This study aims to examine whether large language models (LLMs) can improve tourism demand forecasting accuracy compared to established statistical and machine learning approaches.
A systematic out-of-sample comparison is conducted using monthly tourism arrival data from Eurostat for eight European source markets. Benchmarks include optimized seasonal autoregressive integrated moving average (SARIMA), exponential smoothing (ETS), Prophet, Random Forest and seasonal naive models. Three LLM families are evaluated as zero-shot forecasters. Robustness is assessed across a pre-COVID stable period, a future period beyond LLM training cutoffs and an older model test.
LLM-based forecasts demonstrate consistent performance advantages over all benchmark models across countries and evaluation periods. In the main post-COVID period, LLMs achieve mean MASE values of 0.67–0.85 compared to 1.56 for ETS, 2.67 for SARIMA, 3.04 for Prophet and 4.32 for Random Forest. The LLM advantage persists in the pre-COVID stable period (LLM MASE 0.85–0.90 vs SARIMA 1.71 and ETS 1.31), indicating that the benefit extends beyond structural-break contexts. Robustness tests using an older LLM model and forecasts for a future period beyond LLM training data indicate that data contamination does not explain the observed performance patterns. Prompt ablation analysis reveals that providing contextual information (country identity and date labels) is the most beneficial design choice, while seasonal and shock hints offer no consistent improvement. Diebold–Mariano tests show directional evidence favoring LLMs in 31 of 32 pairwise comparisons, though limited test sample size (14–23 observations per country) precludes individual-country statistical significance at conventional levels.
The analysis focuses on monthly national-level arrival data and zero-shot prompting. Future research should examine alternative frequencies, destination-level data, fine-tuned or retrieval-augmented LLM designs and hybrid approaches combining LLM outputs with statistical model adjustments. The limited test sample size (12–24 monthly observations per country) restricts statistical power for individual-country significance tests, though the consistency of directional advantages across all countries provides compelling qualitative evidence.
LLMs offer a scalable decision-support tool for destination management organizations with near-zero implementation cost via API access.
To the best of the authors’ knowledge, this study provides the first multicountry European benchmark comparing LLM-based forecasting with optimized statistical and machine learning models across multiple evaluation periods, with systematic robustness controls for data contamination.
