Hedge Fund Founder and Amazon Researcher Build a Better AI Stock Picker
(Bloomberg) -- A hedge fund manager and a computer scientist have found a promising new way to use artificial intelligence to pick stocks over longer periods than the typical machine-driven approaches favored by Wall Street.
In tests, John Alberg, co-founder of hedge fund firm Euclidean Technologies, and Zachary Lipton, a researcher at Amazon.com Inc.’s AI lab, generated 17.1 percent annualized returns from their technique, compared with 14.4 percent using a standard statistical model, according to a paper they discussed during a Friday workshop at the Neural Information Processing Systems (NIPS) conference, an annual gathering of experts in the field.
The research comes at a time when financial firms are rushing to embrace AI. At this year’s NIPS conference, hedge funds and investment banks vied with big technology companies to recruit specialists in neural networks, a kind of AI loosely based on the human brain that has fueled recent advances in voice and image recognition by computers.
Only a handful of firms have used neural networks for trading and investing. And those that have are mostly focused on complex trading strategies, often over short time periods. Alberg and Lipton’s research suggests a deep neural network – one with many layers – can be effective for longer-term stock picking when it’s fed mountains of fundamental corporate information like profit, revenue and debt levels.
They gave their system 16 common types of data from financial statements and four other metrics on stock price movements over one-, three-, six- and nine-month periods. They did this for all the stocks on the New York Stock Exchange, NASDAQ and American Stock Exchange for at least 12 consecutive months from January 1970 to September 2017.
Originally, the two asked the neural network to take in five years of data and then try to forecast future stock prices one year out. But that performed no better than a standard computer-based trading model.
“The price bounces around a whole lot, independent of anything actually happening, at least in the short run,” said Lipton, who is also at Carnegie Mellon University. This erratic movement is called “noise” in AI and data-science circles and the neural network got distracted by it, Alberg added.
So the duo tried a different technique. Instead of asking the neural network to forecast the stock price a year out, they asked it to predict the future value of company fundamentals, getting a forecast for things like earnings, before interest and tax (EBIT). They then divided that by each company’s current enterprise value, to end up with a kind of AI-powered, forward-looking valuation multiple. They invested in the 50 cheapest stocks based on that metric.
“If you partition the problem into two steps – predicting future fundamentals from historical fundamentals, and then use those future fundamentals to predict price – the complexity of deep learning can be made useful and improve the model,” Alberg said.
The two said they will continue researching this area. One future project: Seeing if their neural network is better at forecasting company fundamentals than human equity analysts. They also want to see if performance will improve when they feed the system other data that contain clues about a company’s future, such as executive comments on earnings conference calls.
©2017 Bloomberg L.P.