Abstract
This study proposes an interpreted machine learning framework for monthly forecasting of non-resident tourist arrivals in Albania, using the XGBoost algorithm in combination with SHAP. The model received training data from 2016 to 2024, which included time variables, lagged values and seasonal components to detect intricate tourist flow patterns. The evaluation results demonstrate strong predictive accuracy, and the SHAP analysis reveals how each feature affects the outcome by showing that time variables and historical values play the most influential roles. The analysis examined external factors, including gross domestic product (GDP) and exchange rates, but these variables had a lesser impact than the dominant seasonal and lagged data. The method provides both precise forecasting capabilities and clear interpretive insights, which makes it ideal for strategic market decisions in developing tourism sectors.