Sign up to take part
Registered users can ask their own questions, contribute to discussions, and be part of the Community!
I'm currently working on a project that involves time series forecasting. My current context length and forecast horizon is set to 120. I've observed that the minimum length required for a time series is 360 for DeepAR and 374 for AutoARIMA in Dataiku.
What I don't quite understand is why there's such a specific minimum time series length requirement for these algorithms. Could anyone provide an insight or explanation ? 240 should be the expect min length right ?
Operating system used: Ubuntu
With an horizon of 120, DeepAR needs at least 240 time steps to train the model (the model expect 120 time steps as input (the context length) and outputs 120 time steps) and an additional 120 time steps to evaluate the model.
For AutoARIMA, the minimum number of time steps for training also includes two times the season length, it's an heuristic that prevent the autoarima training from failing too often.