I was hoping in the Forecasting: Principles and Practice book that there would be some discussion about how to choose the granularity of the time series data for which to train your time series models (or at least I didn't find it).
Let's say I have sales data recorded daily and I want to make both short-term forecasts and long-term forecasts for future sales. If I want to predict how many sales there will be in year 2023, I have many options of how to proceed:
- I can use daily data to make 365 forecasts for Jan 1, 2023 to Dec 31, 2023 and then sum them together to get the forecast for the entire year
- I can transform my daily data to quarterly data to make 4 forecasts for each quarter in 2023 and then sum them together to get the forecast for the entire year
- I can transform my daily data to yearly data and make a singular forecast for 2023
I am quite unsure of what the best approach to take is. I know for my short-term forecasts I will want to make forecasts at the weekly, monthly, or quarterly level so to me it makes sense to use daily or weekly data and aggregate as necessary.
For the long term forecasts (2-3 years out), would I use the same model that I use to make short-term forecasts? Is it better to use less granular data? Otherwise if I use say weekly data, if I wanted to make a 3 year forecast, it would involve making forecasts 104-156 time steps in the future. Would that make sense?