Abstract

The topic of this master thesis is to determine whether product positioning and sales correlation can improve sales forecasting of groceries. Previous studies have stated that the sales of groceries are related to their in-store placement. If this holds, it might be possible to use that relation to perform forecasting of sales. A machine learning framework is applied to perform the forecasting to fulfill this purpose. The machine learning framework consists of several supervised regression models and a neural network. The models are used to forecast sales by first considering product positioning and sales correlation and then not doing so.

One obstacle in forecasting is the need for comprehensive time series. A possible solution is to use augmented data, which was the decision in this project. However, using augmented data requires reasoning about this choice’s effect on the forecasts’ outcome. Other than data augmentation, re-sampling and data-cleaning are topics of this thesis.

The thesis concludes that using product positioning and sales correlation as features in machine learning models does not necessarily improve sales forecasting. Nevertheless, it is found that there are circumstances when the inclusion does improve the forecast. Those circumstances are when there are many products placed in one section and when the turnover in a section is high. More extensive studies will be needed to fully determine whether product positioning and sales correlation, in general, improve sales forecasting.