Feeding news sentiment data into machine learning algorithms can vastly improve trading and investing performance. A white paper focusing on Euro STOXX 50 and S&P 500 companies demonstrates the forecasting potential of Refinitiv News Analytics.
- Machine learning, artificial intelligence, and unstructured or alternative data sets are increasingly being used to improve market performance.
- The application of Refinitiv News Analytics makes sentiment data fit for consumption within machine learning algorithms.
- Research involving 100 of the most liquid S&P 500 stocks shows up to three times more return on average compared to buy-and-hold strategies.
For more data-driven insights in your Inbox, subscribe to the Refinitiv Perspectives weekly newsletter.
In markets driven by intense competition and the need to improve performance, innovation is key. Machine learning, artificial intelligence, unstructured data, and other alternative data sets are increasingly being implemented to meet this need.
My recent white paper, which includes research completed in 2019, describes how the performance of machine learning algorithms can be enhanced by feeding them with alternative data.
In this particular case, the approach is based on news sentiment-based machine learning forecasts that can drive up the performance of trading and investing strategies.
Download the ‘Learning With Sentiment’ whitepaper now
Benefits of news sentiment data
There are many applications of machine learning for market forecasting, but these follow traditional technical analysis methods of looking for patterns in past price behavior, and are limited by a lack of current information.
By adding more immediate news sentiment data into machine learning algorithms, it becomes possible to significantly improve price forecasts and risk-adjusted performance.
This is not necessarily an easy task with unstructured data that cannot be directly consumed by standard quantitative algorithms.
However, the problem can be resolved by using natural language processing (NLP) techniques to digitize the data and make it fit for consumption.
Refinitiv’s News Analytics solution is based on an NLP engine that reads and interprets the news hitting Reuters and other newswires in real time.
The news covers about 33,000 publicly-traded companies around the world and dozens of commodities.
Eric Fischkin, Proposition Director for Machine Readable News at Refinitiv, says:
“Interest in structured sentiment data has risen rapidly in recent years, with the increased adoption of machine learning along with robust research on the benefits of this data to boost investment strategies.”
Two applications of Refinitiv News Analytics and machine learning algorithms that illustrate Fischkin’s comments are intraday forecasting of the EURO STOXX 50 index and daily trading of the 100 most liquid S&P 500 stocks.
EURO STOXX 50 forecasting
The EURO STOXX 50 (SX5E) application is designed to forecast the direction of the index, which consists of 50 of the largest companies by market capitalization in 11 Eurozone countries.
Forecasts are made at five-minute intervals, with prices recorded in the same time span and corresponding to 102 quotes a day.
The application includes a deep learning network, which has been trained using several months of five-minute price and sentiment data, and the Refinitiv News Analytics database.
Only news relevant to SX5E companies is selected by the network, although this selection is changed every three months to reflect changes in the index. Based on this data, forecasts are made by the deep learning network, which can ‘remember’ previous states and use them as input to subsequent iterations of its ‘thinking’.
In this application, the network was used to test machine learning forecasts for four trading strategies during one random week.
The results show returns relative to specific trading strategies and demonstrate that feeding news sentiment into machine learning algorithms can provide better short-term forecasts of index price direction that can be used to improve performance. The improvement is particularly pronounced when forecasting downward index movements.
Trading S&P 500 stocks
This application of news sentiment-based machine learning forecasts is about trading individual stocks, essentially the 100 most liquid stocks in the S&P 500.
Sentiment scores including relevance, novelty and volume were aggregated for each stock over one trading day.
The 10 most newsworthy stocks attracted up to 45 news items a day, while the whole universe averaged 1.4 items a day.
By linking sentiment scores with lagged prices and trade information, the goal was to predict the next day’s price direction and generate trading signals based on the forecasts.
The machine learning forecasting algorithm in this case was based on the NeuroEvolution of Augmented Topologies, which selects the best neural network for each iteration of the algorithm.
Again, the results of using sentiment-based machine learning forecasts show significant gains in the performance of each stock compared with buy-and-hold strategies.
While performance is best for the most liquid and most newsworthy stocks, which typically are not the same, the overall result of trading based on machine learning forecasts generates up to three times more return on average for these stocks compared to buy-and-hold strategies.
Download the ‘Learning with Sentiment’ whitepaper now
Sentiment remains underutilized in algorithmic trading, but our applications underline its potential to improve trading and investing performance, and add value to the business.
Listen to my recent webinar with Refinitiv, ‘The role of sentiment as an early warning risk indicator’ for more on using sentiment in investment strategies.
Enhance your research and trading strategies with Refinitiv News Analytics
