From the paper: >"First, robust regression between each edge in the connectivity matrices and d′ was performed across subjects. The resulting r values were sta

From the paper: >"First, robust regression between each edge in the connectivity matrices and d′ was performed across subjects. The resulting r values were statistically thresholded at P < 0.01 and separated into a positive tail (edges whose strength indexed higher d′ across subjects) and a negative tail (edges whose strength indexed lower d′ across subjects)." If they tried out different thresholds (I bet they did p<.05, p<.001, etc),then information leaked into the holdouts. So it is unsurprising they could get better than chance performance.