Forecasting The Number of Tourist Trips in Pangkal Pinang City: ARIMA vs. LSTM vs. PROPHET

Abstract

Accurate forecasting of tourism demand is crucial for regional economic planning, yet existing studies often rely on univariate time-series models or single forecasting techniques without considering spatial interdependence among regions. This study proposes a correlation-based comparative forecasting framework to evaluate the performance of ARIMA, Long Short-Term Memory (LSTM), and Prophet in predicting domestic tourist trips in Pangkal Pinang City, Indonesia. Monthly data from January 2019 to July 2025 were obtained from Statistics Indonesia, and highly correlated neighboring regions were systematically selected using a heatmap correlation analysis to enhance model input relevance. After applying Min–Max normalization and an 80:20 train–test split, all models were evaluated using Root Mean Squared Error (RMSE) under identical experimental settings. The results indicate that Prophet consistently achieves the lowest RMSE, demonstrating superior capability in capturing non-linear dynamics, seasonal variability, and abrupt structural changes in tourism demand compared to ARIMA and LSTM. These findings provide empirical evidence that decomposable time-series models with automatic trend and seasonality handling offer distinct advantages over both classical statistical and deep learning approaches in medium-term tourism forecasting. The proposed framework contributes a concise, data-efficient, and replicable methodology that supports evidence-based tourism planning and strategic decision-making.