Hi,
I am trying to use a VAR(1) model with 4 I(1) variables to make a 1 step ahead forecast for only one of the variables (ie WTI oil price here). I use the first difference of the variables after checking their stationarity with DF-GLS and KPSS test. I find a lag order of 1 to be used and then i estimate the model:
As you can see, the variables used hold explanatory power for 3 variables (dlwti, dlind, dlprod).
I then check that the residuals are white noise with the varlmar command:
Unfortunately here, the hypothesis of no autocorrelation must be rejected at the 10% level for too many lags, which would indicate that the model is not fit to be used.
Now if i check specifically for the residuals of the equation i am interested in (lwti), the Portmanteau test cannot reject that the residuals are white noise:
Now, does this mean that the model is fit to forecast dlwti for a 1 step ahead time line ? Or should i dismiss totally the model ? Adding lags to the model does not solve the problem.
Furthermore, the main issue here seems to be coming from dinv, which shows a strong seasonality in the ACF and PACF of its residuals. I should probably get rid of this, but i don't know how to deal with the seasonality in the VAR model.
I hope you might help me on this,
Thank you for your time,
Pierre.
I am trying to use a VAR(1) model with 4 I(1) variables to make a 1 step ahead forecast for only one of the variables (ie WTI oil price here). I use the first difference of the variables after checking their stationarity with DF-GLS and KPSS test. I find a lag order of 1 to be used and then i estimate the model:
As you can see, the variables used hold explanatory power for 3 variables (dlwti, dlind, dlprod).
I then check that the residuals are white noise with the varlmar command:
Unfortunately here, the hypothesis of no autocorrelation must be rejected at the 10% level for too many lags, which would indicate that the model is not fit to be used.
Now if i check specifically for the residuals of the equation i am interested in (lwti), the Portmanteau test cannot reject that the residuals are white noise:
Now, does this mean that the model is fit to forecast dlwti for a 1 step ahead time line ? Or should i dismiss totally the model ? Adding lags to the model does not solve the problem.
Furthermore, the main issue here seems to be coming from dinv, which shows a strong seasonality in the ACF and PACF of its residuals. I should probably get rid of this, but i don't know how to deal with the seasonality in the VAR model.
I hope you might help me on this,
Thank you for your time,
Pierre.
Comment