Forecast standard errors
Posted: Tue May 07, 2013 3:40 am
Hi,
It seems to me that the forecast standard errors you get when forecasting a series that is covariance stationary once it has been transformed by taking the percentage change are incorrect. For example, if you generate forecasts for GDP growth using the AR(1) model
@pc(gdp) c @pc(gdp(-1))
the forecast standard errors do not converge to a constant value for large enough horizon (the way we expect them to). (That they grow with the forecast horizon for the level of GDP is of course fine and in line with theory for a unit-root process.)
However, if you transform the series manually - that is, generate the series x which is the percentage change of GDP - and then model x such as
x c x(-1)
the forecast standard errors are correct.
Would you agree?
It seems to me that the forecast standard errors you get when forecasting a series that is covariance stationary once it has been transformed by taking the percentage change are incorrect. For example, if you generate forecasts for GDP growth using the AR(1) model
@pc(gdp) c @pc(gdp(-1))
the forecast standard errors do not converge to a constant value for large enough horizon (the way we expect them to). (That they grow with the forecast horizon for the level of GDP is of course fine and in line with theory for a unit-root process.)
However, if you transform the series manually - that is, generate the series x which is the percentage change of GDP - and then model x such as
x c x(-1)
the forecast standard errors are correct.
Would you agree?