What if you could instantly scan all individual posts on Twitter for one day, cull those snippets together into a cogent whole, and use that information to paint a picture of the global mood?
If that idea alone isn’t enough, imagine making the snap decision to rush out and buy stocks because within three to four days stocks will rise due to the positive “vibe” in the air as foretold by the collective Twitter chatter. Conversely, if the world is having what amounts to a bad hair day, you accordingly sell your stock holdings, knowing that within three to four days, that dour zeitgeist will portend a drop in the Dow.
Does this sound to you like yet another flimsy system to sell traders on the idea that this might be the next big secret? Does it, like all other stock-related get-rich-quick schemes (and let’s face it, just because it comes out of academia doesn’t mask an unmistakble sense of self-interest) seem too good to be true?
For years scientists and speculators have tried to pin down the mysterious changing tides of the stock market. Proof? Search for “stock predictions” to find millions of options from informed analysis and offers of psychic or spiritual guidance. So far nothing has hit home for long enough to be a tried and true standard for evaluating buy or trade decisions.
That “too good to be true” paradigm for market predictions might be upended, however, thanks to our endless tweets and social media updates that indicate our mood, both through words and emoticons–not to mention an expensive array of compute-end tools to tackle massive unstructured data sets in a flash.
The provocative predictive analytics study in question proved a direct correlation between overall stock market performance and the general mood of many thousands of people as gauged from their brief posts on Twitter. While the model for buying and selling described above only works around 86% of the time, this news caught the attention of traders and computer scientists with equal force.
As a recent article noted, “Online surveillance of social networking sites is emerging as a must-have tool for hedge funds, big banks, high-frequency traders and black box investment firms that run money via computer programs.” The author goes on to note that your feelings and general mood, captured and combined with the rest of the Tweeting, Facebooking world, could become the core of decision-making processes at major financial institutions.
Dr. Johan Bollen teaches informatics at Indiana University and is the lead behind the Twitter mood informatics project. He noted in a recent interview that this marks the beginning of a new era of mood collection to measure stock performance, noting that it is indeed like science fiction that we can now have “a large-scale emotional thermometer for society as a whole.”
In an interview this week Bollen told us that while he can’t share specific details about storage, application layers and the like since his team is in the process of further developing and licensing the research, the processing of tweets is happening in “real time” although it all depends on how one defines “real time.”
In Bollen’s words:
There is clearly a lower bound of the temporal granularity at which you can compute these signals. This limit is largely shaped by processing speeds and the amount of data to process. The amount of Twitter and social media data keeps growing very fast, however at some point you could expect all 7 billion people on earth to have a Twitter or Facebook account and since no one can tweet faster than their thumbs will move across a smart phone screen.
We may find some upper limit on the amount of social media data that can feasible be generated by humanity. From that you can work back to determine the temporal granularity at which you can operate the existing computational limits at a particular point in time.”
In Bollen’s experiences with the computational angle to arrival at global sentiment, he says that he remains optimistic that his team will be able to generate these signals at very small temporal granularity. He noted that as of now, they are “easily processing daily feeds meaning we have a daily signal which is suitable given that our research has shown that this signal is predictive of real-world changes 3 to 4 days in advance. From a preliminary analysis it seems we can take this down to hourly or even half-hourly signals without too much trouble.”
Bollen claims that while all of his work is quite CPU-intensive, much of it can be parallelized because they are analyze each tweet in isolation. With this in mind, however, he claims that they do run into some pretty hard computational limits with their social network analysis as some of the existing algorithms simply cannot run over social networks of such size.
As another element in the computational depth involved in such an undertaking, Bollen told us:
In terms of data intensiveness you do need large-scale data to counter-act noise and other distortions, but there is definitely a law of diminishing returns. Many of these data sets follow very skewed distributions. In general terms you will have very few people making very large and significant contributions, and very many making small and insignificant contributions. By capturing the right subset you can therefore arrive at a much smaller data set that still provides 90% of your signal, and thereby greatly boost your ability to perform your analysis at very short time intervals.
Outside of CPU and signals, there are other challenges that might stand in the way as this idea potentially takes off for financial companies.
This data, which comes in from across an array of global social networks creates a massive pool of unstructured data, could prove a stumbling block for the widespread viability of this kind of real-time data analysis.
Xenomorph is a data analytics and management firm with roots in financial services. According to its CEO, Brian Sentence, “Real time social media feeds give some insight into the human behavior that really drives the markets. However, in addition to the challenge of processing such a large amount of data, correct understanding of the data is the biggest challenge. For instance, seeing “Hathaway” mentioned on Twitter might mean some news on ‘Berkshire Hathaway’ or the actress “Anne Hathaway.”
However, this type of analytics goes far beyond general semantics and natural language processing—it looks at more discreet indicators of mood, including emoticons and other less word-bound cues.
Henry Newman, CEO and CTO of Instrumental, Inc., which is a consultancy firm for users and manufacturers of HPC, also weighed in on the data-level challenges of such analytics. He noted, “There are a number of challenges in this areas including the capture and indexing of the data and of course the development of algorithms to correlate the trends to specific market changes. Additional challenges include the long term storage and the analysis model. I am aware of some sites MapReduce to be able to search this type of data but these are still problems.”
Instrumental Inc.’s Henry Newman also made a good point about the use of predictive analytics for financial markets in his speculation about the real value of this kind of technology. As he said, “I am not a sociologist and have not looked to see if this type of analysis will provide trending information for trading. What I am sure of is that if it does work just like every other method used it will not work all of the time and could cause large market swings just like other methods we have seen over the last few decades.”
The takeaway here if you’re not a stock market player: Be careful what you tweet, the world’s economy might just depend on it…well, at least 86% of the time. Allow me to do my part for the vitality of world economies and end this piece with a 🙂