Past target response history gives no clear gain in error forecasts under fixed ridge models in 15 test people in ds006768 v1.1.0

We tested whether past target responses improve error forecasts in a moving dot task. The source was OpenNeuro ds006768 version 1.1.0. Two fixed ridge models trained on early runs from 15 people. They tested later runs from 15 other people. The test had 2368 targets and 130 errors. The history model used five completed past targets. Each person had the same score weight. The log loss gain was -0.000339 nats per target. The 95% bootstrap percentile interval was [-0.004744, 0.003346]. All 10000 draws sampled training and test people separately and fitted both models again. All five planned check intervals also included zero. The tested history model gave no clear gain. This does not prove that all response history has no value. A time audit found 3769 target onsets where the nearest past target was still on screen. The package supplies a guarded behavioral benchmark, all 210 event tables, source hashes, code, trial forecasts, fitted parameters, and all 15000 draw records. Independent checks passed. All 28 numerical files matched after repeat analysis. No EEG signal or cause was tested.

sustained attention · error forecasting · response history · ridge regression · Multiple Object Monitoring · OpenNeuro ds006768