| Abstrakt: |
Abstract In this work, we aim to study the effect of weather conditions on the sales of the FMCG market. For this purpose, we have collected an extensive dataset consisting of weekly category sales of over 1000 stores in the Czech Republic for years 2015 to 2017, coupled with various meteorological variables for over 80 different weather stations. We introduce a novel approach to analysis, using tree-based machine learning algorithms. These flexible non-parametric methods can estimate complex relationships as well as performing an automatic variable selection. Both of those attributes are critical in our work, as the final dataset consists of over 130 variables. The central point of this thesis is to either conclude there is only a negligible relationship or to provide a model with robust performance and explainable results. We manage to show a significant sales reactions based on changing weather conditions for three top-selling categories, producing a model that significantly outperforms both benchmarks, lasso regression and tree-based model trained on non- meteorological variables only. Ultimately we present two conclusions, firstly that linear regression, a commonly used methodology in similar studies, is not a suitable approch for modeling the weather effects and secondly that the weather variables significantly enhance the demand planning. Keywords weather, FMCG, fast moving consumer goods, retail, classification and regression trees, random forest |
