Why Machines Still Can’t Learn So Goodby and
Man AHL took three years to create machine-learning strategy
Algorithms fail at 90% rate in live tests, a quant says
Anthony Ledford and his colleagues at Man AHL spent three painstaking years building a machine-learning model to do something mere mortals often can’t: find fresh ideas in an avalanche of data.
But even Ledford, chief scientist at the $19 billion Man AHL in London, rolls his eyes when he hears people say that machine learning, a type of artificial intelligence, is going to transform hedge funds tomorrow. To Ledford, a lot of the buzz smacks of hype. The technology is more robust than its predecessors but hardly revolutionary.
“There is some real science here, but it’s not the way it’s been portrayed,” said Ledford, who holds a Ph.D. in mathematics. “Some of it is really marketing, and that’s the bit that annoys me.”
Hedge funds, stung by eight years of underperformance, are latching onto machine learning as a high-tech answer to their woes. But Wall Street’s heady search for the perfect money machine has collided with a sober reality. The technology, which learns on its own to find investment ideas by hunting through troves of data, requires a heavy commitment of time and money, and a high tolerance for failure, since most algorithms turn out to be duds.
“I’m concerned that people may have unrealistic expectations of what is possible with the current state of the art,” David Siegel, co-founder of quant pioneer Two Sigma Investments, said at a Bloomberg summit in September. Siegel, a Ph.D. in computer science who says he’s impressed with AI’s performance in many fields, cautioned: “Machine learning systems can easily, with high confidence, make mistakes.”
The performance of AI, which doesn’t have a long track record, isn’t eye-popping. The Eurekahedge AI Hedge Fund Index, which tracks 12 pools that utilize AI as part of their core strategies, has returned an average of 9.3 percent from 2011 through 2015, and is up 13.8 percent this year through September. Those gains mostly beat the average hedge fund in the five years, but often lost to the S&P 500 Index.
Giants like Bridgewater Associates and smaller firms such as Highbridge Capital Management and Simplex Asset Management in Japan are developing machine learning or investing in it. The next-generation algorithms, which build on the statistical tools quants have used for years, plow through financial, Internet and satellite data to find unusual patterns. A certain default premium combined with a particular yield-curve slope, for example, might produce a high probability that a stock price will rise or fall. Finding such “signals” to wager on is the holy grail.
Many say AI will shake up the industry. Fifty-eight percent of managers in a recent KPMG survey said the technology will have a medium-to-high impact on the way hedge funds operate in the future. Jeffrey Tarrant, chief executive officer of Protege Partners, said machine learning and data science stand to affect investment management as radically as Uber has disrupted transportation.
‘Bunch of Hokum’
“Most of the talk about machine learning is a bunch of hokum,” said Douglas Greenig, a Ph.D. in mathematics who started quantitative hedge fund Florin Court Capital. “When people talk about machine learning, I suppose it’s 65 percent a marketing gimmick and 35 percent substance. But that 35 percent can be very good.”
Hedge funds need a team of scientists and researchers and many months, even years, to shepherd a single algorithm from development and testing to live trading. That’s because most algorithms fizzle out along the way.
Martin Froehler, former senior quantitative analyst at Superfund Asset Management GmbH, said in his experience the failure rate with machine-learning algorithms in live tests is about 90 percent.
“There are an infinite number of potential ways you can fail,” said Froehler, founder of Quantiacs, a marketplace where investors can access systems based on machine learning. “That’s exactly what quants need to be able to stomach and overcome, and of course this can be hard for personal motivation.”
The three-year effort at Man AHL, a unit of Man Group Plc, to build a machine-learning strategy didn’t get off to a rollicking start. When Ledford initiated the research, his colleagues pushed back. They were skeptical of the technology, which over the decades hadn’t lived up to its hype.
“There was quite a lot of internal resistance to machine learning initially,” said Ledford, who leads strategic research at the unit.
Man AHL’s scientists and engineers developed and tested the technology on historical data, often expecting it to flop, and only overcame doubts as the strategy produced solid returns in trials. That prompted the hedge fund to eventually allocate hundreds of millions of dollars in client money to it.
“Only once we had seen these techniques perform in our very controlled environment of building research systems and then testing them rigorously did we have sufficient comfort to implement them in client portfolios,” Ledford said.
Trading 300 Markets
Greenig of Florin Court says a good trader will still usually outperform an algorithm in a single market. But the technology does have one major advantage: it can look for patterns across hundreds of markets while humans can manage only about a dozen positions at a time.
“That diversification gives you an edge,” said Greenig, who uses the technology to test the predictive value of technical patterns that people spot with their eyes. “If you are trading 300 markets around the clock, you can’t look at all these charts. Making things systematic allows you to test signals properly and apply the acceptable ones broadly and without emotion.”
Man AHL’s machine-learning strategy helps diversify the $4.9 billion AHL Dimension fund. It has returned an annualized 7.1 percent in three years through September, compared with a 3.2 percent gain for the average hedge fund, and an 11 percent rise for the S&P 500 Index. For the year, the fund was down 3.1 percent, and the average fund was up 4.2 percent.
“It’s not about applying machine learning techniques to historical data and being satisfied if we rediscover what we know already,” Ledford said. “It’s about discovering new things.”