Paul, the xcorr plot do run in both directions. In the notebook with the link strength the optimal lag times are as you describe, but I think that is because they the xcorr they are mostly random fluctuations. The xcorr peaks at about 0.1. That is no bigger than than the tails of the autocorrelation plot of the enso34 signal. The peak of the random forest xcorr is around 0.6 and that model is predicting 6 months in the future while the link strength is being compared with the present.