Forecasting Individual Brand Choice with Neural Networks: An Empirical Comparison

In this study we compare the forecasting performance of a multinomial logit (MNL) model with that of an artificial neural network (ANN). We also implement a hybrid model, which uses an ANN as a diagnostic tool to detect nonlinearities in the data and then incorporates the nonlinear relations found into an MNL model. We use a scanner data set containing information on 2798 Catsup purchases with four different brands. Price and whether the product was on display or featured in an advertisement at the time of purchase is given, and we also construct a brand loyalty term. Data is split up into a training and test set and different data partitions are evaluated. Forecasting performance is measured with accuracy, the negative prediction ratio and evaluation of confusion matrices. The MNL with an 80/20 data split yields an accuracy of 72.1% on the test set, while the ANN with data partitioning of 65/35 performs slightly better with an accuracy of 72.3%. The hybrid model outperforms both with an accuracy of 73.2% and thus we conclude that this is the most suitable model to forecast individual brand choices on Catsup.