Quest for Robo-Yellen Advances as Computers Gain on Rate Settersby
‘Capability is here’ for AI to inform officials in five years
Future policy to mix human judgment with self-taught computers
Move over Janet Yellen, automation in the workplace is about to get personal.
Instead of relying on the Federal Reserve chair, imagine using a computer to transform mountains of raw economic data into reliable predictions for unemployment, inflation and gross domestic product. What’s the best level for the federal funds rate? Press <Enter>.
“The capability is here,” says Andrew Lo, director of the Laboratory for Financial Engineering at the Massachusetts Institute of Technology, near Boston. “The biggest hurdle is the cultural barrier. You’ve got a lot of central bankers who are not as open to technology.”
Artificial intelligence, or AI, may be on the verge of improving the economic forecasting that serves as the foundation for monetary policy making. The Bank of England, behind Chief Economist Andrew Haldane, has quietly become a pace setter among central banks in exploring the technology’s possibilities.
Paul Robinson, who heads the bank’s two-year-old Advanced Analytics unit, says the goal is to assist rather than replace human economists. He predicts AI will make a meaningful contribution to monetary policy making “certainly within five years.”
Improvements would be welcome. Economists are, by their own admission, notoriously bad at making predictions. Consider the forecasts for 2015 GDP issued by Fed policy makers at the end of 2014. All 17 overestimated growth.
Central banks have long used computers to construct their most advanced models of the economy, which are useful for generating answers to specific what-if questions: What will happen to unemployment if 10-year Treasury yields rise by 2 percentage points? But when it comes to making real-world predictions, models are often worse than humans.
Enter machine learning, the dominant sub-field of AI. It refers to technology that allows a computer to acquire a skill for which it hasn’t been explicitly programmed. At its heart, it’s an automated process for recognizing patterns in data and transforming them into possible solutions for a given problem. The more data, the better.
Google’s self-driving car is an example. No team of programmers instructed the car how to respond to every potential scenario on the road. Instead, it learns to drive by detecting patterns in vast amounts of data generated by real drivers.
Hedge funds, such as Two Sigma and Renaissance Technologies, use machine learning to help make investment choices. Amazon.com Inc. uses it to predict customer purchases and Netflix Inc. to recommend movies.
At central banks, the principle task is to set an interest rate on short-term borrowing that guides an economy to a sweet spot between overheating and stalling. Because rate changes work with a lag, getting them right depends on forecasting economic conditions 6 to 12 months down the road. If machine learning makes forecasting more accurate, central banks can do a better job.
That’s a lot more complex than recommending movies or even picking stocks. The number of variables at work is much broader -- ranging from the strength of a dollar to more subjective factors, such as the possibility of a Donald Trump presidency. The world is also highly dynamic: Even long-understood connections between economic variables can change.
For example, the inverse relationship between unemployment and inflation -- what economists call the Phillips Curve -- seems to have disappeared in recent years, much to the puzzlement of economists. Making it even trickier, by changing interest rates, central bankers are intervening in a way that alters the very conditions they’re trying to predict.
All that makes some Fed officials, despite their enthusiasm for technology, skeptical the technology is anywhere near matching the performance of humans at economic forecasting.
“There will remain an irreducible amount of uncertainty that is very large,” says David Wilcox, director of Fed’s division of research and statistics. “Moreover, the phenomenon of structural change is pervasive in complicated economic systems. For the foreseeable future, the best approach will involve a combination of empirical rigor captured in models, together with human judgment.”
Michael Kearns, a computer science professor at the University of Pennsylvania, doesn’t buy that. He suspects there’s nothing about human judgment that a machine can’t learn. Even intuition, he says, is simply knowledge derived from past experience that can’t be articulated. “And what is experience other than data?”
Therein lies another catch. Even if the technology is ready, are macroeconomic data rich enough to allow computers to produce superior economic forecasting?
No, says Hal Varian, Google’s chief economist and an AI champion. “The data sets are so small. GDP is released quarterly, so 50 years of data is only 200 observations and only seven recessions,” he says.
Varian concedes that contemporary data are growing exponentially. Quarterly aggregated GDP could soon be supplanted by information on billions of transactions reported in real time. Web scrapers, such as MIT’s Billion Prices Project, already comb the Internet for real-time price points relevant to inflation. Still, he says, machine-learning algorithms would need equally dense historical data to make machines “dramatically better” than humans. And that just isn’t available.
Should machine learning surmount that hurdle, don’t expect the Fed or any other central bank to turn monetary policy completely over to a machine. Economists and computer scientists agree there will always be a role for humans in making monetary policy.
As Michael Feroli, chief U.S. economist at JPMorgan Chase & Co., puts it: “I don’t see why, in principle, you couldn’t have a computer set monetary policy. Having it testify before Congress is another matter.”