Extended Data Fig. 4 |. Selecting stimulus window length for prediction.

A window size of 510 ms was chosen to maximize linear model performance with the minimal number of parameters. We fit multiple models, each with a different number of time-lags (window size), from 60 ms to 760 ms. Each model was trained with the full list of predictors shown in Fig. 1c on all electrodes selected by the selection criteria described in Methods (n = 242), and only differed from the other models in the number of lags. Error bars indicate standard error (SE) over electrodes. To compare two different sizes, we perform a paired-sample one-tailed t-test on the cross-validated out-of-sample prediction r-values to determine whether the larger model improves upon the smaller one. The 510 ms model (dashed line) showed a significant improvement over all smaller models (60 ms – 410 ms, p < 0.001; 460 ms, p = 0.023). No larger model showed significant improvement over the 510 ms model (p > 0.5).