Small sample effects in time series analysis: A new asymptotic theory and a new estimate (Q1113599): Difference between revisions

For a real-valued stationary process X(t) in discrete time, with zero mean and spectral density \(f_{\theta}(\lambda)\), where \(\theta\) is a p- dimensional parameter, if observations \(X_ 0,...,X_{T-1}\) are available \textit{P. Whittle} [Ark. Mat. 2, 423-434 (1953; Zbl 0053.410)] suggested to approximate the likelihood function under normality by \[ (4\pi)^{-1}\int^{\pi}_{-\pi}\{\log 4\pi^ 2f_{\theta}(\lambda)+I_ T(\lambda)f_{\theta}^{- 1}(\lambda)\}d\lambda, \] where \(I_ T(\lambda)=(2\pi T)^{-1} | \sum^{T-1}_{t=0}X_ t \exp (-i\lambda t)|^ 2\) is the periodogram. Although the Whittle estimate of \(\theta\) is asymptotically Fisher efficient, the small sample behaviour may be poor if \(f_{\theta}(\lambda)\) contains peaks. The model \(f_{\theta_ 0}=g_ Tf^*_{\theta_ 0}\) is introduced for the spectral density, with a ``strong peak'' part \(g_ T\) (whose peaks may increase with T) and a part \(f^*_{\theta_ 0}\) which is independent of T and whose parameter \(\theta_ 0\) has to be estimated. The tapered version of the periodogram \[ \{2\pi H_{2T}\}^{-1} | \sum^{T-1}_{t=0}h_{tT} X_ t \exp (-i\lambda t)|^ 2 \] is considered, where \(H_{2T}\) is the sum of the squares of the \(h_{tT}\), to avoid a leakage effect. The paper contains: a) A proof of the consistency of the (exact) MLE of \(\theta_ 0\); b) A proof of the fact that the corresponding conditional MLE may be inconsistent; c) An analysis of data tapers; d) A proof that a suitable choice of the data tapers in the Whittle estimates also leads to an efficient estimate of \(\theta_ 0\); e) An analysis of the classical nontapered Whittle estimate.