A previous error correction increased mean one-state forecast loss in 32 selected ds004021 v1.0.1 people after 160 training trials

We tested whether the previous measured error improved a forecast of the next motor adaptation trial. We used both days from 32 selected people in OpenNeuro ds004021 v1.0.1. We fitted a smooth one-state curve to the first 160 rotation trials. We then fitted a correction from adjacent training residuals. The test used the next 160 trial positions. The correction increased mean squared error from 556.73 to 561.19 degrees squared. The mean increase was 4.46 degrees squared, or 0.80 percent. The 95 percent person bootstrap interval for the increase was 0.33 to 9.28 degrees squared. Only 9 people had lower loss with the correction. Thus, this fitted correction did not improve mean later forecasts under the stated model and split. The result does not show that motor learning has no serial structure. The package preserves source files, forecasts, bootstrap draws, code, and independent checks.

motor adaptation · visuomotor rotation · forecasting · serial error · one-state model · held-out prediction · motor learning