this question is for people who use maxbfgs and
ssfpack. I am
a grad student and dealing with
single
source of error state space models. it is well
known in the state space literature
that the random walk plus noise single source
model is equivalent to
an arima(0,1,1) model where beta1 = theta1 + 1.
( beta1 is the term
multiplying the error term in the structural
model ). also,
it is known that the local level plus trend
model is equivalent
to an arima(0,2,2) model where beta1 = (1 -
theta2) and
beta2 = 1 + theta1 + theta2. I wrote code
using ox, arfimasim,
and ssfpack to generate arima(0,1,1) models and estimate
the parameters using
arfima.estimate and ssfpack and
the parameters I get map over to each other
perfectly
in the sense that beta1 hat = theta1 hat + 1.
yet, when I do the same
thing by generating (0,2,2)
models and estimating them
in both worlds, the
parameters do not map to each other and they
aren't
even close. ( the parameters did map when I used
the
gnp data in harvey's 1989 book but that's the only
time.
whenever I generated data they don't map ).
I have looked at everything ( except
writing my own
version of Ssflik ? ) and it
looks like I am doing everything correctly.
Does anyone have an opinion on if I
should
I expect the parameters to map over to each other
or is this asking too much because the algorithms
( Ssflik versus arfima.estimate ) are so
different.
thanks a lot for any
opinions. if anyone has code for
estimating single source of error state space
models
in the state space
world,
I would be interested in comparing my code to
theirs. or
if anyone is
interested in my code , I can send
it to them but I had to
a) create a state space estimate class ( like what
the
authors did with probit except for state
space estimation )
that estimates both multiple source and single source
state
space models
b) create a myarfimasim
class which inherits arfimasim
but saves the data ( because I had to then write
a function
to forward difference it (
either once or twice )
because d equal to an integer doesn't work in arfimasim
)
so there would be some extra files but I can
definitely send them to anyone who wants the
code.
I am leaving for a few days
but
will be back next week to reply to anyone
who
answers this.
Mark
Mark Leeds
