Comparison of category-level, item-level and general sales forecasting models

Abstract

Sales forecasting is the process of estimating future sales. In this thesis, multiple methods are tested out for achieving best forecasting accuracy with lowest computational requirements. Three families of methods are investigated: a traditional statistical forecasting approach (ARIMA), classical machine learning techniques (specifically ensemble methods) and a third one based on deep learning methods (specifically recurrent neural networks with LSTM architectures). The study uses real-world sales transaction data from a large retail company in a Baltic country and the aim of this thesis is to improve their current sales forecasting system. Here we show that improving on their current sales forecasting is possible and additionally analyse the influence of promotional sales to prediction accuracy. The results show that using a combination of multiple item-level decision tree-based ensemble models yields the best prediction accuracy with regard to training complexity. Additionally, when comparing accuracy of forecasts for promotional sales and non-promotional sales, a variant of ARIMA achieves the most accurate results when forecasting promotional sales.