An Extended Evaluation Framework for Neural Network Publications in Sales Forecasting

Abstract

While artificial neural networks (NN) promise superior performance in forecasting theory, they are not an established method in business practice. The vast degrees of freedom in modeling NNs have lead to countless publications on heuristic approaches to simplify modeling, training, network selection and evaluation. However, not all studies have conducted experiments with the same scientific rigor, limiting their relevance to further NN research and practice. Consequently, we propose a systematic evaluation to identify successful heuristics and derive sound guidelines to NN modeling from publications. As each forecasting domain of predictive classification or regression imposes different heuristics on specific datasets, a literature review is conducted, identifying 47 publications within the homogeneous business domain of sales forecasting and demand planning out of 4790 publications within the domain of NN forecasting. The identified publications are evaluated through a framework regarding their validity in experiment design and reliability through documentation, in order to identify and promote preeminent publications, derive recommendations for future experiments and identify gaps in current research and practice.