An adaptive agent architecture for exogenous data sales forecasting

Abstract

In a world of unpredictability and complexity, sales forecasting is becoming recognised as essential to operations planning in business and industry. With increased globalisation and higher competition, more products are being developed at more locations, but with shorter product lifecycles. As technology improves, more sophisticated sales forecasting systems are developed which require increasing complexity. We tum to adaptive agent architectures to consider an alternative approach for modelling complex sales forecasting systems. This research proposes modelling a sales forecasting system using an adaptive agent architecture. It additionally investigates the suitability of Bayesian networks as a sales forecasting technique. This is achieved through BaBe, an adaptive agent architecture which employs Bayesian networks as internal models. We develop a sales forecasting system for a meat wholesale company whose sales are largely affected by exogenous factors. The company's current sales forecasting approach is solely qualitative, and the nature of their sales is such that they would benefit from a reliable exogenous data sales forecasting system. We implement the system using BaBe, and incorporate a Bayesian network representing the causal relationships affecting sales. We introduce a learning adjustment component to adjust the estimated sales towards closer approximations. This is required as BaBe is currently unable to use continuous data, resulting in a loss of accuracy during discretisation. The learning adjustment additionally provides a feedback aspect, often found in adaptive agent architectures. The adjustment algorithm is based on the mean error calculation, commonly used as sales forecasting performance measures, but is extended to incorporate a number of exogenous variables. We test the system using the holdout procedure, with a 5-fold cross validation data-splitting approach, and contrast the accuracy of the estimated sales, provided by the system, with sales estimated using a regression approach. We additionally investigate the effectiveness of the learning adjustment component.