I am a grad student in statistics
at Penn State University and
I have the following time-series problem
that
i have been working on for a
while with
no luck. Eevn if anyone can tell
me that it is not possible to solve
a problem like this, this is
fine also. Atleast I will know
that I should stop trying.
thanks.
Here's the problem :
I have the following MA(5) model ( this is the
model
you get if you write Harvey's bsm in the single
source
of error form but that doesn't matter ) :
y_t = epslion_t +
(alpha1 + alpha2 - alpha3 -1)*epsilon_t-1
+
(alpha2+alpha3)*epsilon_t-2 + (alpha2)*epsilon_t-3
+ (alpha2 +
alpha3-1)*epsilon_t-4 + (1-alpha1-alpha3)*epsilon_t-5.
So, it is an
MA(5) model but it only has 3 unknown parameters.
I
usually use ox to estimate ma models but for this
special ma
model,
I think I can write the likelihood myself and then
estimate the parameters
by, say, minimizing the residual variance in
the likelhood etc.
my problem is
the following : I have
noticed in the software package called
ox, if they get non invertible
ma parameters after the minimization
, they then deal with the
invertibility issue by
finding the inverse
roots which are
greater than 1, flipping these roots, and then
rebuilding the
ploynomial. ( they have functions like
polyroots which finds the roots
and polymake which
can rebuild the polynomial given the new roots ).
I
think that they can do this because they don't have this
dependence
between the parameters that I have. I don't think
I
can do what they do, in terms of fixing the
invertibility
problem, because of the dependence between my parameters
?
I can flip the roots but when I reconstruct the
polynomial
using polymake I will lose the
relationships.
If I am correct in what is say, do you
think it is possible
in some way to estimate an invertible model of
this strange type ? I
really appreciate your answer. thank
you
very much for taking the time to read this.
I really
appreciate your
help.
Mark