Assess whether fine-tuned sequence models improve scheduled-music arousal forecasting

Determine whether a large fine-tuned sequence model improves dynamic arousal forecasting with scheduled future music relative to the frozen MuQ representations and simple regression heads evaluated in the study.

Background

The study uses frozen MuQ music representations followed by PCA and simple regression heads to forecast arousal innovations under strict leakage-control and scheduled-future observation rules. Although these models yield substantial improvements over affect-state baselines, the authors explicitly note that their results do not determine whether a substantially larger, end-to-end fine-tuned sequence model would provide additional forecasting benefits. Establishing this would test whether the observed gains are limited by the representation and regression architecture or reflect a broader property of the forecasting task.