Waters Technology 10/28/11
Author: James Rundle

The impact of social media platforms such as Twitter and Facebook on personal, commercial and political levels is well established, but what about their implications for financial services? Can the sheer volume of sentiment data generated by these networks translate into profits for canny investors? Some hedge funds, technology vendors and academics seem to think so, but others aren’t convinced.

In 1942, Isaac Asimov penned the first in a series of short stories that would together constitute his magnum opus, Foundation. A central mechanism in the book is psychohistory—the idea that while individual predictions of future behavior are impossible, statistical laws as applied to large groups of people can indicate the direction of future events. Asimov’s novel is science fiction, but academic research in recent months has been using measurements of mass common sentiment expressed through social media platforms, with results that can potentially predict the movements of the stock market.

Computational scientists Johan Bollen and Huina Mao, from Indiana University Bloomington, and Xiao-Jun Zeng from the University of Manchester, published the key academic paper in this debate near the end of 2010. In the paper, Twitter Mood Predicts the Stock Market, the authors claim that by analyzing the messages—which can be up to 140 characters in length and are known as tweets—posted by users on the microblogging site Twitter, and categorizing them into different “moods,” they were able to use specific sections of this data to predict the daily up and down changes in the closing value of the Dow Jones Industrial Average (DJIA) with a time lag of several days. Their accuracy was measured at a startling 87.6 percent, with certain factors added in. Other studies have also used sentiment analysis on Twitter, Facebook and other social networks to predict subjects as diverse as movie box-office returns, the rise in stock prices related to specific companies such as Coca Cola and Starbucks, and more.

The Twitter Fund
It wasn’t long before the financial services industry took note of the paper. In early 2011, Derwent Capital Markets started the Absolute Return Fund, with an initial capital value of £25 million ($40 million), to trade on the researchers’ work. After its first month of operation, it returned a 1.85 percent profit to investors, beating the S&P 500, which fell by 2.2 percent that month.

Following the research, Bollen set up Guidewave Consulting, which has licensed patent and software rights to Indiana University Research and Technology Corp., and inked a consulting deal with Derwent. Paul Hawtin, the fund’s co-founder, declines to comment, citing legal reasons pertaining to hedge fund marketing and the amount of publicity it has already received. In a statement regarding Derwent’s partnership with Guidewave, however, Hawtin says, “Investors accept that financial markets are driven by greed and fear, so it’s hugely valuable to monitor and understand global sentiment in real time.”

With the performance of this enterprise and the growing body of academic research in mind, others have taken note. Centigage is a company that offers social media analytics with the express purpose of informing investment decisions, using what it calls a human-curated algorithmic approach to mine social media platforms such as Twitter for indications as to the movements of stock prices.

“We envision a world where every trading website has a social sentiment indicator, or every financial institution or professional is looking at sentiment from a social media outlet to see what’s going on,” says Robert Logan, co-founder at Centigage. “Currently, there’s a small group of financial professionals looking at social media. It is hard to say that it is a prominent indicator, but that’s the way it is headed. The data itself, if we’re talking specifically about Twitter, isn’t old enough for us to run back tests on decades’ worth. But it’s such a vast and vibrant amount of information that it seems foolish to pass up trying to decipher some sort of message from what’s being said.”

The amount of raw sentiment data being projected on a second-by-second basis is the key attraction for these kinds of providers. But social media also has a reputation for being noisy, in that much of the content produced has no specificity to it, some of it is sarcastic and can be misread, and a lot of it is personal and not relevant to stock markets, corporate actions, political events or other areas that can impact share prices. For Logan, however, this isn’t a deal-breaker. “It’s important to understand that we’re looking for a general sentiment of all social media users. Even if something appears to be not relevant—specific statements including a keyword that we are focusing on may not indicate an obvious emotion—we find that in mining these 200 million or more Twitter users, you’re getting a good indicator as to where the market will be in the next 12 to 24 hours.”

Logan says what allows the company to sort through this broad sentiment is experience in the financial markets. Centigage is a small startup founded by, as he puts it, two investment professionals in different areas of the financial world who thought social media was the new frontier. The human-curated approach with that background allows them to differentiate the data in a more tailored manner than a pure algorithmic method of data mining would, although he acknowledges the impact that Bollen and his scientists’ work has had.

Critical Mass
Not everyone agrees that social media has come of age, however. John Coulter, president and CEO at Titan Trading Analytics, says it still has some way to go before it can be a truly useful source for informing trading decisions. “We see it as an overlaying data element at this point. The basis of what we do is very quantitative on price and volatility,” he says. “The Twitter data and the sentiment that goes along with social media in different aspects ranging from blogs to chat forums doesn’t really have enough critical mass in our opinion to be able to generate alpha strictly from that basis alone. So we’re using it as an indicator on top of our quantitative signals, just to give the trader a heads-up to the pulse on the market, so they can make their own decisions. This isn’t something that they auto-trade on by any means. What Derwent has done, and what others might be attempting to do, we just don’t see the critical mass of data being there at this point. It may be there some day, and it probably will be, but not at this particular moment in time.”

Rich Brown, global business manager, machine-readable news at Thomson Reuters, says the difference in approaches to sourcing information from, say, news from tier-one or tier-two media, and the wider view from social media makes it an additional rather than a prime source of investment indication. Indeed, the vendor will be adding social media to its machine-readable news business some time this month.

“It’s more common to use it as a complementary factor than the sole factor. The Twitter fund itself, I believe, uses it as a sole factor rather than someone who’s using a statistical arbitrage strategy, with five or 10 other primary factors—pricing, volume, upgrades and downgrades, comparable company analysis, pure analysis, and so on. When you add news to that, it’s an orthogonal source of alpha, which means it’s differentiated, and it’s bringing something else to the party,” he says. “When you start to look at social media, you’re adopting a crowd-sourcing approach, because you’re not going to buy every time there’s a good tweet on something, or a good blog post, because when you start to bring in the fire hose of the internet, you’re burning through transaction costs when you’re buying and selling all the time without changing the fundamental reason why you’re buying and selling. In social media, you have a lot more consumer-type postings, which, when taken in aggregate, like the Twitter fund does, you use to predict general market movement as opposed to microstructure-type movements, like single stocks with buy and sell signals.”

Titan factors in what it calls the “emotional” element for its analytics, for which it partners with other vendors. In addition to what Coulter describes as a lack of critical mass, he also sees another potential hazard for those using social media platforms alone to inform buy or sell signals. “When you look at these things, there’s a high probability with the amount of data right now that a lot of it can be gamed,” he says. “Because there isn’t critical mass yet, you could create programs to put out false indicators, especially if it’s not a very highly-traded stock. I think a good analogy is if you look back into the Web 1.0 days, when people were gaming the notion of banner ads from search engines. People were creating these automated programs that were creating click fraud, and they were inflating or manipulating results. That same sort of gaming is probably going on right now. I think a firm relying solely on Twitter data, unless it has figured out ways to do this where others haven’t yet, will have to be aware of a heavy amount of gaming taking place.”

Behavioral Economics
It’s not just the social media sentiment that plays into new developments in capital markets, but the interconnected nature of the technology itself. Various online shops now offer foreign exchange (FX) trading on a social media basis—some where participants “follow” a star trader’s strategy for returns, others where the market works on a collaborative basis. Returns from this can be mixed, and every company carries stark warnings about the dangers of trading without a full understanding of the markets. The primary idea behind the recent popularity in sentiment analysis and behavioral finance, expressed through Twitter, is not new. The Elliot Wave principle has long argued that psychology and a host of other factors can play a part in the movements of stock prices, its key advocate today being the Socionomics Institute.

Socionomics is the study of how social behavior and economics interact. The Institute sees these developments and the nascent market in social media sentiment analysis as being in line with its own research. “Research into the predictive possibilities of social media is in the early stages,” say Alan Hall and Matt Lampert, analysts at the Socionomics Insitute. “So far, most of the models that we’ve seen aim to forecast only a few days at a time, but researchers are working on models that may one day generate longer-term forecasts, specific indicators and other metrics.”

Investors can leverage the cache of raw information provided by Twitter, but Lampert and Hall say it is important to have a proper mechanism in place to guide the analysis of the data, rather than searching ambiguously for whispers in the machine. “With the popularity of social media, we now have a large-scale, high-frequency data repository that researchers may be able to use to develop new metrics of social mood. Quantifying and tracking the changes in mood can give analysts a leg up on forecasting the trajectory of financial markets,” they say. “But every data source comes with its limitations: For example, Twitter posts can be distorted by emotion and short-term reactions to events, such as what happened upon Michael Jackson’s death. It is also important to have a theoretical perspective to guide your interpretation of the data. From a socionomic perspective, people’s tweets don’t cause the stock market to go up and down. Rather, we propose that social mood is an underlying factor that influences financial market valuation and the character of social sentiment that shows up in Twitter posts. The underlying variable social mood is what produces the correlation between the two data series.”

Infancy
The technology to exploit social media platforms in financial services is still in its infancy, and the methods of harnessing the data they generate have barely been conceived. The confluence of consumer technology and high financial analysis is intriguing and raises further questions. If the practice of using social media sentiment as a data stream for investment decisions continues and develops, does it in turn become inherently unreliable?

A key tenet of Asimov’s psychohistory was that the basis of the prediction and analysis ceases to become relevant when the subjects realize that its behavior is being monitored in this fashion. If Wall Street begins to take a high-profile accounting of what is being said online, will that too lead to gaming, deliberate misinformation and other areas of high variance that render the data set unusable? Whatever the answer, social media analysis, and factoring in our collective behavior, is something that seems destined to stay and grow.