[kurumo]

I work in this field. That is not the first paper written about this system; the previous was "Sentiment Analysis of Financial News Articles", Schumaker et. al. This is not the only academic project describing something like this, see papers by Mittermayer and Lavrenko for other examples. Short summary of the results is basically this: you would have to be insane to trade purely on output of a system like this. There are multiple issues with this, the primary being that most of these systems do not and cannot distinguish documents leading the trend from those lagging the trend. Furthermore, to actually be able to trade on the indicators derived from news you need to encode the market expectation into your model, which is a separate (and difficult) problem. The margins these systems produce in simulated trading are small, and for some reason nobody takes transaction costs into account. This is not to say that systems like this do not exist in the real world - they certainly do, but the people that build and/or use them generally will not discuss the details, for obvious reasons.