GARCH Forecasting Using @MVGarchFore

[Gregory]

Hello everyone,

I'm having a bit of trouble with the forecasting procedure.

I have 310 weekly observations of three variables, x, y, and z. I want to treat the first 208 as my estimation or in-sample, and the remaining 102 as my forecast or out-of-sample.

I successfully estimated a multivariate BEKK GARCH model for the in-sample, saving the covariance matrices and residuals.

garch(p=1,q=1,mv=bek,asymmetric, $
mvhseries=hv,hmatrices=hh,rvectors=rv,$
method=bfgs,iters=200,pmethod=simplex,piters=20) 2000:01:12 2003:12:31 x y z

I then took a shot at forecasting, using 100 steps (close enough to 102) like this. (I'm guessing that the term "steps" corresponds to the frequency of the data being forecast, so that 100 steps means 100 weeks in my case.)

@MVGarchFore(mv=bek,steps=100) hh rv

Unfortunately, I get the error message

## MAT2. Matrices with Dimensions 6 x 1 and 3 x 1 Involved in .* Operation
The Error Occurred At Location 1949 of MVGARCHFORE
Line 67 of MVGARCHFORE

However, if I run the forecasting procedure without the model specified, that is, without mv=bek, as

@MVGarchFore(steps=100) hh rv

it appears to run OK, although there is no result returned to the screen. I am able to access what are apparently forecasts of the covariance matrix by displaying hh beyond my in-sample of 208.

display hh(209)
display hh(210)

etc.

But when I get to about hh(225), which would be 17 steps into the forecast period, the covariance matrix stops changing, so I figure that I do need to specify the model in the procedure's options.

Any advice would be most appreciated.

Gregory

[TomM]

Gregory:

If you're still having problems with this, email a copy of the program and data file along with your RATS version number and serial number to support@estima.com and we'll take a look.

Tom Maycock

[Gregory]

Well, I appear to have MVGarchFore working in that I don't get any error messages, and the series of hh matricies contains elements that change for from one date to the next. Very pleased.

Now I have a question about forecasting as opposed to using RATS to generate the forecasts. Is it to be expected that forecasts converge, that is, become more similar or change less, the farther you forecast? I'm sure that in retrospect I'll be embarrassed for asking. But what I notice is that the element-wise difference of each new hh forecast becomes progressively smaller, the farther out I go.

Any thoughts on this would be most appreciated.