As Twitter's initial public offering approaches, Wall Street analysts are building models to calculate what it might be worth – with a figure of up to $15 billion doing the rounds.
But while the IPO is sure to be one of the biggest financial media events of the year, traders are starting to look at Twitter not as a buy or sell in itself but as a way to generate hot tips on other stocks.
The idea is a simple one: the 500 million daily messages on Twitter put online the sort of gossip usually only available in snippets by eavesdropping in bars or the office elevator. Apply some moderately sophisticated computer filtering, and out should pop market-moving news and views investors can use.
(Related video: Twitter's new connection to television)
Twitter has already demonstrated its potential to move markets, with the "hash crash" – named for the "#" symbols used on the micro-blogging service – knocking 145 points off the Dow Jones Industrial Average in April on the back of a false tweet from a hacked wire service account.
This week brought another example, with some traders blaming a misunderstood Israeli military tweet referring to bombing Syria in the 1973 war for a $1 per barrel drop in the oil price on Thursday. There are plenty more examples of tweets offering advance notice of news which takes minutes, hours or days to reach the mainstream media and move prices.
The problem is how to sift the information from the tweets about Justin Bieber, Halloween and the latest online games.
"You have to be happy with a lot of noise in your data," says Richard Peterson, managing director of Los Angeles-based MarketPsych, which uses computers to sift social media and news feeds for market signals. "You get a lot of alphanumeric gibberish if your machines don't know what to look for. The more you get into it the more you realize that it's very complex."
The latest attempt to prove that Twitter contains useful information comes from a PhD student at University College London, Ilya Zheludev, and academics Robert Smith and Tomaso Aste. Their work, undergoing peer review, shows that tweets can contain useful information – but that it is far from guaranteed.
"The proponents of this idea really do exaggerate it," Mr Zheludev says. "If it really was as possible as they think it is, first of all they wouldn't publicize it and second they would be rolling in cash."
"I'm not saying there's nothing here, but I'm not saying you can print money either."
He analyzed tweets in real time using a standard psychological dictionary to give them a positive or negative sentiment score, filtering only those containing specific company names or the $-tags used to identify stock symbols. They also looked at the S&P 500 and the FTSE 100, without filtering the inputs.
For most stocks there were not enough tweets to generate statistically significant volumes, while for the FTSE and many large companies – including Google, Intel and Bank of America – the sentiment had a stronger link to past price moves than future ones.
But 11 of the 50 financial instruments tested showed a statistically significant link, which Mr Zheludev said was not merely "data-mining" for past correlations as the system had been set up then run in real time.
"There's enough there that could potentially be useful," he says. "But we're still at the very beginning of what's possible. I wouldn't personally be happy trading from this data but I would be happy using it to contribute to a wider model."
There are already several companies analysing Twitter data and selling it to hedge funds and other traders desperate to find an edge in the markets, in what is proving to be a lucrative business.
(Related video: Twitter's value? Depends who you ask)
Joe Gits, founder of Chicago-based Social Media Analytics, says he expects a turnover of $60m in two years from his system. SMA measures how far sentiment is from the norm for 8,000 US equities, from tweets by 400,000 people identified as financial professionals, active traders or market influencers.
"We're not listening to the average Joe, we're following the guy who's been trading for 20 years," he says. Customers are mostly computer-driven hedge funds using the analysis as an extra input to their models, or active traders looking for ideas on which stocks interest will soon be focused. Mr Gits says that on average social media leads the news wires by 12 minutes, but that the indicators work better for sectors than they do for the overall market.
He has big expectations. "I think it's going to become every bit as big as earnings estimates in the not-too-distant future," he says.
Previous attempts to trade purely on the back of social media data have come unstuck. Mr Peterson ran a hedge fund based on signals from social media, but gave up after finding it did not work for the biotech sector. In London Derwent Capital set up the first Twitter hedge fund in 2011, but quietly closed down again just a month later.
Some of the big quantitative hedge funds which have examined the idea of using tweets and social media have given up, or simply maintained a watching brief. "We've looked into this in great details and the noise just becomes so overwhelming that you can't really predict very much," says one manager of a large fund.
Mr Zheludev's approach is rather more complex than the simple algorithms used by some early traders to extract signals from the online noise. But even the most advanced language processing has trouble with much of Twitter: they cannot process sarcasm.