A new time series forecasting business approach.
This is an idea I get when I was working as a Data Scientist for Bouygues Telecom (for an internship).
Many upgrades / issues have been discovered :
-
Sometimes for a period, it don't find a model (stationary issues). Knowing that we are computing a mean of time series forecast, if for one period, it can't generate a model, then our weighted mean is not going to work (generate errors).
-
For more complex time series, the fbprophet seems to well performed. We should implement it.
Last update 2020-15-08