Description

This notebook runs AutoMFLES and Lightgbm. Both algorithms are based on Nixtla. Once we have obtained forecasts from multiple inference runs on validation sets (or hold-out sets after fitting / tuning), we want to find optimal weights to create a blended ensemble. We do this by minimising our desired error metric via scipy optimise based on the sample data we generated. For the final inference, we apply the obtained weights. 

The core idea is to combine a statistical learner and a global ML model and to combine them to reduce regression to the mean (under-forecasting top-sellers, over-forecasting long-tail articles), which occurs in the global ML model. At the end we show how we increase the average forecasted values for top-sellers, while reducing forecasted values on average for low-selling articles and thereby getting closer to the real distribution of the test set (when ensembled compared to Lighgbm alone).