David Tweed wrote:

> Hi, just as a general note I'm slowly incrementally adding code towards doing linear/bilinear regression [here](https://github.com/davidtweed/multicoreBilinearRegression)...

Thanks, great! But I don't really know what you're doing. I guess it's something new or nonstandard, because I'm guessing should be some software you can just take "off the shelf" that does _approximately_ what I'm asking for... right?

Are you trying to implement some of the ideas [discussed in your blog article](http://www.azimuthproject.org/azimuth/show/Blog+-+Exploring+regression+on+the+El+Ni%26ntilde%3Bo+data)?

I'm hoping someone will help me use some "off the shelf" software to make an initial attack on the El Niño prediction problem, just to get some sense of how hard it is. This could be someone less trained in machine learning than you but more friendly with computers than me.

> In terms of things to predict, the only real thoughts I’ve had so far is that I’m reluctant to try to predict a 3-month average based El Nino 3.4, purely because it’s likely to be noisy and hence errors aren’t necessarily indicative of errors on the bigger problem. I’m inclined to try something like the (1+5+1)=7 month average El Nino 3.4 index for the following period immediately after the observations used for prediction, but that’s not really more than a rough guess.

I guess you're saying this because an El Niño occurs when the 3-month running average of the Nino 3.4 index is over 0.5 °C for at least 5 months in a row, which involves 1+5+1 months of data?

Of course the 3-month running average being above some value for 5 months in a row is a bit different than the 7-month average being over that value for one month. The former is 5 inequalities while the latter is one.

To some extent we have a choice between predicting things that are easy to predict and predicting things people care about. People care about when there's an official El Niño. But they also care especially when it's a "strong" El Niño. It's hard to predict an El Niño more than 6 months in advance. But for that very reason, this is what people most want to do.

> A quick note about prediction errors: it may be worth considering things other than squared-error as predictors.

I have no strong ideology about this sort of thing, so it will be good if you provide us with one.