Software “detects CEO emotions, predicts financial performance”

Perhaps they could also incorporate data on CEO horoscopes, and which coffee outlets they frequent.

Nate Silvers wrote an entire book on this subject called "The Signal and the Noise" [1]. Humans are so often taken in by people claiming to be able to make predictions by combining new data points. The more unusual, or unrelated to the subject matter, the better. They make good headlines but (not surpsingly) almost always turn out to be heavily flawed in practice.

You can basically measure how much a pundit/expert is going to be wrong in their predictions by how ideological they are in their analysis. The best indicator is when they use only one or two metrics as a basis of a prediction of an otherwise very complex scenario.

One example from the book is how a researcher became famous before the 2000 US presidential elections by claiming to predict races with 90% accuracy [2]. He claimed that by measuring a) per-capita disposable income combined with b) # of military causalities you can determine whether democrat or republicans get elected. He said historical data backs up his theory. He then proceeded to fail to predict that years election and faded into obscurity.

Nate did his own historical analysis and demonstrated it was only 60% accurate instead of 90%. Plus that was only if you ignore 3rd party candidates as the model assumes a two-party system.

Plenty of other examples are provided in the book which makes me highly suspicious of the value of the predictions made in this article.

The general idea is that we need to stop looking for simple one-off solutions to complex problems. Instead we should adopt multi-factor approaches which suffer from fewer biases and are better grounded in reality. Otherwise these predictions are just another form of anti-intellectualism.

[1] http://www.amazon.com/Signal-Noise-Many-Predictions-Fail--bu...

[2] the "Bread and Peace" model by Douglas Hibbs of the University of Gothenberg http://query.nytimes.com/gst/fullpage.html?res=9803E5DD1F3DF...

Which is actually why you shouldn't invest based on this algorithm if you care about the research. I think, ideally, you should:

1) Generate a few hypothesis algorithms, including one that invests at random.

2) Publish a cryptographic commitment for each algorithm.

3) Never actually invest any money. Alternatively: let someone else invest your money for you, without knowledge of your hypotheses.

4) Run your algorithms privately, without updating them at all. Capture the data the algorithms use (including random choices taken).

5) 5, 10 or 20 years later, publish all your algorithms, the data they had as input and their results, see if any of them would have predicted the actual performance of the market in an statistically meaningful way.

I imagine the main reason most researchers are unlikely to do that is the 5-20 years project requirement. It is a lot easier and faster to just take historical data from the market and then produce algorithms that would have predicted performance after year X, based on information before year X. Of course, the problem is that you run into over-fitting and survivor bias (in that only positive results are generally published).

Btw, having your algorithm be run by a fund and having that fund succeed, then publishing the algorithm, would also be susceptible to survivor bias.

Yes and no - AIUI they're "dogs" based on having limited future growth potential rather than having had a bad quarter.