This content is from:Portfolio

Can Twitter Predict Stock Market Moves?

London-based Derwent Capital Markets' Cayman Island-domiciled Derwent Absolute Return Fund began trading last week. The fund appears to be the first to make bets based solely on sentiment analysis derived from Twitter.

What happens when a 28-year-old British currency trader who previously sold a hair care business teams up with an international trio of informatics professors who publish an academic paper that purportedly unlocks the potential for Twitter to predict stock market moves?

They launch Europe's first social media-based hedge fund, of course!

That, in a nutshell, is the improbable story behind London-based Derwent Capital Markets, whose Cayman Island-domiciled Derwent Absolute Return Fund began trading last week. The fund appears to be the first to make bets based solely on sentiment analysis derived from the microblogging platform, which celebrated its fifth anniversary in March and, by one recent count, now produces more than 140 million messages, or Tweets, on average, every day.

Investors are enamored: The fund’s launch, which was originally slated for February, was delayed twice in part because Derwent attracted far greater interest than its founders had anticipated. In April, the entrepreneur and trader who is Derwent’s founder and fund manager told IR Web Report that he had raised $40 million and that Derwent’s “master list of interested parties” stood at $95 million.

“For years investors have widely accepted that financial markets are driven by fear and greed,” Hawtin said in a statement announcing the fund’s launch. “But we’ve never before had the technology or data to be able to quantify human emotion. This is the 4th dimension.”

Derwent’s founders aren’t the first to extol the predictive powers of Twitter, of course. The microblogging site has already been used to forecast everything from movie box office sales to elections in the U.S. and the UK. And algorithmic traders have been experimenting with Twitter sentiment analysis for several years as one of many potential clues that might help inform their fast-moving trading strategies. Still, Derwent Capital’s experiment represents the leading edge of a flurry of new academic studies that stand out for their bold assertions on just how profitable trading strategies gleaned from Twitter’s fire hose can actually be. The firm’s target: annual returns of 15 to 20 percent.

Derwent’s secret formula originates in academia. In a widely read paper, Indiana University's Johan Bollen, who is advising the fund, along with Indiana colleague Huina Mao and the University of Manchester’s Xiao-Jun Zeng, used two different mood tracking tools to analyze the text content of nearly 10 million Tweets. The first tool, OpinionFinder, is an open-source software package hosted by the University of Pittsburgh. The second is an algorithm developed by the authors that is based on the “Profile of Mood States,” a methodology used by psychologists to monitor the effects of treatment changes or the impact of drug regimens on patients’ mood states. The authors expanded the 72 mood descriptors in the standard POMS questionnaire to a universe of 964 terms—a far richer mosaic of the range of human emotion, they say—by tapping Google to analyze “word co-occurrences” in 2.5 billion sequences of terms scanned by the search engine on publicly accessible web pages. The researchers’ expanded lexicon was then mapped to six mood dimensions: Calm, Alert, Sure, Vital, Kind, and Happy.

Their key finding: Only the Google-enabled POMS measure of "Calm” has predictive value. But astonishingly, this mood state alone predicts daily up and down changes in the closing value of the Dow Jones Industrial Average three to four days later with 87.6 percent accuracy.

Not everyone is sold. “Do people Tweeting have the money that drives the sentiment? Probably not,” says Wesley Gray, assistant professor of finance at Drexel University and co-founder of quantitative hedge fund Empirical Finance, which trades in part based on the flow of information through private social networks used by professional investors. Gray questions whether the relationship between measures of calm and future market returns makes intuitive sense and, in a recent blog post, concluded that “using mood to predict stock prices doesn’t pass the sniff test.”

A second research paper, published in April, focuses its analysis on individual securities. Written by two academics at the Technical University of Munich, Timm Sprenger and Isabell Welpe, the paper analyzes stock sentiment in Tweets that include ticker symbols marked with dollar signs—such as $AAPL for Apple Computer Inc. or $GOOG for Google Inc.— the protocol that Twitter users employ to identify messages about specific names. The study borrows yet another sentiment methodology, this time from computational linguistics, to sort nearly 250,000 Tweets about companies in the S&P 100 index according to their buy, hold, or sell signals.

Like the academics behind Derwent Capital’s models, the authors uncover a link between Twitter sentiment and stock prices. One finding, in particular, stands out: that investors could have earned as much as a 15 percent return over a six-month period by going long and short the top and bottom three stocks ranked according to the strength of their Twitter buy sentiment. The trading signals from Tweets are both concentrated and short-lived, however: the optimal holding period is just one day, the authors found, and trading more than just the top and bottom three stocks seemed to hurt returns. In April, the authors launched a website based on their methodology, called Tweettrader.net, which aspires to be “the most innovative financial information market for stock microblogs” (tagline: Making $ense of it All).

这种情绪指标可能来源于thousands of investors Tweeting about specific securities seems plausible. But this study, like its predecessor, is challenged on several fronts. As traders have pointed out, Twitter sentiment algorithms—like sentiment readings generally—are unlikely to predict the unforeseen events that seem to plague markets with alarming frequency, a fact that Bollen and his colleagues explicitly acknowledge in their paper. And developing models that accurately process the nuance and context of language contained in a Tweet is extraordinarily tricky. “It’s difficult to get a ‘street-wise’ dictionary that doesn’t have industry or firm-specific problems,” says Tim Loughran, a University of Notre Dame finance professor who cites, by way of example, the terms “vice” and “crude.” These words typically have negative connotations but, in financial contexts, are more likely to contain no value judgments at all: think “vice president” or “crude oil.”

In a Journal of Finance article published in February, Loughran and a co-author, fellow Notre Dame finance professor Bill McDonald, demonstrated this complexity. The authors examined a large sample of 10-K filings from the mid-to late-1990s and found that nearly three-quarters of the words identified as negative by the Harvard Psychosociological Dictionary, a widely used textual classification system in other disciplines, were in fact terms that are not typically negative in the context of financial reporting.

When it comes to diligent, ethically sound investing there are rarely any shortcuts. Twitter’s most important contribution to financial analysis may prove to be a lesson of a different sort: that forserious investment professionals, there is simply no substitute for using reasonable care, exercising independent professional judgment, and having a reasonable and adequate basis, supported by appropriate research and investigation, when making investment decisions.

Len Costa is director of innovation and emerging media at CFA Institute.