Performance Comparison of State Space and RBF-based Models for Daily Sales of Small Restaurants

• Activity Number: 368
• Type: Contributed
• Date/Time: Wednesday, August 10, 2005 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics and Marketing
• Author(s): Rui Yamaguchi*+ and Tomoyuki Higuchi
• Companies: Kyushu University and Institute of Statistical Mathematics
• Keywords: micro marketing ; daily sales prediction ; restaurants ; state space model ; radial basis function network model

Abstract

Daily sales of restaurants are affected by several factors: day of the week, national holidays, weather, events taking place near the site, etc. It is beneficial to construct a model to predict future sales based on the inherent information of each store. For such purpose, we can consider dynamic and static models. The former, such as time-series models, involve dynamic structure of time series, explicitly. The latter, such as linear regression models, do not. Although it seems natural to use a dynamic model to predict daily sales, there exist many static models with high representational power. Therefore, it is meaningful to compare predictive abilities of both types. We compared predictive abilities between a dynamic model (a state space model) and two static models (a multiple linear regression model and a radial basis function network [RBF] model) by applying them to two-year daily sales data of a restaurant adjacent to a large convention center. The result suggested that although the RBF model earned a good mark at a few points, the dynamic model was more advantageous in many aspects.