Optimizing Long Short-Term Memory Forecasting with Boruta Feature Selection: A Case Study on Tourist Arrival Prediction

Abstract

Accurate forecasting of international tourist arrivals is crucial for strategic planning and policy-making, particularly during the post-pandemic recovery phase. This study compares three Long Short-Term Memory (LSTM) modeling approaches: univariate, multivariate with all features, and multivariate with feature selection using the Boruta algorithm. The dataset includes monthly tourist arrival records in Indonesia (2008–2025) and tourism-related search indices from Google Trends. The results show that the univariate LSTM model performs best (R² = 0.552), while the multivariate model with Boruta-selected features performs worst (R² = -0.913). These findings underscore that adding features without considering temporal dynamics may reduce prediction accuracy, and that simpler, single-variable models can be more effective for time-series data. This research offers both practical and theoretical contributions to developing more accurate and context-aware AI-based tourism forecasting systems.