I am working on a forecasting project with a univariate (at the moment) multi-seasonality time series.
I want to go step by step and improve the forecasting model step by step.
As of beginning, I am struggling (potentially with my understanding) and surely with R to use expanding forecast.
Data set: 3 years and 3 months of hourly observations
Problem definition: forecasting 24-step ahead
Training set: first 3 years of hourly observations (26.298 obs)
Test set: last 3 months of hourly observations (2.160 obs)
Goal: I would like to forecast 24 observations, then use the 24 corresponding real observations from the test set to forecast the next 24 observations, and repeat this until I have forecasted my entire test set.
Anyone knows how I could proceed to do so?
Other question: when I have to forecast 24-step ahead, should I fit my model to do so?
I mean from what I have read so far, the models is trained not knowing how many step-ahead he will have to forecast. I find it quite surprising. Is there anything I am missing in my understanding?
I look forward to reading from you about this interesting subject,