A second order predictor may benefit from it, giving less gains more reliably guessing algorithm behavior instead of stock trends.
(University of Amsterdam)
With that said, I had between 60-75% success rate on predicting the direction, depending on how I set different things like requiring multiple sources, confidence required to execute prediction, NLP tweaking, etc. The tiers was worse, but I think it was around 40-45% on the last iteration. Obviously, I would want to get better at predicting magnitude before I used it in any real life trading. It also was highly dependent on the company/industry. It performed better on something like Apple or Google, but worse on some smaller companies in industries that are not near as news heavy.
The main goal of the project was just to satisfy my curiosity with algo trading, web scraping and data analysis while giving me another cool project to talk about in interviews, as I'm a college student. I used scrapy, NLTK, pandas mostly.
Shorting has no true floor to the amount of money you can lose. This isn't true of going long, as an instrument can only go to zero. There is a natural imbalance between the max risk associated with shorting vs going long (and there are lots of other risks associated with trading, this is very simplistic) that means that no one who has done this sort of thing before uses a predictive ability without risk as a serious basis for a trade.
In fact, predictive ability is not that special. There are trivial algorithms that can predict movements with high levels of accuracy (much higher than 60% for instance), but don't perform well on a risk adjusted basis in modelling or in reality.