It looks like you're new here. If you want to get involved, click one of these buttons!
I will post a blog article about what I learned at NIPS when the videos of talks appeared. The best ideas I got were these:
1) To what extent have people systematically checked that Niño 3.4 is the best quantity for predicting other El Niño-related quantities, e.g. ones that actually matter to farmers and other people who need weather forecasts? There are a number of El Niño indices, but maybe we could seek "optimal" ones.
2) I should attend the workshop on Climate Informatics during September 25-26 in Boulder, Colorado.
3) In a 2013 paper in J. Climate, Jia DelSole and Tippett apparently discovered a low intrinsic dimensionality for the behavior of world climate models, due to El Niño and a few other features! Looking for a low-dimensional attractor in a high-dimensional phase space sounds fun.
In a similar vein, Kawale et al, SDM 2011 and Steinback et al, KDD 2003 did automated discovery of pressure dipoles like ENSO, NAO, etc. Also related to this is Ebert-Uphoff et al's paper A new type of climate network based on probabilistic graphical models: results of boreal winter versus summer J. Clim. 2012, which used Bayesian network ideas to infer causal relationships between the 4 biggest teleconnections.
Maybe there's a way to combine these ideas and start from climate data and build up a simplified but interesting model of the world's main variable climate elements.
4) The idea of rating the main existing El Niño forecasts, and/or developing a better one, still sounds very interesting to me. We've developed a certain amount of expertise in this subject and we haven't fully exploited it yet. For example: maybe we could try to predict Niño 3.4 using the entire matrix of link strengths, not just the average link strength. How much better is this?
5) I saw a rather shocking graph of world temperatures as predicted/retrodicted by the main climate models used by the IPCC. What was shocking is that these models stay close together in the best but then spread out wildly in the future. This suggests something like overfitting - the models seem to be fitting the known data "too well" compared to how much they agree about the future. Is there a way to quantify how bad this "overfitting effect" is, or prove that something bad is going on? This could have a lot of impact if it could be done well.
Needless to say, we won't have the energy to do all 5 of these things! I would love it if people could talk strategically about what we want to do. People tend to be more interested in discussing their latest favorite idea, but I'd like this thread to be about sketching a plan. It's possible that doing more of what we've just been doing is the best plan... but I plan to take a break from the energetic work I've been doing, and think a bit about what's good to do next.