<![CDATA[Background: Reliable forecasts of new student registration are essential for religious higher education institutions because admission volumes affect capacity planning, resource allocation, promotional strategies, and program sustainability. Historical admission data are often used descriptively rather than to generate forward-looking and interpretable estimates. Objective: This study developed an explainable multivariate time-series framework for forecasting new student registration by integrating domain-based feature engineering, comparative machine-learning and deep-learning models, CNN-LSTM forecasting, and SHAP interpretation. Methods: Historical admission records from 40 study programs at UIN Ar-Raniry for 2017–2025 were analyzed, comprising 360 program-year observations. The workflow included data validation and preprocessing, construction of share, pressure, and lag features, two-year sequence construction, temporal train-validation-test splitting, train-only scaling, comparative model evaluation, 2026 forecasting, and SHAP-based global and local interpretation. Results: CNN-LSTM achieved the best average performance across five iterations, with an MAE of 6.0394, an RMSE of 8.6185, a MAPE of 27.7548%, and an R² of 0.9167. Its advantage is consistent with the structure of the data: convolutional layers extract local relationships among multivariate admission features, while LSTM captures temporal dependencies. SHAP showed that recent historical registration variables, especially reg_span_lag1, reg_um_lag1, and total_reg_lag1, together with capacity-related features such as jml_dt, were major contributors to the forecasts. Conclusion: The study demonstrates that combining domain-informed feature engineering, CNN-LSTM, and SHAP provides both accurate and interpretable registration forecasts. The framework contributes an explainable decision-support approach for admission planning, quota setting, resource allocation, and targeted recruitment in religious higher education.]]>
Copyrights © 2026
