Comparing Statistical and Machine Learning Approaches to Forecasting Tourist Overnight Stays: The Case of Greece

Abstract

Tourism is one of the key sectors of the Greek economy, and the number of tourist overnight stays is an important indicator of demand volume and accommodation capacity utilization. This paper analyzes and forecasts the monthly number of tourist overnight stays in Greece based on official Eurostat data for the period from January 2000 to October 2025. Greece is a particularly challenging forecasting case, as around 57% of annual overnight stays are realized in only three summer months, while the COVID-19 pandemic introduced a severe structural break in the series. Three methodologically different forecasting approaches are compared: the seasonal ARIMA model (SARIMA), the Prophet model, and the XGBoost regression algorithm, with a Seasonal Naive model serving as a benchmark. Model performance was evaluated using MAE, RMSE, and MAPE metrics on a test period of 24 months, from November 2023 to October 2025. Among the developed models, XGBoost achieved the lowest forecast error, with a MAPE of 16.95%, compared to 42.42% for SARIMA and 52.36% for Prophet. The Diebold-Mariano test confirmed that its advantage over SARIMA is statistically significant. The results also show that a simple Seasonal Naive benchmark remains difficult to beat in stable periods, which has direct implications for the choice of forecasting models in tourism practice.