This study develops a Skip-Connected Temporal Contextual Deep Learning (SC-TCDL) model to forecast monthly inbound foreign tourist arrivals to South Korea, targeting demand volatility and structural shocks such as COVID-19 while supporting planning-oriented decision making. SC-TCDL adopts a dual-stream architecture that disentangles inputs by function: an LSTM branch encodes a 12-month rolling history of arrivals with calendar indicators, while an encoder-only Transformer processes forward-looking exogenous variables with positional encodings. The LSTM temporal representation is injected into the Transformer and fused with the Transformer output via an MLP through skip connections. COVID-period distortion (Mar 2020 Dec 2023) is addressed by virtual demand restoration using a counterfactual LSTM trained on pre-pandemic data. Probabilistic forecasts are generated via Monte Carlo Dropout. Using monthly data (Feb 2013 Apr 2025), SC-TCDL outperforms SARIMA, vanilla LSTM, and a Transformer on the test period (May 2024 Apr 2025), achieving MAE 78,626, RMSE 94,019, and MAPE 6.94%, reducing MAE by 30.5% relative to SARIMA, 28.3% relative to vanilla LSTM, and 24.9% relative to the Transformer, with statistically significant improvements by Wilcoxon signed-rank tests. By structurally separating temporal and contextual learning while enabling controlled fusion and uncertainty quantification, SC-TCDL offers a robust framework for tourism demand forecasting in shock-prone environments.

1. Introduction
2. Related Work
    2.1 Hybrid Deep Learning Models for TourismDemand Forecasting
    2.2 The Mechanics of the LSTM Network
    2.3 The Transformer Model
3. Proposed Model Architecture
    3.1 Disentangled Input Streams
    3.2 Skip-Connected Fusion and Prediction
4. Experimental Design
    4.1 Data Description and Preprocessing
    4.2 Restoration of Pandemic-Period Data andVisual Validation of Reconstructed Demand
    4.3 Forecasting Configuration and TrainingStrategy
5. Forecasting Results and Analysis
    5.1 Comparative Evaluation of Overall andTemporal Forecast Accuracy
    5.2 Component-wise Contribution Analysis:LSTM and Transformer Modules
    5.3 Contextual Variable Contribution Analysis
    5.4 Forecast Visualization and UncertaintyAnalysis
6. Conclusion
    6.1 Summary of Contributions
    6.2 Limitations and Future Directions
References

• Kyeongmin Yum(Department of Global Business, Tongmyong University) | 염경민 (동명대학교 글로벌 비즈니스학과) Corresponding author