Five-song offset calibration gave lower squared arousal error on average over all MET v1.0.4 10/10 person splits

We tested whether five personal music ratings improve forecasts for 15 other songs. We used 20 people in MET, OpenNeuro ds008701 version 1.0.4. The model added half of the personal calibration offset to peer song means. Ordinal scores were treated as equally spaced. The initial fixed 10/10 person split had an arousal MSE gain of -0.1762. Its approximate 95% person bootstrap interval was [-0.3852, +0.1999]. The planned reverse split had a gain of +0.3423. We then made an exploratory audit of all 184,756 balanced person splits. Average arousal MSE fell from 2.59788 to 2.50731, a 3.49% decrease. It fell in 81.43% of directed splits and 95.05% of 92,378 complementary two-fold partitions. Average valence MSE increased by 1.42%. These exact split counts share the same 20 people and fixed songs. They are not confidence estimates for new people or songs. The benchmark requires peer ratings of the test songs. It does not test EEG forecasts or infer a neural mechanism. The package preserves source data, hashes, local plans, code, forecasts, all split scores, figures, and reproduction instructions. Independent calculations checked every saved split value and the bootstrap arrays. A full replay used freshly retrieved source files and compared all numeric result file hashes.

music emotion ratings · personal calibration · arousal · valence · mean squared error · person split · benchmark