With a historical two-day market shutdown in the wake of Superstorm Sandy, emotions on Wall Street are running high. Perhaps it's only fitting, then, that we talk about the use of sentiment analysis in the investment and trading industry.
Last Friday, in the calm before the storm, I talked with Rich Brown, head of Elektron Analytics at Thomson Reuters, about how sentiment is coming into play in making buy/sell and other investment decisions. Brown, who is in New Jersey, had been scheduled to speak at Tuesday's Sentiment Analysis Symposium in San Francisco. Aleksander Sobczyck, head of quantitative research-machine readable news, presented in his place.
As I wrote in a March post, Thomson Reuters News Analytics (TRNA) scores digital content for sentiment, relevance, and novelty, helping its newsfeed clients better capture market opinion and determine their next moves based on "what's going on out there." In March, the news was that Thomson Reuters had extended its machine-readable news offering with a sentiment scoring system for social media.
Use of the social media scoring remains "really niche" for now, Brown told me. After all, as Brown himself pointed out, "Lengthier content lets you get better sentiment signal than does 140 characters. Tweets don't give you enough context to understand the implications." Blogs are better, but news analysis, closed captioning feeds, conference call transcripts, and other, more traditional content types, are even more so, he said.
Sentiment analysis in the investment and trading sector is still leading-edge stuff, Brown said. He placed the number of clients using sentiment in the scores rather than the thousands, and said fewer than 100 clients are exploiting the social media sentiment scoring at this point.
But, as he noted, the numbers have been, and will continue, growing:
As more and more firms are writing research to prove how you can make money off of sentiment analysis and use sentiment in your trading models, it's becoming more and more common. Eventually, sentiment will be a signal everybody has to have in their models.
TRNA delivers sentiment scores on a 100-point scale, showing content's positivity or negativity. An article talking about a company's "challenging management environment" might get a score of negative two, while a piece discussing how results are "exceeding expectations" might get a positive three. The engine uses entity-level scoring, so if a news article disses Microsoft while singing Apple's praises, then Microsoft would get a negative sentiment score and Apple a positive one, Brown added.
To date, Thomson Reuters has seen clients use the sentiment scores in a variety of ways. The most common are:
- As a circuit breaker. "If news comes out on IBM, the client will stop trading on it till a human can evaluate whether it should continue buying or selling -- whether it changes your investment hypothesis," Brown explained.
- To change trading patterns. The rate at which news flows can be a predictor of future volume and volatility spikes. "So, you might trade faster when news comes out to get a better price on the market before it runs away."
- To "generate alpha." This is the staple "buy the good news, sell the bad news -- and many different versions of that," Brown said.
Of course, according to Brown, this is all "amazingly sophisticated." I'd add, "complex," too. TRNA uses 82 variables, which get combined in any number of ways. Then take into consideration the number of different time horizons, the number of different content sources, and so on -- and, well, you can see how "complex" easily substitutes for "sophisticated."
Some help comes through a visualization toolkit Thomson Reuter also offers. "It can put thousands and thousands of stories in greater context, and see how that can confirm or go against an investment hypotheses," said Brown. A client, for example, might color code a country or sector for an at-a-glance look at sentiment. Tech sector companies might appear in green, financial services in an orangey color, and healthcare, with the nebulousness surrounding the impending election, yellow.
"If a picture is worth 1,000 words, an intelligent visualization on news analytics can be worth a million articles of 1,000 words each," Brown quipped.
How will the horrors of the hurricane translate into content tone picked up and scored by TRNA, and used in making investment decisions? I guess that all depends on who's doing the decision-making, whether they're doing so with the short term or long term in mind, and, oh, about a jillion other factors. Let's leave that to the analytics engines.