With one trial of feedback delay, mixing two fixed models limited added log loss in 15 final SEEG sessions

We tested response feedback for fixed task and SEEG forecasts in 15 final numbered sessions from OpenNeuro ds004752 version 1.0.1. The test had 690 trials and 59 wrong answers. With equal priors and one trial of feedback delay, a probability mixture limited added cumulative log loss to ln(2), about 0.693 nats, relative to either model at every prefix. The largest observed added loss relative to the task model was 0.691262 nats. Mean log loss was 0.2910 for the mixture and 0.2868 for the task model. The mixture did not show a clear gain over fixed equal averaging: the mean difference was -0.00192 nats per trial, with a 95% person bootstrap interval from -0.00679 to 0.00201. We provide a causal forecast tool, a delay proof, figures, full source provenance, fixed input models, and code. The general mixture method is established; this SEEG benchmark is new. The result concerns cumulative probability loss. It does not establish better memory or a stimulation benefit. Task phase use assumes fixation alignment of source epochs. Delayed feedback tests keep the input models' task history available.

probability forecasting · log loss · forecast aggregation · stereoelectroencephalography · working memory