A Hybrid Approach for Predicting Tourist Arrivals in Sri Lanka: Integrating Machine Learning with Time Series Modelling

Abstract

This study aims to refine forecasting models for tourist arrivals in Sri Lanka, a sector pivotal to the nation's economic and social vitality. Utilizing advanced hybrid modelling techniques, the study integrates traditional statistical models (SARIMA, VAR) with machine learning approaches (SVR, ANN, LSTM) to enhance predictive accuracy and robustness. A dataset spanning from January 2012 to December 2021 is utilized for this purpose. The methodology involves a comparative analysis of single models against hybrid configurations that capitalize on the strengths of both linear and nonlinear modelling techniques. The models were rigorously evaluated using the root mean squared error (RMSE) and the mean absolute error (MAE) across different forecasting horizons (short-term, mid-term, and long-term). This evaluation was conducted using a diverse array of features, including climate data, TripAdvisor reviews, and Google Trends data, offering a comprehensive view of the factors influencing tourist arrivals. Results indicate that hybrid models, particularly those combining linear and nonlinear methodologies, substantially outperform single model approaches. These models demonstrated a superior capacity to capture the complex interdependencies and nonlinear patterns affecting tourist behavior, thus providing more accurate and actionable forecasts. The relevance of this research extends beyond academic circles to practical applications, aiding policymakers, tourism operators, and economic planners in strategic decisionmaking. By enhancing forecasting accuracy, this study contributes significantly to the sustainable growth and resilience of Sri Lanka's tourism industry.