How to get rid of autocorrelation for VAR model that requires high lags


New Member
I am trying to develop a model yield to forecast yield curve 6 months in the future, loosely based Diebold Li term structure model. I am using a Vector Autoregression to regress the factors at time t against factors at time t-6 (which is the method proposed in the Diebold Li paper). However the results show significant autocorrelation of the errors at lag 1 (using PACF). However, as the model needs to be used to forecast yield 6 months in the future, using t-1 as a regressor would defeat the purpose of the model. How can I get rid of this autocorrelation? I have tried using lag 6 and lag 7 but it still doesnt get rid of the autocorrelation. Would the forecasting ability of this model be poor because of this autocorrelation?
Thank you!:wave:
Models are best presented using equations, especially the multi-dimensional ones. Please rewrite rigorously in 1-2 paragraphs... Yes, if the sample size is small, autocorrelations has to be removed. Otherwise all p-values are off.


Fortran must die
In ARIMA, and regression with ARMA errors, you get rid of autocorrelation by specifying lags of Y I believe (certainly you identify the AR structure and specify a factor to correct for that). I am brand new to VAR methods so I do not know how this is addressed in that.