In a paper by Lemos and Stokes (1998) they review a number of classic tests
for nonlinearity (with respect to serial order correlations) and for their
simulations they generate data from a number of nonlinear time-series
divided into two specific groupings.
In the first group, they list: (i) AR, (ii) Bilinear, (iii) Threshold
Auto-regressive (TAR), (iv) Sign Autoregressive (SGN), and (v) Nonlinear
Autoregressive (NAR) processes.
Models (ii) thru (v) are clearly nonlinear; however, I don't see how (i) is
nonlinear...
They use the following AR process given by:
y(t) = 0.6*y(t-1) + epsilon(t)
where the innovations, epsilon(t) ~ N(0,1).
This is a stable AR process and does not contain multiplicative interaction
terms between the past values and/or innovations...
any comments?
p
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
