Machine learning and natural language processing are creating new categories of alternative data, as exemplified by Accrete.AI which offers a service which identifies and ranks sources of M&A rumors including blogs, social media, articles and chat rooms. There are multiple research applications for AI but the most interesting are those which derive fresh insights.
Accrete defines an M&A rumor as chatter about a potential acquisition target that has yet to be verified by either the target or the acquirer. Accrete’s Rumor Hound currently reads and ranks over 114,000 sources to identify potential takeover rumors, generating an average of 25 rumors a day. Rumor Hound was initially seeded with less than 500 sources and has since organically grown its sources by intelligently crawling links based on the likelihood of containing market moving M&A rumors.
“Rumor Hound not only has expanded its coverage universe, it has also become more sophisticated in its ability to identify market moving M&A rumors buried in increasingly complex linguistic structures,” said Andrés Diana, Head of Product for Accrete.AI, in an interview with Integrity Research. “As source awareness continues to grow, so does the percentage of announced deals captured by the system, now averaging between 30% and 40% of announced deals.”
Reliability of Rumors
Rumor Hound’s learning engine measures the reliability of sources based on factors related to how rumors published by sources propagate in terms of speed and breadth, as well as how those rumors correlate to excess risk adjusted returns in target stocks over various periods of time. Rumor Hound has identified hundreds of low popularity sources, such as twitter handles, that are more reliable than high popularity sources such as Bloomberg or CNBC in terms of publishing rumors that gain enough momentum to move prices. Nearly 60% of the 29 deal announcements that Rumor Hound has identified early emanated from low popularity sources that have high reliability scores.
Users can access not only the source of each rumor but also the rumor language itself, providing transparency on the origins of each rumor. Rumors are pushed to users via API, dashboard or daily report. Users can set sector specific watch-lists as well as email and SMS alerts. The service provides rumor snippets, source links, source popularity and reliability scores, rumored acquirer names if available and all rumor history on the acquisition target. Users can filter rumor alerts based on source reliability, popularity and other metrics.
Accrete is backed by a strategic partnership with IBM that provides enterprise product distribution and secure cloud computing and infrastructure and delivery for enterprises. The firm’s patent pending capabilities span intelligent web crawling, multi-dimensional contextual unstructured data analytics (text, images, audio and video) as well as semantic search.
According to an IBM Research and Watson case study, Accrete has developed a proprietary framework based on an iterative collaboration between domain-specific experts and deep learning systems. Human domain experts create contextual seeds that define semantic boundaries within which proprietary deep learning systems work to find statistical patterns in unstructured data pertaining to specific, expert-enriched dictionaries and type systems. “Our company’s novel approach to seeding deep learning systems with context allows Accrete’s tools to get smarter and uncover hidden features in data despite having sparse training data sets,” said Diana.
Besides M&A rumors, Accrete has applied to its platform to analyze FOMC statements, an application first pioneered by Prattle in 2014; earnings call transcripts — an application launched by Prattle in 2017 and now more widely copied; and supply chain analysis. Accrete is expanding its product line to other applications including extracting drug test results from unstructured data, as well as patient chatter as it relates to clinical drug trials.
According to Accrete, traditional methods to analyzing sentiment in text misattribute weights to irrelevant language like opening remarks by moderators on earnings calls. Accrete’s learning engine parses language in on a topical basis, accounting for topical distributions and bias in language about industry specific topics and then analyzes the ratio of positive and negative keywords and phrases pertaining to that particular topic. Accrete weights management sentiment on a topic by importance placed by analyst questions and the system learns from the interplay between analyst questions and management responses at the topical level.
Accrete was founded in 2017 by Prashant Bhuyan, who spent over a decade in high frequency trading, automating manual order processing tasks traditionally performed by floor brokers. Bhuyan believed that volatility was changing due to the digital revolution and shifted focus to automating specific complex cognitive tasks traditionally performed by human analysts to find alpha in unstructured data.
The firm has 29 full time employees and is recruiting for five additional AI scientist and Industry analyst positions including a biotech/healthcare analyst position to help develop company-specific ontologies and validate AI results.
Accrete is based in New York with a wholly owned research and development subsidiary in India with offices in Mumbai and Bangalore; wholly owned sales and marketing subsidiaries in the UK and Ireland; and a majority interest in a subsidiary called Shinr.ai that helps enterprises unlock the intrinsic value of proprietary data and expertise to generate operational alpha. The company’s expansion has been funded by private investment.
One of the reasons asset managers are spending more on AI than alternative data is that it has applicability beyond the front office. Even within the front office, there are multiple applications for AI. Trading has been a key area, as exemplified in the background of Accrete’s founder. Within research, there are a few different use cases, ranging from automating maintenance research to enhancing earnings analysis. AI is also transforming alternative data.
Accrete.AI provides a good illustration of the synergy between AI and alternative data. The combination of natural language processing and machine learning extracts new varieties of alternative data from the raw material of blogs, social media, news sources and chat rooms. As AI techniques improve, additional alternative data use cases become feasible and data quality improves over time. We get an inkling why S&P splurged over a half billion dollars on Kensho Technologies.
Accrete has come far fast. In eighteen months, it has launched four different alternative data offerings. In part this reflects introducing products originally pioneered by others, such as central bank and earnings call analysis first commercialized by Prattle. However, Rumor Hound is a fully differentiated product and Accrete is furiously working on new applications such as extracting information on clinical drug trials. The firm intends to continue to push the alternative data envelope.