Försäljningsprognoser i detaljhandeln : En utvärdering av geografisk kontext och kalendereffekter i närbutiker med hjälp av Prophet och Google Places API

In the competitive retail sector, forecasting sales can be challenging due to volatile, location-specific consumer behavior and time-based fluctuations. This thesis examines how geographic context and calendar effects impact demand at Pressbyrån stores in Stockholm. Using the Google Places API, ten point-of-interest categories within 200 meters of each store were identified. Stores sharing identical binary point-of-interest profiles were grouped into clusters, yielding three large clusters whose daily sales were modeled with Facebook’s Prophet. For each cluster, a baseline Prophet model was compared to one enhanced with custom regressors for popular Swedish calendar events. While adding custom regressors slightly improved accuracy (2-5% reductions in Mean Absolute error, Root Mean Square Error, and Mean Absolute Percentage Error values), other percentage-based errors (Symmetric Mean Absolute Percentage Error) were inconsistent, and significant variability remained within clusters, suggesting other factors (for example, local demographics) influence demand as well. The findings highlight the limitations of a purely calendar and point-of-interest based approach and underscore the need for additional covariates (e.g. promotional activities) or more granular, store-specific model tuning. Future work should explore hybrid or alternative time-series methods and incorporate cross-validation across multiple forecast horizons. By integrating richer spatial, operational, and demographic data, subsequent models may achieve more robust, actionable forecasts tailored to each store’s unique context.