Fig. 3.

The pipeline of the analysis uses windowed time-lagged cross-correlation (WTLCC) (Upper panel) WTLCC is applied to two series I₁(n) and I₂(n) of length N, creating a matrix with K rows and J columns. (Lower panel) A custom convolutional neural network (CNN) was used as a general approximator for the task of finding a mapping from the WTLCC matrices to a binary variable of unfavourable/favourable outcome. The details of WTLCC and CNN are presented in the Supplementary materials