Sales Forecasting of Perishable Foods With Multiple Stores and Communities-An Empirical Study of Convenience Stores in Taiwan

Abstract

Purpose: In recent years, smart retailing has been gradually acknowledged and has slowly impacted countless industries. This study takes smart retailing as the starting point and explores the increased attention on the perishable food (fresh food) market in convenience stores. Remarkable improvements in the general food consumer population in Taiwan and the advent of an aging society with changes in social structure and consumer style are also taken into consideration. As perishable goods (fresh food products) easily spoil, the scrap rate of fresh food in convenience stores has always been a topic of great importance. Design/methodology/approach: Perishable food from well-known convenience store chains in Taiwan is taken as the subject of this study, and sales data of stores in different regions are used to establish sales forecast models for the convenience store chains. The optimal sales forecast model for each store's products was established through a data analysis, and the characteristics and differences of all products are explored to establish decision-making advices for enterprises in ordering perishable food. Multiple stores and perishable foods with two-year sales data of 11 perishable foods in every six stores in different county are adopted to establish the predictive models in this study, including time series, support vector machine (SVM), Lagrangian support vector machine (LSVM), random forest, neural network, generalized linear, and generalized linear mixed models. Findings: Results show that the time series model and SVM have lower prediction error values and better prediction results among all the established sales forecast models. Given the influence of region and population characteristics, varying models are applicable to stores in different regions. Thus, the difficulty of prediction will also vary. Practical implications: Different commodities will have varying levels of prediction difficulty due to dissimilarities in commodity attributes. Convenience stores are generally willing to predict the sales of fresh food through an artificial intelligence model. Different forecasting models should be selected by stores. If one or more forecasting models are used for prediction, the model with a stable forecasting error should be selected for implementation. Originality/value: This research analyzes the products sold in multiple stores and applies different time series and machine learning algorithms to build predictive models. The results show that the most suitable algorithms for stores of products are different. If a convenience store wants to build a predictive model, differentiated models must be established for different stores and commodities.