Food Sales Prediction: "If Only It Knew What We Know"

Sales prediction is an important problem for different companies involved in manufacturing, logistics, marketing, wholesaling and retailing. Food companies are more concerned with sales prediction of products having a short shelf-life and seasonal changes in demand. The demand may depend on many hidden contexts, not given explicitly in the form of predictive features. Even if some changes are known to be seasonal, predicting (and even detecting) when season will start and end remains to be non-trivial. In this paper we present an ensemble learning approach that employs dynamic integration of classifier for better handling of seasonal changes and fluctuations in consumer demands. We focus our research on studying how the business is currently operated, and how we can improve predictions for each product by constructing new groups of predictive features from (1) publicly available data about the weather and holidays, and (2) data from related products. We evaluate our approach on the real data collected by food wholesaling and retailing company. The results demonstrate that (1)our ensemble learning approach can perform better than the currently used baseline, (2) we can handle seasonal changes with ensemble learning better if feature set for a target product is complemented with features of related product (having similar sales pattern), and (3) an ensemble can become more accurate if information about the weather and holidays is presented explicitly in a feature set.