Improving milk sales quantitative estimation by using POS data

In this study, experiments on milk sales estimation were conducted in two phases to optimize stock ordering. Sales data of two brands of milk between 2007 and 2008 were obtained through 9 stores in Ishikawa Prefecture. In the first phase, several variable reduction methods were investigated to prevent overfitting of prediction models using methods of the Principal Component Analysis and Decision Tree. In the second phase, the estimation experiments were conducted by using learning data generated through the variable reduction methods. Neural network, k-NN algorithm and RBF network were used for prediction models. The results showed k-NN algorithm with original non-reduced variables and neural network with cumulative contribution rate of 90% yielded higher accuracy in sales estimation.