Comparison and evaluation of time series forecasting models and their application in beauty retailing

Abstract: The beauty industry has shown steady growth within the last decade and as more retailers move their product selections online, a lot of sales data can be generated. To improve sales and to meet the ever-increasing competition in the beauty retail sector, time series forecasting has potential in providing knowledge and guidelines in decision making to improve business performance. Time series forecasting, or extrapolation of time series data, is the process of fitting a model on historic data and using it to predict future values. The purpose of this thesis is to compare and evaluate multiple established time series forecasting models to find the most accurate model. The models that were selected as candidates for times series forecasting were chosen based on a literature review of scientific articles. Based on the literature review, SARIMA, Prophet, and LSTM were chosen as forecasting models. Following a research strategy of experiments, the chosen models were implemented at a specific representative company of beauty retailers in the Nordic countries, Lyko. The results showed that there was no model that performed the best in every case. It is possible to forecast the sales of beauty products at least somewhat confidently with the given models. By using forecasting models and accounting for the forecasting error, beauty companies can have a solid basis for business decisions and gain a competitive edge in the beauty market. The main takeaway from this thesis is that there is no one-fits-all model. Instead, if possible, all models should be tested on the data to see which model performs the best. The information of this thesis is useful for the decision maker of a beauty retailer when it comes to applications of forecasting algorithms. A model could be chosen based on the specific demands of the business based on accuracy, runtime, or complexity.