An implementation of a feature based algorithm inspired by Hyndman's work on the FFORMS.
https://robjhyndman.com/papers/fforms.pdf
The idea is to compute rolling features and use them to compute the t+h next points.
The computation of the rolling features is based on pandas and the forecasting use the scikit-learn API.
To use the FBE, one just have to specify his sklearn model, the forecast horizon, the seasonal length (used to compute rolling features) and the freq of the datas.
Tests suggests that increasing the size of the seasonal length will increase performance on forecasts but also increase the performance time.