27 January 2011

Fun with Artificial Intelligence, Part II

As promised when I initially wrote about artificial intelligence (and specifically about Eureqa), I have been playing around with economic data and have come up with some interesting results. After running dozens of experiments using different combinations of data and allowing Eureqa to use different computational operations, I have come to a fascinating conclusion: in terms of what helps the AI fit equity performance data points to a model based on all data available to it, it seems equities don't 'care' about anything but how we, the consumers, think about the economy, regardless of how well- or ill-informed we are about the economy. Give the AI GDP data, CPI, unemployment, CD rates (to provide opportunity cost), even prior year stock performance and price-to-earnings data: if you also provide consumer confidence data, it will systematically discard every other data type save that one. In some simulations with solutions that have similar complexity and fit, it might also use Fed Funds rate and/or CD rates, which I see as opportunity-cost stand-ins. But for the most part, it just wants to know, "How do you feel about the economy, regardless of how well or poorly it is actually performing and regardless of how much you even really understand it?"

As I said in the prior post, I am far from convinced that such data-mining can offer realistic predictive models. But ironically, the fact that the AI prefers the least rigid, least 'rational' data type, in fact makes me less skeptical about its predictive power. My reasoning is this: if the AI chose models that were based strictly on 'hard' data such as CPI, etc., I would suspect it was simply data-mining and that the results would be useless outside the confines of the already-given universe of data points, since equities markets are inherently irrational and are driven by things far less quantifiable than, say, GDP; but the very fact that it chooses the least 'rational' data type, consumer confidence, tells me that it may indeed be coming up with a decent predictive model, since it seems fitting that it has chosen the one data type that combines rigid metrics but applies them to decidedly 'soft' data (that is, consumer sentiment).

So the proof will be in the pudding, I guess. But that pudding will take quite a while to cook, so don't hold your breath. Meanwhile, for what it's worth, I will shortly add some of the predicted values from various experiments. Then we'll sit back and see what happens!

My next project is GDP. I want to see what data types the AI most prefers for predicting the performance of the US economy.

No comments:

Post a Comment