time series cross validation and forecast origins

[nickcox896]

I have a monthly series from 2001m1-2026m7 and I want to make annual forecasts for the next three years (h=36). I use the time series CV as given here link to compute the RMSEs.

In this context, I am not sure where my first forecast origin should begin.

1. should I use, say, the first 70-80% of initial observations( that is approx 2001m1-2021m5) for model estimation and then start the CV for the next 36 months?
2. should I go back say 4 years (that is approx 2022m7) and then start the CV for the next 36 months?

Is there any suggestion that I could follow?

[robjhyndman]

The initial training set needs to be as small as possible while still being able to reliably estimate the model. That way, you get the most test sets to average over, and therefore get a better measure of accuracy. With monthly data, you can normally fit a model relatively easily with 3 years of data. But if you need more for a sensible fit, adjust this as appropriate.

[nickcox896]

Thank you very much. Assuming that I use the first 3 years for the estimation of the model, then what h should I consider for CV RMSEs? Should obtain these RMSEs for h=12, 24, 36,42,... (that is every 12 months)?

[robjhyndman]

Choose whatever horizons for which you want to measure accuracy