Rolling forecasts where horizon is larger than step-size

Question

Is it bad practice to perform rolling time-series forecasting where the forecast horizon is greater than the step-size? For example, if I have a model which produces weekly forecasts on a rolling day-by-day basis, then it gives 365 weekly forecasts a year, as opposed to 52 forecasts if my step-size is equal to my forecast horizon.

On one hand, having daily forecasts gives the model more training data, but on the other hand, errors will be serially correlated.

I've not been able to come up a good intuition about setting up the problem, and would appreciate any help on how frame the problem / best practices.

Answer 1

You should produce your forecasts for the horizon for which they are necessary. If you need forecasts for one week ahead, then that's fine. How often you produce such forecasts is completely independent from the horizon. You could produce a new forecast every minute if you wanted. However, this will probably be a waste of resources, so computing a new forecast each day, as you propose, is probably more reasonable.

You are correct that the errors will be serially correlated and this will affect your backtesting, e.g. vie Diebold-Mariano tests, but they can be adapted to this situation without much of a headache.