Here is a must read from the WSJ on the pitfalls of automated investing. James Mackintosh explains the pitfalls of algorithmic trading and the problems with artificial intelligence. He identifies three serious problems with using AI for investment.
Ten years ago, computer-driven traders pulled the plug after their algorithms ran amok, leading to billions in losses and the eventual closure of Goldman Sachs ’s flagship quantitative fund.
A decade on, artificial intelligence and machine learning are the buzzwords in automated investment. But for all the hype, applying AI to investment has three serious problems: it works too well, it is often impossible to understand, and it only knows about recent history. Worse, it will be self-defeating if it proves popular, as algorithms face off against each other in the market.
Machine-learning systems are now really good at spotting patterns. Unfortunately, computers are just too good, and frequently find patterns that aren’t really there.
Michael Kollo, chief strategist at Axa IM Rosenberg Equities, points to the neural network—a type of AI loosely modeled on brains—developed by three University of Washington researchers to distinguish between pictures of wolves and dogs by associating wolves with snow.
“It can easily identify something of an intransigent nature and learn one rule from it,” he says. Train an AI on the last 35 years of markets, and it might well develop a single simple rule: buy bonds. With 10-year Treasury yields down from 13.7% in July 1982 to 2.31% on Monday, it worked beautifully in hindsight—but yields can’t possibly fall that much again in the next 35 years.
In the industry, spotting patterns that don’t repeat is known as “overfitting”—picking up on the irrelevant snow in the picture of a wolf, or chance patterns in past stock prices that bear no relation to the future.
Read more here.
