Finding Novel Ways to Trade on Sentiment Data

It’s no secret that hedge fund managers are always looking for new sources of data that will help them in their never-ending quest to beat the market. Quantitative researchers at Bloomberg have been developing innovative methods to help reveal embedded signals in one of the more popular sources of unconventional financial data: sentiment analysis of news stories and social posts.

“Everyone is looking into alternative data sets, sometimes without really understanding their value,” says Dr. Arun Verma, Ph.D., a researcher who leads the Quant solutions team within Bloomberg’s Quantitative Research group, which is headed by Bruno Dupire. “They are looking at data like sentiment, supply chain relationships, and even things like satellite imagery. Often Machine Learning methods are applied to optimize alpha from such data, but a lack of scientific rigor can lead to poor out of sample performance. We avoid the trap of extreme data mining by using robust statistics.”

Since 2010, the Bloomberg Terminal has offered sentiment data based on news stories about companies. Over time, the data set has been further expanded through the addition of more sources for news stories, other entities, as well as sentiment on social media.

In some early experiments, Bloomberg’s Quantitative Research team devised sentiment-driven strategies that could beat the market by double digits. The researchers caution that traders in the ‘messier’ real world should not expect to do quite as well due to transaction costs and other inefficiencies. “We wanted to show some ideas for how sentiment can be used to make money,” says Xin Cui, another quantitative researcher on the team. “What we have shown is that the data has value on its own. It’s up to the portfolio managers to integrate the sentiment signal with their existing proprietary strategies.”

Verma recently presented some of his group’s findings at The Trading Show in Chicago on May 17th. He’ll also be speaking at AI, Machine Learning and Sentiment Analysis Applied to Finance in London on June 28th.

Using Sentiment to Build a Portfolio

The process of leveraging sentiment data starts with using machine learning to assign a rating to every news story about a stock: positive (+1), negative (-1), or neutral (0). The scores of every relevant story over the past 24 hours are then averaged together, resulting in a single daily sentiment score that is available before the market opens. Sentiment data can also be derived from Twitter, with each tweet treated as a very short news story.

Verma’s team then designed some simple long/short trading strategies based on the sentiment scores. In one strategy, they used the Russell 2000® stocks as their universe. “The daily sentiment ranking can be considered as a lagging indicator,” says Cui. “And for large cap stocks, the market is very efficient. There’s less chance to capture anything after a day. However, for small cap stocks, the market is not that efficient, and so, it takes longer for the market to digest all the information.” This is illustrated in Figure 1 below.

The Bloomberg team proposed the following strategy: buy stocks in the top third of their sentiment rankings and short those in the bottom third. At the end of each day, they would clear out the portfolio. In a variation, they bought the top five percent of stocks and shorted the bottom five percent. Then they back-tested these strategies from January 2, 2015 to March 31, 2016. “Diversification across a large universe of stocks is the key to getting a better signal to noise ratio,” says Verma. “Sentiment factor can be too noisy to extract a good result on a single stock or a small basket of names.”

The raw results were impressive: Just trading on the top and bottom third of sentiment scores resulted in an annualized return of 23 percent with a highly significant Sharpe ratio. Other strategies returned as much as 38 percent, without dramatically increasing volatility (keep in mind; these results are not fully achievable in practice due to transaction costs).

Figure 1: Sentiment strategies perform better on small caps stocks

They also tried incorporating a dispersion score – a measure of confidence in sentiment – and combining sentiment signals from News & Twitter. Figure 2 shows that all the sentiment strategies outperform the market. Using the dispersion indicator helps get better performance, but even better result can be achieved with a combination of News and Twitter signals.

Figure 2: A dispersion indicator helps get better performance, but even better performance can be achieved with a combination of News and Twitter signals

Another strategy focused on earnings reports. Using S&P 500® companies, the researchers bought stocks with positive sentiment and shorted those with negative sentiment the day before earnings, and conducted a similar back-test. “With large cap companies, you have a lot of analysts following them and there are more chances to have rumors around earnings,” says Cui.

Using results from both Twitter and news sources, this strategy achieved an annualized return of 156 percent coupled with higher volatility still producing significant risk adjusted returns.

Despite the strength of these numbers, Verma and Cui aren’t exactly recommending that anyone use these strategies directly, since they involve very large portfolios, and trading costs can eat away much of the returns. “These are high turnover strategies,” says Verma. “There might not be enough liquidity to execute them every day and one may not be able to exactly obtain the open price that is used in back-testing this strategy. However, a trader might be able to reduce their transaction costs, since stocks could remain on the long or short side for a few days at a stretch. We assumed the portfolio manager was paying the full bid-ask spread and clearing the portfolio every night.”

Verma and Cui are continuing to look for new wrinkles in sentiment data, seeking out corners of the market where signals might be even stronger. Still, the early results are encouraging. “The sentiment signal, derived directly from unstructured textual data, has low correlation to other traditional alpha factors,” says Cui. “That’s really important since that’s what portfolio managers are looking for – something that can bring additional uncorrelated alpha to their strategy, making it easy to deploy within their larger Quantamental Factor Model setup. Essentially, we are recommending sentiment as a factor in a smart beta strategy framework, though on a much shorter time horizon than what is typical in the factor investing context,” adds Verma.