Good days ahead : forecasting ticket sales for Go Fjords using weather data

Abstract

The purpose of this study is to evaluate whether public weather data from MET Norway can be used to improve ticket sales forecasts for the travel company Go Fjords, and to demonstrate how such a forecasting model can be technically implemented to provide value over time. The problem statement is defined as follows: Can weather forecast data make demand forecasts more accurate for Go Fjords, and how can business value be derived from such a forecast? Throughout the study, a wide range of methods were used. This thesis outlines how to retrieve historical weather data from MET Norway’s ‘Frost API’, how to scrape weather forecast data off yr.no, and how to assemble the data for use by forecasting models. The following model types and frameworks were tested: A Generalized Additive Model (Facebook Prophet), a Dynamic Generalized Linear Model (PyBats), and a Random Forest Regression model created by Microsoft Azure’s automated machine learning functionality. Performance metrics are discussed in depth, and Root Mean Squared Error was chosen as the basis for evaluation and comparison. A set of univariate ‘benchmark’ models were created to answer the problem statement: a naïve forecasting model and a seasonal ARIMA model. The Facebook Prophet model was used to demonstrate deployment and was implemented to run daily in Microsoft Azure. The forecasts were pushed daily to Go Fjords’ database, and made visible in their Microsoft Power BI dashboard, along with actionable advice on the optimal number of buses to rent, taking future weather into account. The Prophet model performed worse than expected, and the PyBats model performed very well. Potential causes and ways to adjust the models are discussed. ARIMA and Random Forest Regression had similar RSME scores, strengthening the validity of their results. To conclude: It is possible to create better demand forecasts for Go Fjords by using weather data, rather than by basing forecasts on sales data alone. By optimizing the models for RMSE, variance is minimized, consequently minimizing the frequency at which Go Fjords deploys the wrong number of buses, thus capturing more revenue and achieving cost savings.