Stock traders are fallible. Sometimes they act based on gut instinct rather than sound logic. That’s why Sentient Technologies (known as Sentient.AI) of San Francisco is working on Sentient Investment Management, an artificial-intelligence program that decides which stocks to buy and sell.
“Our AI system can be more consistent and reliable,” says IEEE Fellow Risto Miikkulainen, the company’s vice president of research. “It buys and sells stocks based on an entire history of data rather than, for example, on a single unreliable piece of information that a stockbroker might fixate on.”
The company’s researchers have been developing the program for the past 10 years to manage its investments. It relies on evolutionary computation, in which computers are fed massive amounts of data and, after hundreds of thousands of trials at making decisions, the machine evolves to make better ones. With Sentient’s program, that data includes a history of changes in stock prices and past returns on its investments.
The program, which learns to make better trades as it goes, now makes all the trading decisions for Sentient’s investments.
YEARS OF TRAINING
Evolutionary computation is a form of machine learning, Miikkulainen notes. With standard machine learning, computers are fed a large data set containing hypothetical situations and told what the right answers should be. “Evolutionary computation is more like learning to play a video game,” he says. “We didn’t know what the right actions were at the beginning. Instead we learn by exploring and trying out different things.”
Evolution uncovers what Miikkulainen calls candidates. Different candidates emerge as the system is tested, and the candidates learn by trial and error. If they make bad trading decisions—making small gains or, worse, losing money—they’re discontinued. But if they make smart, profitable choices, they live to make another simulated trade.
At its peak, Miikkulainen says, Sentient’s system was testing 40 trillion candidates per year. Each one is tested hundreds of thousands of times, and if it passes them—essentially, if it makes mostly intelligent, profitable decisions—then it is deployed to make actual trades.
“These candidates wake up each morning and buy and sell stocks, just like day traders, only there’s no human oversight,” he says. “They are completely autonomous.”
HUMANS STILL NEEDED
AI lacks at least one important quality: common sense. “Automated stock traders are basically idiot savants,” Miikkulainen says, meaning that they can do only the thing they’ve been trained to do. Sometimes they don’t recognize that something has gone wrong.
During a major event that could negatively affect stock prices, there always will be a “panic button” that a trader can press to shut down the automatic system, Miikkulainen says. That might include the outbreak of a war or the unexpected results of an election. Because such events were not part of the AI program’s training, it could not factor them into its trading strategies.
Sentient Investment Management is getting smarter by the day, Miikkulainen says. For now, the product is not available to customers; it is still being tested internally. It could eventually be applicable elsewhere, though. “After proving the technology, we’ll expand to other markets,” Miikkulainen says.
No matter what, minders with expertise in engineering and finance will still be required. “For the foreseeable future, these systems will need to have a human companion, providing the common sense,” he says. “AI is simply an advanced decision-making tool.”
IEEE membership offers a wide range of benefits and opportunities for those who share a common interest in technology. If you are not already a member, consider joining IEEE and become part of this worldwide network of more than 400,000 students and working professionals.