Time-series forecasting with TabPFN-TS markedly improves prediction of next-episode emotions and perceived proficiency over a static model (e.g., anxiety R² 0.458→0.866; boredom 0.596→0.953)¶

Claim · All claims
Evidence · 1 study (2 entries) · associational r1–r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Training TabPFN-TS on each learner's first 85 episodes and forecasting the final 20 raised within-dataset R² from the static model: anxiety 0.458→0.866, boredom 0.596→0.953, self-perceived proficiency 0.515→0.832, enjoyment 0.651→0.934. → Peng Wang 2026 q2 i? Against an individual-mean baseline, TabPFN-TS achieved skill scores of 0.740–0.782 and MAE of 0.26–0.32 on the 0–10 scale across the four targets. → Peng Wang 2026 (2)

Evidence¶

Peng Wang 2026¶

Peng Wang, Lesya Ganushchak, Camille Welie, Roel van Steensel. (2026). Emotions Are Not Random: Machine Learning Reveals Predictable Patterns in a 21-Day Second Language Learning Trajectory. Educational Psychology Review. https://doi.org/10.1007/s10648-026-10141-8

q2 · i? · associational · r2

Time-aware TabPFN-TS forecasting using recent history and dynamic emotional profiles, trained on the first 85 episodes per learner; the article notes R² is used only for within-dataset comparison.

"Relative to a static model, out-of-sample fit improved markedly, with R² increasing as a within-dataset comparator: anxiety: 0.458→0.866; boredom: 0.596→0.953; self-perceived proficiency: 0.515→0.832; enjoyment: 0.651→0.934."

Peng Wang 2026 (2)¶

Peng Wang, Lesya Ganushchak, Camille Welie, Roel van Steensel. (2026). Emotions Are Not Random: Machine Learning Reveals Predictable Patterns in a 21-Day Second Language Learning Trajectory. Educational Psychology Review. https://doi.org/10.1007/s10648-026-10141-8

q2 · i? · associational · r1

Benchmark of the sequence model against an individual-mean baseline; the article states the MAE of 0.26 to 0.32 is small relative to the 0–10 scale, indicating good short-horizon accuracy within this dataset.

"The model consistently achieved high relative skill scores and low errors across all measured states—anxiety (0.771/0.319/0.765), boredom (0.782/0.283/0.689), enjoyment (0.772/0.277/0.674), and self-perceived proficiency (0.740/0.263/0.635), reported as skill score/MAE/RMSE."

Discussion¶

Related Claims¶

• Adult L2 learners' momentary emotions cluster into five profiles, dominated by a low-arousal 'Routine but Pleasant' state (57.8% of episodes), challenging anxiety-centred paradigms — related
• In static concurrent prediction, TabPFN outperformed Random Forest, XGBoost and Linear Regression across the four emotional and proficiency targets — related
• During the course, curiosity, enjoyment, confusion and anxiety decreased over time, boredom increased, and surprise and frustration remained stationary. — related