Large windows and small overlap means you don't have enough initial condition associations.

[djpasseyjr]

When this happens the reservoir does not learn to associate training data states with internal states. We could add lots of initial condition association points to the end of the R and U matrixes to solve this problem. This could potentially sidestep the need for windows and overlap. Maybe call it the anchor matrix, because it anchors one system to the other.

[djpasseyjr]

You probably still need windows because otherwise you end up with one orbit exploring the reservoir space and a bunch of associations of single points without learning how those points evolve in the space.