We just published a new paper on how to properly initialize reservoir computers when only a limited history of observations is available. Check out the code here and the paper here.