The last response time reduced forecast log loss by 1.14% in 144 PURSUE ds007139 v1.0.0 test people with fixed offline models

We tested whether the last response time improves a forecast of a wrong Flanker response after task conditions and prior performance. We used public PURSUE data from OpenNeuro ds007139 version 1.0.0. Fixed logistic models trained on 143 people and tested on 51,842 selected trials from 144 other people. The added last-time terms reduced equal-person mean log loss from 0.307149 to 0.303663 nats per trial. The gain was 0.003487 nats per trial, or 1.14%. The conditional 95% interval was 0.001950 to 0.004824. This interval assumes independent people within ID groups and holds the training models fixed. All five planned model checks had positive gains. The test selected valid response pairs after the records were complete. It does not measure forecasts of omissions or establish a neural cause. The optional training-refit check did not finish. The package preserves 582 source files, provenance, code, predictions, model inputs, sampling draws, figures, and reproduction instructions. A fresh-source replay matched all 43 numerical result files and all four figure exports. The result supplies a behavioral reference for future neural forecast tests.

Flanker task · response time · error forecasting · logistic regression · log loss · PURSUE · OpenNeuro · behavioral benchmark