Forecasting Food Sales in a Multiplex Using Dynamic Artificial Neural Networks

Abstract

In India, food sales are emerging to be a major revenue generator for multiplex operators currently amounting to over $367 million a year. Efficient food sales forecasting techniques are the need of the hour as they help minimize the wastage of resources for the multiplex operators. In this paper, the authors propose a model to make a day-ahead prediction of food sales in one of the top multiplexes in India. Online learning and feature engineering by data correlative analysis in conjecture with a densely connected Neural Network, address the concept drifts and latent time correlations present in the data respectively. A scale independent metric, $\eta _{comp}$ is also introduced to measure the success of the models across all food items from the business perspective. The proposed model performs better than the traditional time-series models, and also performs better than the corporate’s currently existing model by a factor of 7.7%. This improved performance also leads to a saving of 170 units of food everyday.