Tourism demand forecasting remains a critical yet complex task for data-driven destination management organizations. Despite recent advances in time series modelling, the application of Foundation Models (FMs) to structured forecasting tasks remains underexplored. This study addresses this gap by assessing the performance of FMs in predicting monthly overnight stays in Lisbon, a city where tourism plays a central role in the local economy. A 15-year dataset was developed, integrating historical overnight stays with heterogeneous exogenous variables, including macroeconomic indicators, Google Trends search indices, meteorological data, and event-based signals. A comparative evaluation was conducted using PromptCastGPT4.1, TimeGPT, TimeLLM, and TimesFM, benchmarked against AutoRegressive Integrated Moving Average (ARIMA) with automatic configuration (AutoARIMA), Seasonal ARIMA with exogenous variables (SARIMAX), and Machine Learning (ML) models such as Random Forest (RF), Extreme Gradient Boosting (XGBoost), Gradient Boosting (GB), and Light Gradient Boosting Machine (LightGBM). The results show that TimeGPT and PromptCast-GPT4.1 consistently outperformed all baselines across these metrics, particularly excelling in capturing seasonal demand patterns. The study contributes to academic research by extending the use of FMs to structured time series problems. It offers practical insights into their applicability in operational forecasting, even in zero-shot configurations.

Dissertation presented as the partial requirement for obtaining a Master's degree in Data Science and Advanced Analytics, specialization in Business Analytics