Publicação
Hotel Demand Forecasting using Foundation Models
| datacite.subject.fos | Ciências Naturais::Ciências da Computação e da Informação | pt_PT |
| dc.contributor.advisor | António, Nuno Miguel da Conceição | |
| dc.contributor.author | Pimenta, Diogo André Domingos | |
| dc.date.accessioned | 2025-11-14T11:10:06Z | |
| dc.date.embargo | 2028-10-30 | |
| dc.date.issued | 2025-10-30 | |
| dc.description | Dissertation presented as the partial requirement for obtaining a Master's degree in Data Science and Advanced Analytics, specialization in Business Analytics | pt_PT |
| dc.description.abstract | The hospitality industry is increasingly reliant on accurate demand forecasting to support revenue management and strategic planning. Traditional forecasting methods often struggle to capture the nonlinear and seasonal patterns typical of hotel booking data. To address these limitations, this study evaluates the performance of three types of models: statistical models (e.g., ARIMA, Prophet), which rely on predefined mathematical structures; machine learning models (e.g., Random Forest, XGBoost), which learn patterns from historical data; and Foundation Models (e.g., TimeGPT), large-scale pretrained models applied in a zero-shot forecasting setting. The objective is to evaluate how pretrained FMs perform in forecasting hotel occupancy rates across different horizons (1-day, 7-day, and 28-day), using real reservation data from 100 hotels. The CRISP-DM methodology and the Kedro framework were employed to facilitate standardized ingestion, data cleaning, feature engineering, and modeling. Forecasting targets were derived from the occupancy rate. Results show that XGBoost consistently achieved the best balance between accuracy and computational efficiency across all horizons. Random Forest offered solid performance but with higher computational costs than XGBoost. TimeGPT, despite operating without local training or finetuning, performed competitively, especially in larger hotels, with more financial activity and longer booking windows. Prophet performed well in smaller hotels with clear seasonal patterns, while ARIMA was most effective in the simplest environments. TimeGPT had the highest inference time, and while recent API updates allow the extraction of SHAP values for partial interpretability, it remains a black-box model. No single model outperformed across all contexts, confirming that model suitability depends on hotel characteristics. Limitations of this work include the absence of variables such as events or weather; limited number of Foundation Models tested due to computational constraints. Still, the study provides evidence supporting the potential of Foundation Models,such as TimeGPT, in hotel demand forecasting and offers a replicable methodology for model comparison in time series forecasting tasks within the hospitality industry. | pt_PT |
| dc.identifier.tid | 204073375 | |
| dc.identifier.uri | http://hdl.handle.net/10362/190723 | |
| dc.language.iso | eng | pt_PT |
| dc.rights.uri | http://creativecommons.org/licenses/by/4.0/ | pt_PT |
| dc.subject | Hotel Industry | pt_PT |
| dc.subject | Hotel Demand Forecasting | pt_PT |
| dc.subject | Artificial Intelligence | pt_PT |
| dc.subject | Large Language Models | pt_PT |
| dc.subject | Time Series Models | pt_PT |
| dc.subject | SDG 8 - Decent work and economic growth | pt_PT |
| dc.title | Hotel Demand Forecasting using Foundation Models | pt_PT |
| dc.type | master thesis | |
| dspace.entity.type | Publication | |
| rcaap.embargofct | Publicação em jornal cientifico | pt_PT |
| rcaap.rights | embargoedAccess | pt_PT |
| rcaap.type | masterThesis | pt_PT |
| thesis.degree.name | Mestrado em Ciência de Dados e Métodos Analíticos Avançados, especialização em Business Analytics | pt_PT |