We introduce five regression models for the modeling of expressed emotion in music using data obtained in a two alternative forced choice listening experiment. The pre-dictive performance of the proposed models is compared using learning curves, showing that all models converge to produce a similar classification error. The predictive rank-ing of the models is compared using Kendall's τ rank cor-relation coefficient which shows a difference despite simi-lar classification error. The variation in predictions across subjects and the difference in ranking is investigated vi-sually in the arousal-valence space and quantified using Kendall's τ .