Developing a hybrid model for forecasting transportation demand based on SARIMA and XGBoost intelligent methods

Abstract: Forecasting demand for road freight transportation is a highly complex task due to the nonlinear dynamics of time series, pro-nounced seasonality and the influence of other exogenous fac-tors. Traditional statistical time series models effectively iden-tify autocorrelation structures and seasonal patterns; however, they do not always show satisfactory results when modeling nonlinear dependencies. Machine learning algorithms, in con-trast, successfully identify hidden patterns in multidimensional data, but require significant amounts of training samples and demonstrate instability during extrapolation. Based on the above, the aim of this study is to develop a hybrid model for forecasting transportation demand that combines the ad-vantages of statistical and algorithmic approaches. The meth-odology is based on the sequential integration of SARIMA to isolate the temporal data structure and XGBoost to model resid-ual variation, taking into account the constructed set of features (time, lag variables, aggregated statistics). The empirical base consists of data from the digital cargo transportation platform for the period January 2024 to December 2025, including about 10,000 daily observations. On the author's test sample (Octo-ber–December 2025), the hybrid model proved to be more ac-curate than isolated SARIMA, XGBoost and Random Forest across all metrics: the average absolute percentage error was 5.21% with R2 = 0.928, which is 38.1% lower than SARIMA and 11.2% lower than XGBoost. Among the predictors, it was found that weekly seasonality and medium-term trends contrib-ute the most to the model's predictive ability. The results are applicable for optimizing the planning of transport capacities and route networks by logistics platform operators.

Keywords: machine learning, time series, xgboost, sarima, hybrid model, cargo transportation, demand forecasting