Machine learning for forecasting visits to recreational zones in mountain areas considering calendar cycles and meteorological conditions

Abstract

Introduction. This study addresses the problem of forecasting tourist flow in mountain areas under conditions of pronounced seasonality and high sensitivity to weather variability. The research focuses on the relationship between visitation dynamics, calendar structure, snow cover, temperature, precipitation, wind speed, and delayed effects of previous tourist activity. The relevance of the study is determined by the need for more accurate short-term forecasting tools for mountain tourism territories, where traditional statistical approaches often fail to capture sharp peaks and rapid declines in demand. Methods. A combined observational dataset was formed for four complete tourist years and included 1461 daily records. The dataset contained information on tourist visits, air temperature, precipitation, wind speed, snow depth, weekends, holidays, and seasonal indicators. The data were synchronized by time step, cleaned from technical gaps and abnormal records, and transformed using lagged variables and cyclic encoding of seasonality. Several forecasting models were compared, including linear regression, seasonal autoregressive modeling, Random Forest, Gradient Boosting, XGBoost, LightGBM, and an LSTM-based neural network. Results. The results showed that tourist flow in mountain areas has a strong weekly and seasonal structure. Saturday visits were 1.82 times higher than weekday visits, while Sunday visits were 1.56 times higher. Winter peaks reached 5200–6900 visits per day, whereas summer maximum values ranged from 4100 to 5400 visits per day. The most accurate models were LightGBM and XGBoost, with MAPE values of 9.1–9.4% and determination coefficients of 0.86–0.87. Winter visitation was mainly affected by snow depth, temperature, and holidays, while summer visitation was primarily influenced by precipitation, thermal comfort, and route conditions. Conclusions. The study demonstrates that machine learning methods are more effective than traditional statistical models for forecasting tourist flow in mountain areas. The use of lagged weather and visitation variables, as well as cyclic representation of seasonality, improved forecast accuracy and allowed nonlinear relationships between weather, calendar factors, and tourist demand to be captured. The proposed approach can be used for short-term planning of visitor flows, route load assessment, and operational management of mountain tourism infrastructure.