What I usually do to forecast is:
- Train a model (i.e. bats, stl + arima|ets, ANN, ELM, etc ...) through a model function obtaining a trained model object, based on information criteria or cross-validation for time series.
- Use a forecast|predict method to obtain iterative h step-ahead forecasts using the trained model.
- Evaluate forecast accuracy (for each horizon up to h)
All the above considering rolling forecasting origin techniques that allow to move the training and test sets in time and specify a model for each window.
I have noticed that several methods found in R packages have a
model parameter that (if specified) apply an existing model to a new data set. Considering the rolling forecasting scheme described above:
- When it is appropriate to use the
- Is this option an alternative to fitting a model for each horizon as in the rolling window scheme?
- What does the ''If model is passed, this same model is fitted to y without re-estimating any parameters'' description implies to forecasting procedures?