Interpreting Alternative Data in Novel Ways Using AI

March 7, 2019

How smart businesses utilize continuously learning AI systems to automate analytical workflows and extract hidden knowledge in proprietary data to make predictions about the real world that impacts financial markets.

Alternative data, which was once conceptually new and novel, has become commoditized as every institutional investor utilizes the same types of datasets in their investment calculus. Alternative datasets were once very expensive and difficult to obtain but an explosion in alternative data vendors has exerted pricing pressure and increased accessibility. The widespread adoption of alternative data has consequently left institutions scrambling to find new ways of extracting alpha.

Rather than focusing on the obtainment of exclusive alternative data, quantitative investors should place more emphasis on interpretation and analysis. Traditional, static quantitative methods (factors defined a priori) will not be as effective as continuously learning AI systems that can unlock novel features hidden beneath complexity that surface predictive insights invisible to both humans and machines employing static models.

Dynamic vs. Static Learning Models

Most “AI” in the space offers static learning models as if they are truly autonomous and generally intelligent. However, the problem is that static “AI” produces only a linear return on the cognitive effort required to train the system. This is because the system relies on experts to identify features. Such static models fail to produce accurate insights when the context of information changes. Since it is impossible for a human expert to account for all the possible factors that could impact a specific future event, static machine learning models require continuous training and are unable to generate sustainable alpha.

Truly autonomous learning systems capable of accurately identifying essential features across multiple domains (i.e., general AI) is still science fiction; however, dynamic continuously learning models are real and useful in interpreting alternative data in novel ways. Unlike static AI, dynamic learning systems are constantly evolving in terms of feature richness because they are constantly adapting to changing contexts with continuously decreasing amounts of human interaction. Ever-evolving, cognitively generated features help quants modulate signals produced by proprietary models built to predict behavioral impulses in relation to future events.


For more information on how Accrete is employing dynamic continuous learning to transcend alternative data, please visit our website.