Do Forecasting Algorithms Need a Crisis-Mode? Machine Learning Based Sales Forecasting in Times of COVID-19

On average, algorithmic forecasts outperform human forecasts by 10%. This disparity increases further thanks to cutting-edge machine learning (ML) algorithms. As business decisions based on sales forecasting are regarded as particularly important and a variety of other activities rely on them, accurate sales forecasting is critical to companies’ profitability. At the same time, being able to predict the next day’s sales more accurately can significantly reduce food waste and help fulfilling sustainability. Thus, sales forecasting is one of the primary value propositions of artificial intelligence (AI). However, it is crucial for the acceptance and adoption of ML-based sales forecasting algorithms to perform reliably during pandemics such as the covid-19 pandemic. Although governments’ containment measures highly impact the sales of a bakery’s products, no study has yet scrutinized incorporating the stringency of containment measures as an input variable for sales forecasting. Hence, this paper examines the performance of a ML sales forecasting system for baked goods in times of covid-19 and proposes incorporating a covid containment measurement stringency index as an additional input variable to increase forecast accuracy in times of pandemics. This way, prediction accuracy increases by 4.61% on average. Consequently, a containment measures stringency variable should be used to increase accuracy in future pandemics. By simulating an upcoming pandemic, it is further demonstrated how learnings from the covid-19 pandemic could be meaningfully transferred. For this study, real data is used: A Swiss bakery chain provides real sales data covering 5 years including 2 years of the covid-19 pandemic.