Exploring CNN-BiLSTM in Forecasting Tourist Arrivals

The tourism industry makes a significant contribution to the global economy, including Indonesia, through various sectors such as hospitality, transportation, and services. Accurate tourist arrival forecasting is crucial for governments and industry players for strategic planning. In this study, a CNN-BiLSTM model was developed as a more accurate forecasting method using historical airport and port arrival data. This research explores ten different hyperparameter combinations to find the best configuration, with hyperparameters including dropout rate, batch size, learning rate, optimizer, activation function, and data split proportions. The experimental results indicate that the optimal hyperparameter combination, particularly with a small batch size (10), a higher learning rate (0.005), the Adam optimizer, and the Tanh activation function, and 60:40 train-to-test data ratio, delivers the best performance with the lowest RMSE and MAPE for the Jakarta and Riau Islands regions. The CNN-BiLSTM model has proven effective in handling seasonal data and fluctuations in tourist arrival patterns, especially in Jakarta, which exhibits a more stable pattern. Conversely, the model's performance in the Riau Islands is affected by more unstable seasonal pattern variations and extreme anomalies, resulting in a higher MAPE. This study demonstrates that selecting the appropriate model configuration and adapting to data patterns are essential for improving tourist arrival forecast accuracy, which can support better decision-making in the tourism sector.