For Michael Chrytonov, setting up a hedge fund based on machine learning was an unusual challenge: it was three times more difficult, and took three times more time than he expected. “Most of the things we’ve tried have failed,” says the founder of Voleon Group of Berkeley, California.Machine learning, an array of techniques that allows computers to detect patterns in data without applying pre-written laws by humans, produces advances in a variety of fields, from robotics to weather forecasting and automatic translation. It’s also a technique behind an autonomous car.The sense that machine learning could be cracked with capital markets led to an “arms race” in which investment companies that were already focused on mathematics began to recruit the smartest computer engineers and statisticians they could find.This bet apparently succeeded in two of the best-performing hedge funds this year. Quantitative Investment’s largest fund rose 68% this year, it says, thanks to machine learning. Teza Capital Management recorded a similar increase of over 50%.Still, the cases where machine learning becomes a continuing success in investments are rare. One can understand the difficulties when examining the difficulties of Voleon, one of the first investment companies to decide to go big on this field, which has been going on for many years.
One of the company’s early lessons was that machine learning might not be suitable for trading in financial assets, an area where patterns are not always clear.The idea that we can simply take machine learning techniques in speech recognition and computer vision to make better predictions simply does not work,” said Chrytunov, a computer scientist who emigrated from Russia to the US at the age of 18.Institutional investorsHaritunov, 54, and his second co-founder, John McAuliffe, 43, have a PhD in computer science and statistics, respectively. They both went into finance and became researchers in the D.E. group. Shaw, one of the oldest and most successful “hedge fund” hedge funds. Chrytunov was at some point subordinate to Jeff Bezos, before he left and set up Amazon.For several years, Haritunov and McAuliffe believed that machine learning tools, both of which they had learned, were too crude for investment. The methodology was serious, but the computers were not fast enough, and there were no data sets large enough to scan.In contrast to the more common case, in which the scientist has a note and writes an algorithm to the computer to perform, in learning the machine, the scientist sows large quantities of data in the computer, and asks the computer to find patterns on its own. The computer actually writes its algorithm and uses it to make predictions.