EViews.com

Hi,


I have a model that tries to link the yield spread to future changes in GDP/industrial production.


The specification that I have is:

%change in industrial production over t, t+48 = c + beta*yield spreadt + errort

and this data is run over a monthly frequency and estimated using OLS with HAC variance.


It is obvious from the specification that a shock to industrial production will last in the model for 48 periods and induce an autocorrelative behaviour.

I've attempted to correct it using the HAC variance however it did not help.

Also, if I put in AR(1) structure in the estimates, for obvious reasons, the AR term dominates and renders other variables insignificant.


With the simple model with HAC variance and without the AR(1) term, the regression output show high levels of adjusted R squared however due to the highly autocorrelative specification, I highly doubt its usefulness.


Could you advise me of any specifications that I could make?


Many thanks in advance :)


David

The HAC variance doesn't have anything to do with the R^2, nor does it change the serial correlation properties of the model. HAC corrects the standard errors.

You might try an MA specification, although it appears that EViews doesn't permit 48 MA terms.

Oh yes, I referred to the R^2 just to point out that it might be due to the autocorrelation

Hmm MA sounds like a good alternative, I'll give that a try. Thank you :)

I've tried and eviews does indeed allows for MA(48) structure. However, the MA structure is found to be insignificant

In EViews notation, MA(48) means a single MA term at lag 48. If you have 48 overlapping periods, you want MA(1 to 48).

But what exactly did you mean when you said "HAC variance did not help?"

Thanks for the reply :)


I think I will have to clarify on the concept, so if Newey West HAC variance is used, are test results will be robust against autocorrelation and heteroskedasticity?

So in that case, is there no need for me to worry about autocorrelation once I use the HAC variance? (although R square may still be inflated by autocorrelation)


Many thanks!

That's right, although you need to remember that HAC requires a large sample

awesome,

so since the quantities of interest for my research is both the slope coefficient (and its significance) and also the R squared, I was wondering if there are any other way of treating autoregressive behaviour other than using ARMA structure.

thank you for answering my questions!

Another approach may be to use lags and build in an explicit dynamic structure.