A SARIMA-Based Rolling One-Step-Ahead Framework for Monthly Forecasting of Prefecture-Level Guest Nights in Japan

Abstract

Accurate monthly guest-night forecasting supports tourism planning and accommodation management. This study compares local seasonal autoregressive integrated moving average models with global multilayer perceptron, support vector regression, and extreme gradient boosting models for 47 Japanese prefectures using monthly observations from January 2007 to January 2026. The benchmark considers model type and training scope, although their individual effects are not isolated. Performance is evaluated over January to December 2024 using a rolling one-step-ahead protocol. Each forecast uses only observations available up to the end of the preceding month; the actual value is incorporated after prediction. The local seasonal model achieves the lowest normalized mean absolute error, normalized root mean squared error, and symmetric mean absolute percentage error of 5.12%, 9.42%, and 6.16%, respectively. These results support local seasonal modelling for this dataset, which exhibits seasonality, regional heterogeneity, and pandemic-related disruption. Supplementary projections for February to December 2026 use models refitted on observations through January 2026 and are not treated as an accuracy evaluation. Overall, the local seasonal model outperforms the evaluated global machine-learning configurations, highlighting the importance of matching forecasting approaches to data characteristics.