Competing Forecasting Models to Study Crisis Periods: The Case of Sweet Snacks Sales

The COVID-19 pandemic significantly affected the purchasing behaviour of Portuguese families, compelling them to reduce their shopping expenditures. This socioeconomic crisis necessitated that food retailers adapt their strategies to evolving consumer preferences, emphasizing digitalization, sustainability, and safety. This study examines the sales evolution of the Sweet Snacks category at two major Portuguese retail banners1 from January 2018 to June 2023, segmented into three forecasting periods: pre-crisis, crisis and post-crisis. The project’s primary objective is to infer forecasting models for these periods, using the ARIMA and Prophet time series models, and compare them to assess consumer preference changes. Additionally, this work forecasts Sweet Snacks sales beyond June 2023 to extend the appraisal of sales performance in the post-crisis and detect potential anomalies in the sector. Using the CRISP-DM methodology, the research developed an integrated BI solution, employing Power BI for data preparation and R Studio for multidimensional data modelling and forecasting analysis. In the pre-crisis period, Sweet Snacks sales progressively increased until the onset of lockdown, declining in the crisis period. At the end of the crisis, consumption patterns normalised, but post-crisis, retailers diverged due to their adaptability to new trends. The results indicate that while ARIMA models generally offer higher accuracy, Prophet provides more precise future forecasts. ARIMA predicts a steady future trend, whereas Prophet captures post-crisis sales patterns more effectively. This project’s main contribution is the development of a BI solution and a comprehensive forecasting report for a Consulting organisation in the Food Retail sector.