This pet-project addresses the boosting uncertainty problem by using the implementation of the virtual ensemble described in this article A. Ustimenko, L. Prokhorenkova and A. Malinin, Uncertainty in Gradient Boosting via Ensembles” (2020), arXiv preprint arXiv:2006.10562.
As a dataset, time series of product sales is used, where target variable is a number of sales per day.
metrics_validation.py contains the metrics for evaluation the model and the class for group time series validation written from scratch.
In those high-risk tasks where machine learning applied is is crucial to estimate uncertainty in the predictions to adoid mistakes. A virtual ensemble comprised of the sub-models from the one trained gradient boosting model can resolve this problem. 
By using the virtual ensemble, the knoweldge uncertainty was estimated:
