<![CDATA[University ranking prediction requires adaptive models to capture temporal dynamics and handle data anomalies. This study develops a time-adaptive ensemble framework integrating outlier-aware scoring and hybrid feature selection. We collected data from Times Higher Education from 2011 to 2024, applying windowed outlier detection with clipping and masking, and using ANOVA F-tests, permutation importance, and SHAP values to select dynamic feature subsets. The framework trains linear moving-average, temporal Random Forest, and LSTM models, then ensembles their forecasts using dynamically optimized weights. Experimental results on rolling forecasts (2016–2024) demonstrate a mean rank deviation of 1.2 positions, Top-1000 classification accuracy of 0.96, and reduced MAE and RMSE compared to single-model baselines. SHAP-based analyses reveal evolving feature importance across time windows, highlighting the impact of changing indicators. The findings indicate that integrating outlier handling, dynamic feature selection, and ensemble learning enhances prediction robustness and interpretability. This framework can support strategic decision-making, policy formulation, and resource allocation in higher education]]>
