Time series analysis via rank order theory: Signed-rank tests for ARMA models (Q1182743): Difference between revisions

For \(n\) time series observations, a linear, serial signed-rank statistic of order \(k\) is \(\displaystyle (n-k)^{-1}\sum_{t=k+1}^ n a_ +^{(n)}\), where \(a_ +^{(n)}\) is a score function of the signed ranks of the observations, or of some adequate transformation. Then various functional forms of the \(a_ +^{(n)}\) will produce, for example, the ``classical run test statistics'', the ``turning point'' or ``runs up and down statistic'', or a signed version of the Spearman-Wald-Wolfowitz rank serial correlation coefficient. A particular type of these statistics are the \(f\)-rank autocorrelation coefficients, analyzed by the same authors [Ann. Stat. 16, 402-432 (1988; Zbl 0659.62111)]. In the present paper an asymptotic distribution theory is developed, and then used to derive asymptotically locally optimal tests (in the maximin sense) for testing an ARMA model against other ARMA models. The asymptotic relative efficiencies of the proposed procedures with respect to each other, and with respect to their normal theory counterparts, are provided. In one example, the AR(1) model with autoregressive coefficient \(\rho=\rho_ 0\in(-1,1)\) and unspecified, symmetric innovation density against an alternative \(\rho>\rho_ 0\), and innovation density \(f\) is studied. In another example the alternative is an \(\text{ARMA}(p,q)\) model with \(1\leq p\leq 2\) and \(q\leq 1\); and still in another example the innovations have a Cauchy density.