Variance cap 10 decreased SEEG log-size loss but increased mean RMS error against Gaussian forecasts in eight people

We tested a fixed latent variance cap of 10 on public SEEG response forecasts. The test used 83 sites in eight people, with five fit and five later test pulses per site. The cap gives finite RMS moments under the model. Mean log-size loss was 0.231954 nats, against 0.432403 for a Gaussian and 0.228996 for Student t. All eight people had lower loss than the Gaussian. This does not prove equivalence to Student t. Normalized MSE for the mean RMS forecast was 6.962812. The Gaussian mean MSE was 0.191386. The common median MSE was 0.176787. A larger cap changed loss little but raised mean error sharply. Finite moments and good probability scores did not give reliable mean RMS forecasts in this test. The package includes exact compact data, source hashes, a new full-source rebuild for one person, code, checks, and figures. The cohort is reused and quality labels are retrospective. This is an offline response test, not a session-wide deployment or stimulation safety test.

SEEG · forecasting · latent variance · Student t · RMS voltage · brain-computer interface · prediction error · OpenNeuro