Improvements of Machine Learning forecasting for perishable food sales in real-world grocers

Abstract

Machine Learning is a useful tool for a wide range of forecasting applications that can differ largely in available data and suitable methods. We do a deep investigation of Machine Learning for demand and sales forecasting in retail and wholesale domains. Using Machine Learning in these scenarios helps to optimize the ordering process and helps to reduce food waste through improved inventory management and ordering. Although this problem has been studied by several researchers in the literature, the reproducibility of the results is usually lacking because of the unavailability of the data, and there is a potential for advanced features and prediction methods. We collaborate with three large Austrian retailers and wholesalers who provide real-world data and insights into their business processes. Based on the data, we propose, collect, and create a dataset containing a novel combination of features that includes extensive long-term real-world sales and other business data, contextual data, weather data, and movement data retrieved from cellular towers. Using this, we propose a combination with state-of-the-art machine learning techniques to improve the performance of forecasting in diverse real-world retail and wholesale environments. We make anonymized datasets available to facilitate future research and reproducibility of results. Our evaluations show that combining extensive datasets with state-of-the-art algorithms improves our forecast performance to help generate more accurate orders.