AutoPilot for Forecasting
Hi there,
IHAC who asked a question regarding the possible use of AutoPilot for solving Forecasting problems. They don't have the knowledge to play with DeepAR and they are running tests in parallel with Amazon Forecast. Their questions are:
- Is it possible to use AutoPilot for Forecasting problems? (my answer would be yes, since regression problems can be solved by XGBoost, which also won a bunch of competitions on Forecasting)
- Which kind of pre-processing should the customer do and which pre-processing is done by AutoPilot which could simplify transformation of data for solving forecasting? In particular: are there any transformation to be done on the timestamp column? Should we introduce lagged entries - or is it done by AutoPilot?
Thanks to those taking the time to answer these questions :)
Best, Davide Gallitelli
- 最新
- 最多得票
- 最多評論
Although it is possible to model the forecasting as a regression problem in Autopilot, there is no time-series capability built in Autopilot. So, you need to do the preprocessing tasks such as time-series windowing, lag differencing, etc. in order to generate the training/test datasets for the autopilot experiment.
Additionally, time series forecasting usually requires a model which can detect the pattern in a sequence of features. So, services such as DeepAR or Amazon forecast provide better capabilities to address this challenge.
