The predictive ability of media and particularly social media for financial market developments is receiving increasing attention, as the first funds are entering the novel territory of investing on the basis of sentiment analyses of Twitter, Facebook and traditional news media content.
Sentiment trading relies on the premise that stock prices, in addition to factors like fundamental company data, such as earnings, earnings forecasts, dividends or share price history, are also influenced by investor attitudes towards a specific company, its products and stock.
According to this view, investor attitudes are influenced to a large extent by news media and social media, as well as personal experience and information obtained in personal conversations from trusted sources.
If this is indeed the case, the correct analysis of company news and social media sentiment should make it possible to predict investor attitudes and behaviour and to design a corresponding investment strategy that exploits this predictive ability.
An explosion in the use of internet search engines, social networking sites and digital communication has resulted in the creation of an enormous amount of data each day, which should yield valuable information on customer, consumer and investor attitudes and behaviour. The question is how to mine this “big data” as it called.
This is where sentiment trading, as a subsection of algorithmic trading comes in.
Company news and share price effects
The traditional argument against news events having a measurable effect on share and asset prices was that news stories were published with some delay after the events they reported, eliminating any impact they could have on the stock market. This delay, however, shortened when traditional news media began moving online and started to publish information closer to real time.
In addition, certain media have the perceived ability to move share prices directly. In 2006, Eugene Plotkin, a research analyst in the fixed income division of Goldman Sachs, and David Pajcin, a former employee of Goldman Sachs, were charged by the Securities and Exchange Commission with several counts of insider trading. One of their alleged schemes involved infiltrating a printing plant used by Business Week to obtain the information contained in the widely read IWS column of the magazine prior to publication.
The SEC complaint alleged that the column generally impacts the price of the securities mentioned in it and that Plotkin and Pajcin had either acted on the insider information or passed it on to other traders, resulting in over $345,000 of illicit trading profits.
Since then academic papers have extended this news effect to special interest websites such as Seeking Alpha. A 2011 paper by Hailiang Chen, Prabuddha De, Yu (Jeffrey) Hu and Byoung-Hyoun Hwang notes the website associates strongly with contemporaneous and subsequent stock returns, and helps predict earnings surprises. The authors also find that the “social media effect is stronger for articles that receive more attention and for companies held mostly by retail investors, the primary generators and consumers of social-media content”.
Reuters trading dashboard
Buoyed by recent advances in computational linguistics and natural language processing, which have improved the automated analysis of electronic text and the ability to assign favourable or unfavourable ratings to the content, many news analysis tools have emerged.
Thomson Reuters’ financial trading dashboard incorporates text and sentiment analysis software from Lexanalytics. The Lexanalytics system can sift through texts from multiple sources, filter it for keywords, tone, freshness and relevance, and analyse both the meaning of the text in a textual analysis as well as the positive, negative or neutral connotation in a sentiment analysis.
The system feeds Thomson Reuters news through this blackbox to produce data across 80 different variables for all articles. Algorithmic trading systems then employ the analysis output to make trades on the expected effect of the news on the share price.
This type of automated trading attempts to take advantage of the effect that news has on the share price as well as the ability to analyse and act on the information more quickly than any human could by reading the news.
The advent of social media has extended the possibilities for gauging investor sentiment and the public mood by applying sentiment algorithms to what people are saying on Twitter, Facebook or message boards. Given that Facebook has over 900 million active members and Twitter more than 250 million, the appeal of a broader content set and new metrics, for example for consumer confidence, is clear.
The type of analysis is essentially the same that is used for traditional media and the search of messages can be narrowed down to conversations referring to a specific stock, commodity, company or product.
Ravenpack is one of several news analytics providers that combines traditional news media content and social media sentiment analysis. The firm recently released a set of algorithms to analyse foreign exchange and commodities-related macroeconomic and geopolitical news items.
Other analytics firms focus their efforts solely on social media. Topsy Labs Inc. announced in January it would release a Twitter trading tool later in the year.
The firm, a provider of a real time Twitter search engine and sentiment analysis platform, also published a whitepaper on predicting stock prices using social sentiment in January of 2012 using the example of Netflix.
Topsy selected all tweets that included the terms “Netflix”, “NFLX” and “Qwikster” and compared the positive, neutral or negative sentiment of the messages to the end of day Netflix share prices. The paper found a correlation between positive and negative sentiments expressed in Twitter with the movements of the company’s share price. The effect was more pronounced for share price declines, the study said.
“You have to look at Twitter sentiment as a giant opinion poll that you can take for products and companies all the time,” says Rishab Ghosh, Topsy Labs Inc co-founder in an interview with CNBC.
He claims that if investors had used Topsy sentiment data on the day of the Apple 4S launch in October 2011 to buy Apple stock, they would have made a 20 per cent return in two weeks, even though the initial negative media and analyst coverage knocked 5 per cent off the share price on the opening bell.
Topsy’s analysis found a much more favourable sentiment expressed on Twitter by potential consumers. “Since consumers, not analysts, are those who actually drive revenue for the iPhone, traders would have really benefited from getting this kind of insight into consumer opinion,” Topsy Labs said. When the iPhone 4S was released days later, sales in fact soared.
The jury about the trading success of sentiment analysis based investments is still out.
Ghosh said: “I have to caution, we are not in the business of trading stocks. If we came up with a magic formula, we would be a hedge fund and not a technology company. We are just trying to show that there is signal out there that we are able to analyse.”
Derwent Capital Markets
Meanwhile others have tried to put social media sentiment analyses into action. Cayman-registered Derwent Absolute Return Fund, which was dubbed by some media the “Twitter fund”, became the first fund that attempted to predict market movements on the basis of the sentiment expressed in Twitter messages. By analysing all Twitter messages worldwide according to their mood state, the fund looked to predict changes in the equity markets.
The idea for the fund came from a research paper Twitter mood predicts the stock market by Johan Bollen, Huina Mao and Xiao-Jun Zeng from Indiana University, who in 2010 examined the correlation between the mood states in Twitter messages and the Dow Jones Industrial Average closing values.
Not only did the paper find a correlation, but a predictive correlation, that revealed a three day lag between a change in overall sentiment and a corresponding movement of the Dow Jones. Significant changes in sentiment were reflected in the magnitude of the correlated Dow Jones movement, Paul Hawtin the fund’s founder and fund manager claimed. Hawtin said the daily direction of the Dow can be predicted with 80 to 90 per cent accuracy.
The analysis employed by Derwent did not filter the approximately 1 billion Twitter messages a week, for example by focusing on individual Twitter users such as CEOs or traders or those tweets that are referring to a particular financial instrument or company. Consequently the results included anything from the 12-year-old girl gushing over Justin Bieber to specific stock related tweets.
Bollen admitted that even though his research strongly indicates a predictive correlation between the moods expressed in tweets and the equity markets, he does not know the details of what causes the effect. “Something real is going on, even if we are not sure how.”
The tweets were searched for key words in seven languages according to a lexicon of terms that relate to six different dimensions of mood states (calm, alert, sure, vital, kind and happy) and run through an algorithm. The output came in the form of a graph on a daily basis, expressing the standard deviation from the previous day.
This information was then used to construct a basket of liquid instruments, including futures on indices and the most widely traded stocks in the S&P 500, Dow Jones and FTSE100, to take advantage of the expected movements by going long or short on the market two to four days after a significant mood change on Twitter.
According to Hawtin, the system was not an automated black box for trading. In the case of obvious global events that were unrelated to the financial markets, the mood signal still needed to be validated and would not always be acted upon.
Although the hedge fund showed promising yields during a trial period and made a 1.87 per cent return in the first month after launch, it was closed down shortly after. Hawtin blamed the market timing for the closure of the fund. “2011 was one of the worst years in history for the hedge fund industry which made it extremely difficult for us to raise sufficient capital to run the fund effectively,” he said in his newsletter. The company is now focusing on a sentiment analysis technology for the retail sector.
It appears that while the relevance of news and social media for investing is widely recognised, it is a totally different issue to effectively incorporate sentiment analysis in quantitative models and devise a successful trading strategy.
Whether social media sentiment, as Bollen said, is really “the canary in the coal mine” for the equity markets will depend on a number of factors, including the ability to analyse text accurately for meaning and sentiment, the ability to identify the sources of traditional and social media that have the strongest indicator effect and to devise trading strategies that are able to translate the predictive ability into actual investment returns.