Getting sentiment-al about algo trading

Marc Adler posted an interesting question on his Magmasystems blog asking how CEP and/or sentiment engines might have determined if the United Airlines bankruptcy story was old and could have avoided the actions taken by algo trading engines that caused the stock to take a hit.

The problem with considering sentiment alone is that sentiment algorithms perform a discreet function. They read a story contextually, then run a set of sophisticated algorithms to determine if a story is positive or negative along with determining other thematic attributes. Any trading strategy that is pulling the trigger on trades would also have to factor many non-sentiment attributes such as levels of trust and confidence. Trust might be high if the source of the story was verified as a company PR event on PR Newswire. Confidence may be high if the date was verified as today, but low if it is missing or old. If sentiment, trust and confidence (probably among many other factors) align, then the algo might signal a trade.

This is a nice way of saying that any trading algorithm that makes buy/sell decisions based on an undated story that a Google crawler ‘found’ at a small paper in Florida kind of deserves what it gets. I personally am not convinced that news-based trading was the real problem. It’s more likely that enough people fell for the old news that it triggered stop loss selling, level 2 imbalance selling and price algo selling.

To me the question that is equally interesting to ask is how this opens the door to fraud and manipulation. How easy would it be to trap Google’s crawlers into recreating this situation with other solidly negative old news? Or what about a perfectly legal algo that figures out that this is a bogus event and gets into United stock long at $3 in the seconds before trading is halted?