Emotions Are Not Random: Machine Learning Reveals Predictable Patterns in a 21-Day Second Language Learning Trajectory

Abstract Emotion research in second language learning has largely focused on static, trait-based measures, especially anxiety, while neglecting contextual and temporal dynamics. This study reconceptualizes emotions as emergent and context-dependent, shaped by real-time interactions with learning activities and self-perceptions. Using Ecological Momentary Assessment (EMA), 92 adult learners reported their emotional states (anxiety, enjoyment, boredom), perceived proficiency, and contextual features (e.g., task modality, duration) across 6,918 learning episodes over 21 days. We identified five emotional profiles, including a dominant “Routine but Pleasant” state, challenging anxiety-centred paradigms. Emotional variability was primarily intraindividual (62–68%), with weak temporal trends and idiosyncratic cycles. Time-series machine learning (TabPFN-TS) achieved high predictive accuracy for emotional states (e.g., anxiety R² = 0.87; boredom R² = 0.95) and perceived proficiency (R² = 0.83). These findings underscore the value of modelling emotional dynamics in context and suggest that short-horizon forecasting may help identify moments when timely learner support could be explored in future work.