J.P. Morgan’s AI research head: real data overvalued

At a recent conference, Manuela Veloso provided an overview of J.P. Morgan’s research efforts in trading and data management using artificial intelligence (AI). Veloso is the Herbert A. Simon University Professor in the School of Computer Science at Carnegie Mellon University on leave at J.P.Morgan AI Research.

The focus of Veloso’s talk was on AI for intelligent financial systems, and she began by explaining that AI is a “science of components”, differentiating it from machine learning, as the terms are often used interchangeably.

“Machine learning is a fundamental component…but AI is in fact many sub-areas and one day we will eventually put them all together in complete agents and we are doing this also at J.P. Morgan,” she said.

One of the first projects she highlighted came from observations from the trading floor: that decisions are made using visual information. Her team changed the time series data into pixels and used historical data trained on human decisions to train neural nets on the “buy” or “not buy” decision.

Results showed that the learning model performs at 95% accuracy and 94% precision on historical S&P500 data. While that doesn’t mean it is making the “right” decisions necessarily, it is reproducing the decisions humans make by using images of trading signals.

Another project is in creating simulations of the multi-agent environment that represent the markets. The simulator can generate an infinite number of similar market days, she explained, and therefore, one of the outcomes was that the team was able to generate synthetic data.

“Real data seems to be overvalued. Real data is great, but when it’s not available we can still, in AI, make progress and use synthetic data which is close to real…to make all the developments, all our algorithms, all our understanding. And then, eventually, when needed, we apply it to the real data,” she said, adding that J.P. Morgan will be making the data available, so “stay tuned”.

Veloso also discussed projects that automate data translation into standardized formats as well as report generation. Research publications can be found here. 

Related Posts

Previous Post
Reuters: Goldman Sachs restarts cryptocurrency desk
Next Post
State Street Completes First Trade with ICE Clear U.S. as a FICC Sponsored Member

Fill out this field
Fill out this field
Please enter a valid email address.

X

Reset password

Create an account