Our approach combines a diverse ensemble for Subtask 1 with a context-aware transformer aggregation architecture for temporal forecasting in Subtasks 2A and 2B. The ensemble achieved state-of-the-art performance for the Subtask 1 Valence metric, ranking first in Valence prediction. Our Subtask 2B independent architecture ranked second in Valence prediction and fourth in Arousal prediction among competitive submissions. We also report results for Subtask 2A, analyzing challenges our architecture faced with next-entry affect forecasting. These findings underscore the significance of our methodology for affective prediction, achieved without reliance on external affective datasets.