Sentiment Signal And Systematic Financial Applications

Capturing sentiment from Twitter and systematically trading on the derived sentiment signal is all the rage in financial circles these days. At least in theory. But how many buy-side organisations have actually implemented a solution to capture this sentiment systematically and found sustainable returns? And what of the potential for abuse?

The answer to the first question is probably very few, though we may never know given the buy side’s penchant for secrecy. And those using such signals would have to do so cautiously and not in isolation. The answer to the second question remains open.

One of the most visible case studies of the buy side using Twitter sentiment is that of Derwent Capital. Founded in 2008 by Paul Hawtin, it went on to launch an absolute return fund two years’ ago, in February 2011, that claimed to use sentiment mined from Twitter to drive investment decisions. In early 2012, the firm liquidated the hedge fund and launched a trading platform with social media analysis built in. In this platform, the social tools are just that – not the core value proposition – which was pretty much the same as any other retail trading platform. In February 2013, it put the platform up for sale on the Internet. The last bid (Feb 12 2013) was just GBP100,000… Not much for a GBP350,000’ investment. (Incidentally, a MetaTrader license – the world’s largest social trading platform – goes for around USD100,000.) Maybe it’s like e-Bay and all the bids will come rushing in at the end?

The foundation for Derwent Capital’s fund and platform lay in academic research from October 2010 by Johan Bollen, Huina Mao and Xiao-Jun Zeng entitled “Twitter mood predicts the stock market”. The paper established the connection between emotion-related words appearing in Twitter posts and subsequent movements in the Dow Jones Industrial Average.

Since then there has been a plethora of academic research hypothesising on how to extract sentiment from Twitter, and other social media, for use in financial services. And there are at least 20 start-up firms claiming to have the answer as to how it should be done – the sentiment extraction, that is. There are no visible claims to successful systematic trading implementation.

Anecdotally there must be some value in Twitter as an aggregate measure of sentiment trend. Counting the frequency an entity or event is mentioned undoubtedly has value as a proxy for price volatility. Indeed, this works from normal news sources. And there is probably value in aggregating sentiment from positive or negative words in Tweets. But only in aggregate – not as a signal for event-based trading.

There are, however, several well-known firms that extract event-level sentiment from real-time news feeds and other official publications. Their clients use this data in systematic trading and investment applications. As a sales guy in RavenPack, one of the firms in the business of news analytics, I am often asked whether we will analyse Twitter. The answer is no, but it’s a qualified no. The main reason is the number of false positives that would be generated. Clients do things like systematically trade currencies after detecting breaking news of a serious earthquake. Or trade stocks after detecting a rumour of a profit warning in the news. If Twitter were processed in this way, how would we tell the difference, for example, between a tweet about a real earthquake in California and a mother tweeting that her house looked like it had been hit by an earthquake because her boys had been on the loose inside?

There are probably four methods to extract event signals from Twitter. First, only content from trusted sources – or sources of authority (where there’s some degree of editorial control) on a particular instrument might be used. But we may as well then subscribe straight to the source because they are likely to be financial publications. Using Tweets from individuals, even if you could overcome the challenge of maintaining some sort of ‘clout’ metric per user per instrument, would be dangerous.

We might rely, secondly, on hashtags and instrument symbols. But the tweeters are only human and, by definition, prone to inconsistency and error.

Third, we could rely on our ability to analyse the language around the event and entity to ensure it’s accurate. It’s a massive challenge because the sentence construction (or lack of) used by tweeters makes it near impossible to ensure accuracy.

Finally, we might use bursts of activity on Twitter as an indicator that an event has occurred or is about to. But even then we have no way of systematically knowing the burst is down to a confirmed event.

Net result, the accuracy levels our clients expect on detecting events, entities and the direction of sentiment just would not be there with the technology available today. We only need to be wrong two times out of ten to destroy the value proposition for our clients. Hence we rely on journalists to monitor Twitter for us, conduct some good old verification of a story, and publish it on an official news source.

But let me pose a different question. What’s to stop Twitter becoming a favoured tool for market manipulation and spreading misinformation? Most financial services firms have rules in place to try to control their employees from doing such things. But there’s nothing to stop them using an anonymous Twitter handle tied to a trumped up e-mail address from their smartphone or home PC. Pity the compliance department and the regulators who have to follow that trail!

And how about private individuals? Can they be prosecuted for market manipulation or spreading misinformation? Of course they can. But on Twitter? I guess the answer depends on the country and how much effort firms would put in to tracing the source. In the UK, you can get prosecuted for libel against another individual posted on Twitter. Why shouldn’t companies be afforded the same protection? Imagine that, someone other than a banker caught with their fingers in the pie.

The use of sentiment derived from Twitter in financial applications is, no doubt, here to stay. But the technology used to extract that sentiment is still short of being able to help in anything other than decision support. And then there are the legal and compliance issues which haven’t really been addressed yet.

Hugh Taggart is Head of Sales & Business Development at RavenPack, a leading news analytics firm with offices in New York, London and Spain. www.ravenpack.com