theory, and on general properties of threshold autoregressive processes. Some of the results established in the course of the proof are of independent interest. This is particularly true for Lemma~2.2, which in the case of a geometrically \(\beta \)-mixing stationary process \(\{Y_i\}_{i=0}^\infty \) establishes the rate of convergence for the lower-tail probability of \(\{n^{-1}\sum _{i=1}^n Y_i\}_{i=1}^\infty \), and for Proposition~2.1, where that rate of convergence transfers also to \(\{n^{-1}\sum _{i=1}^n{X_{i-1}^+}^2 \}_{i=1}^\infty \) and \(\{n^{-1}\sum _{i=1}^n{X_{i-1}^-}^2 \}_{i=1}^\infty \). In the proof of Theorem~1, the stopping rule \(T\) is shown to be a.s. asymptotically increasing as \(\sigma \sqrt {2q}\). That result is then further detailed in Theorem~2, establishing that for \(q\to \infty\): (a) \((\sqrt{\sigma\sqrt{2q}})^{-1}(T(q)-\sigma\sqrt{2q})\) is asymptotically normal with zero mean, (b) \(ET(q)-\sigma \sqrt {2q}=\) const + \(o(1)\). This theorem is proved using nonlinear renewal theory. The paper is an interesting contribution to the rich literature on sequential estimation in autoregressive processes. During the last two decades, more than twenty authors have been actively involved in that area. In this context, it is a bit regrettable that among the references provided by the authors of the present paper all the four items concerning sequential estimation in autoregressive processes are their own works from the years 1987-1996. Also some other important papers would be worth mentioning, at least from recent years, e.g., \textit{V. Konev} and \textit{S. Pergamenshchikov}, Ann. Stat. 25, No. 5, 2127-2163 (1997; Zbl 0887.62087), or \textit{A.K. Basu} and \textit{J.K. Das}, Sequential Anal. 16, No. 1, 1-24 (1997; Zbl 0879.62071).