M-Estimation for Nearly Non-Stationary Autoregressive Time Series.
Abstract
The nearly nonstationary first order autoregression is a sequence of processes where the autoregressive coefficient tends to 1 as n approaches infinity. M-estimates of the autoregressive coefficient are considered. The process is allowed to be nongaussian, but a 2 + delta moment condition is assumed. The limiting distribution is not the usual normal limit but is characterized as a ratio of two stochastic integrals. The asymptotically most efficient M-estimate is not given by maximum likelihood. However, it is shown that the loss of efficiency in using maximum likelihood is no worse than about 20% whereas the usual least squares estimator can have arbitrarily low efficiency. Keywords: M estimation; time series, autoregressive; non stationary.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 1987
- Accession Number
- ADA181184
Entities
People
- Dennis D. Cox
- Isabel Llatas
Organizations
- University of Washington