Machine Learning for Restaurant Sales Forecast

There are many restaurants that do not have a solid forecast of their daily sales.Often, they neither have the education nor the energy to make a calculated estimation of the sale. At best some restaurants look at the last years sale on the equivalent day and the current date settings. Caspeco is a company that provides different services to the restaurant business. Until recently they have tried different forecasting solutions which usually includes trends of some time interval. In this thesis, we investigate if it is possible to create a forecasting solution based on supervised learning. Two different methods are tested, Extreme Gradient Boosted Trees and Long Short Term Memory Neural Network. The two methods are evaluated against each other and compared to the current uplift model used by Caspeco. The data used for training the supervised learning methods is a combination of data provided by Caspeco, and data collected from the Swedish Meteorological and Hydrological Institute (SMHI). This is data such as temperature, minutes of sunshine, rainfall etc, all of which are known to have an impact on the sales of a restaurant.The results show that the models are dependent on the settings of the restaurants. That is the size of the restaurant, the type of the restaurant and if they have an outdoor seating area etc. Both models show a better result of approximately 10-15 percentage points, with regard to the current unit of measurement of Caspeco's current uplift model. Although the models surpass its comparison metric,opportunities for an even greater result is real. Unexpected sales in the form of events are known to influence the results. If different types of event data would be provided we conjecture that the supervised learning models can give a much higher prediction result