Accurately predicting daily international tourist arrivals is challenging for tourism-dependent economies like Sri Lanka, especially during its post-crisis recovery (2023âÂÂ2025). Existing methods often ignore non-Google search engines, missing critical signals from key source markets like Russia, where Yandex dominates. This paper introduces a multi-model forecasting framework that integrates official daily arrivals with Yandex data, localized Google Trends, exchange rates, and weather records. We evaluate seven architectures, Pure SARIMA, SARIMAX, XGBoost, Random Forest, LSTM, Support Vector Regression (SVR), and a Sequential Hybrid, over a 550-day rolling test period. To prevent data leakage, feature selection utilizes Granger causality and cross-correlation analysis strictly on training data. The results demonstrate that SVR with an RBF kernel performs best, achieving a rolling MAPE of 8.94\% and an RMSE of 736 arrivals/day. Crucially, Yandex queries for tours and flights emerge as the strongest exogenous predictor across all machine learning models, completely outperforming Russian Google Trends data, which proved uninformative. These findings confirm that aligning predictive features with a tourist market's preferred search engine significantly enhances forecasting accuracy, providing highly actionable insights for modern tourism intelligence systems.
Keywords: Natural Language Processing | Machine Learning / Deep Learning | Data Science |
@inproceedings{pererae2026ffect,
title={{The Effect of Search Intent Data on Predicting Post-Crisis Daily Tourist Arrivals in Sri Lanka}},
author={Perera, Tharusha and Mahanama, Janith and Kumarasinghe, Maleesha and Udurawana, Bimsara and Nawarathna, Praveen and Narasinghe, Patalee and Athukorala, Nathali and Wickramanayake, Sandareka and de Silva, Nisansa},
booktitle={2026 Moratuwa Engineering Research Conference (MERCon)},
pages={395-400},
doi={10.1109/MERCon71835.2026.11691456},
year={2026},
abstract={Accurately predicting daily international tourist arrivals is challenging for tourism-dependent economies like Sri Lanka, especially during its post-crisis recovery (2023âÂÂ2025). Existing methods often ignore non-Google search engines, missing critical signals from key source markets like Russia, where Yandex dominates. This paper introduces a multi-model forecasting framework that integrates official daily arrivals with Yandex data, localized Google Trends, exchange rates, and weather records. We evaluate seven architectures, Pure SARIMA, SARIMAX, XGBoost, Random Forest, LSTM, Support Vector Regression (SVR), and a Sequential Hybrid, over a 550-day rolling test period. To prevent data leakage, feature selection utilizes Granger causality and cross-correlation analysis strictly on training data. The results demonstrate that SVR with an RBF kernel performs best, achieving a rolling MAPE of 8.94\% and an RMSE of 736 arrivals/day. Crucially, Yandex queries for tours and flights emerge as the strongest exogenous predictor across all machine learning models, completely outperforming Russian Google Trends data, which proved uninformative. These findings confirm that aligning predictive features with a tourist market's preferred search engine significantly enhances forecasting accuracy, providing highly actionable insights for modern tourism intelligence systems.},
misc={https://tharusha.cse23.org/Search_Intent_Effect_on_Post_Crisis_Daily_Tourist_Arrival_Prediction_for_Sri_Lanka.pdf,NLP:ML,https://goo.gl/iY6aTr}
}
This dataset provides a unified, daily time-series panel designed for forecasting international tourist arrivals to Sri Lanka over the post-crisis recovery period of **January 2023 to December 2025**. It integrates five disparate data streams to capture the macroeconomic, environmental, and behavioural signals that drive tourism demand.
ActivePI: Nisansa de Silva
The Foundational Data Science project investigates how systematic data acquisition, integration, analysis, modelling, and visualisation can be applied to real-world problems to generate reliable and actionable knowledge.