Thanks Jan.
Way back early in the El Nino Azimuth Project discussion, Dara O'Shayda was doing studies using the *random forest* approach, for example in this discussion thread

https://forum.azimuthproject.org/discussion/comment/13797/

It looks like the files are gone from his old website but they may have all moved here:
http://files.untiredwithloving.org/random_forrest_nino34.pdf

This is an example of one of Dara's plots

![](http://imageshack.com/a/img921/5581/aJFtI4.gif)

Perhaps we can pursue some paths by looking at link strengths, etc, especially now that we think we know what the candidate non-linearities are. Correct me if I am wrong, but I am under the impression that one of the issues with deep learning is explaining what the connections mean once they have been discovered. In this case, we may be able to guide the exploratory parameters to reveal the underlying patterns. If we can get a random forest approach to reproduce the patterns, it would boost the model's credibility, as it's likely a less biased criteria for establishing a causal correlation.

AGU meeting session on Machine Learning https://youtu.be/xH3tbCOd_oQ