In FPP3 section 5.10 on time series cross-validation, an example is given on the forecasting performance of 1- to 8-step-ahead drift forecasts. In the section on evaluating point forecast accuracy, it is recommended that the test set should ideally be at least as large as the maximum forecast horizon required (and preferably this comprises at most 20% of the number of observations).
Here is a snapshot of the code from the article:
google_2015_tr <- google_2015 %>% stretch_tsibble(.init = 3, .step = 1) fc <- google_2015_tr %>% model(RW(Close ~ drift())) %>% forecast(h = 8)
To incorporate the '20%' guideline, don't we need
.init = 40 so that the forecast horizon forms at most 20% of the total data? I seem to have trouble fitting
ARIMA without a much larger
In addition, say we wanted to forecast 12 months ahead. Does it make sense simply to do time series cross-validation with 1- to 12-step-ahead forecasts, or is there an argument for going further?