Time Series Sales Forecasting Analysis: A Case Study

This study focuses on the examination of time series data concerning thermal paper consumption at different sales points, motivated by the logistical challenges of a company facing unexpected shortages and the subsequent need for expensive, unscheduled shipments. The research aims to develop a predictive model using a bespoke algorithm to forecast future demand accurately, thereby optimizing shipment volumes and reducing the frequency and cost of unplanned deliveries. Two forecasting algorithms, Prophet and DeepAR, were applied to model the data. Prophet is utilized for its flexibility and ability to adapt to the unique seasonal and trend patterns of individual or grouped time series. In contrast, DeepAR adopts a collective modeling approach, enhancing prediction accuracy by learning from a broad array of related time series. The algorithms underwent testing across various settings, and their performance was assessed using a novel metric that compares the predicted against the actual consumption of thermal paper at each sales point for the forthcoming quarter. The results indicated a superior performance of the Prophet algorithm over DeepAR, highlighting the importance of tailored forecasting models in addressing the heterogeneity of time series data and improving logistical and operational efficiencies for businesses encountering similar distribution challenges.