How much of time series data use to model/forecast

[Mkkl]

Greeting to you all

I was wondering the following: I want to fit a VAR model to two time series.
I would like to base the model on part of the data and subsequently use the model to make a forecast of the last few months of my data.
Now I'm wondering if there is some kind of rule of thumb of how much data to leave out. In this case I'm using monthly data over a time span of 18 years.

Thank you!

[trubador]

I do not think there is one. In predictive analytics, there are such rule-of-thumbs that correspond to splitting of sample with respect to training, testing and validation stages. However, I believe this approach would not make sense in this case for at least two reasons: i) Your data is not as large (or big) as the ones that are used in those studies, ii) Time dimension may introduce parameter instability.

First, make sure that parameters of your model are stable over time. If they are not and your sole purpose is forecasting (not a structural analysis), then you "can" consider a shorter span and/or a longer lag structure. It would be an ambitious task to forecast more than a year ahead with monthly data, but it is always better to experiment with different forecast horizons and to look at alternative forecast evaluation criteria. Consider this as a robustness check. And since EViews will do most of the job, you do not have to worry about the computational burden...