I am regression yearly electricity consumption on GDP and population.
I found that both independent variables are cointegrated with electricity consumption (with Granger causality running from the independent variables on the dependent variable)
I am just interested in long-run forecasts, not in short-term behavior.
Do I still need to run a VECM? Or is OLS fine in this case?
Is there a difference if also GDP and population are cointegrated (e.g. that then I need to use VECM for sure because of endogeneity)?
Thank you!
You have to run VECM because OLS alone is not sufficient for that analysis
I understand that you are only interested in making forecasts and are not trying to make structural or causal interpretations of the coefficients in your model. A good forecasting model has clean residuals. A VECM or VAR without clean residuals is likely to produce bad forecasts. In forecasting, almost anything goes. Simple or parsimonious models forecast better than complicated over specified models. No one can give a general statement that one method of forecasting is better or worse than another without an analysis of your data. How long is your long run?
That leaves you with a variety of forecasts from univariate ARIMA or ARIMAX through multivariate VAR, Bayesian VAR (Minnesota Priors), VECM, etc. How then do you decide which method is best for you? The econometric/statistical software that you are using should have a set of routines that you can use. If it doesn't see if you can get access to a different package. You will find specific details in the software manuals. It is important to remember that the within-sample fit of the data to the model is a bad measure of the ability of the model to forecast.
Say your sample has n measurements and you with to forecast p periods ahead. Take the first m measurements where m+p
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running