When I asked Tim McCusker, CIO at consulting firm NEPC, about active managers’ increasing use of artificial intelligence, he paraphrased the Roman historian Suetonius: “AI investing is not going away.”
The evidence is on McCusker’s side. According to Business Insider, at a recent J.P. Morgan conference, the bank asked 237 investors about big data and machine learning and found that “70% thought that the importance of these tools will gradually grow for all investors. A further 23% said they expected a revolution, with rapid changes to the investment landscape.”
Such investor interest signals both the frustration with current active — and specifically quant — managers and the nascent promise shown by AI hedge funds.
Whatever the cause, AI investing presents consultants and asset owners with a serious challenge.
Our industry takes as a universal truth that investment processes should be understandable. As part of the kabuki theater we call investment due diligence, asset owners and consultants require that a manager be able to explain its strategy and model.
A manager reveals just enough of its investment process to provide allocators with a cairn from which they can then orient themselves and continue their assessment.
We must recognize that a traditional manager’s degree of disclosure reflects a willful act: The manager could reveal more but chooses not to because, it claims, doing so would put its process at risk. It’s probably more truthful that sharing too much would reveal the paucity of the process itself.
However, though an AI manager can provide a general overview of its approach (“We use a recursive neural network”), it can provide no authentic narrative of its investment process — not because of willful deflection but because, unlike a traditional manager, it has not hand-built its investment model. The model built itself, and the manager cannot fully explain that model’s investment decisions.
Think of a traditional manager as Deep Blue, a human-designed program that used such preselected techniques as decision trees and if/then statements to defeat chess grandmaster Garry Kasparov in 1997. Think of an AI manager as DeepMind’s AlphaGo, which used deep learning to beat some of the world’s best Go players. (Go is an ancient Chinese board game that is much more complex than chess and has more possible moves than the total number of atoms in the visible universe.) Without explicit human programming, AlphaGo created its own model that allows it to make decisions better than its human opponents.
With enough time and training, we can explain why Deep Blue made a certain chess move at a certain time. Although we can observe how AlphaGo plays Go, we cannot explain why it made a specific move at a specific point in time. As Yoshua Bengio, a pioneer of deep-learning research, describes it: “As soon as you have a complicated enough machine, it becomes almost impossible to completely explain what it does.”
This is why an AI manager cannot explain its investment process. The requirement of interpretability brings the assessment of — and by extension, the investment in — AI strategies to a screeching halt.
With AI investing, allocators face a new choice. Currently, in an act of complicity, they choose access over knowledge — accepting a manager’s willfully limited disclosure of its narrative but naively believing that the narrative does exist and is known to the manager’s illuminati.
The new choice facing all AI consumers is more fundamental. The choice, according to Aaron M. Bornstein, a researcher at the Princeton Neuroscience Institute, is, “Would we want to know what will happen with high accuracy, or why something will happen, at the expense of accuracy?”
Requiring interpretability of investment strategies is a vestige of old-world assumptions and is entirely unsatisfactory for reasons that transcend investing: We either foreswear certain types of knowledge (e.g., deep learning–generated medical diagnoses) or force such knowledge into conformity, thereby lessening its discovered truths (do we really want our smart cars to be less smart or our investment strategies to be less powerful?). Moreover, this requirement smacks of hypocrisy: Given what Erik Carleton of Textron calls “the often flimsy explanations” of traditional active managers, investors really don’t know how their money is invested. And conversely, who would not suspend this criterion given the opportunity to invest in the Medallion Fund?
We need to better invest beneficial assets. AI investing can help, but its adoption compels us to judge AI strategies not by their degree of interpretability but by their results.
As scientist Selmer Bringsjord puts it, “We are heading into a black future, full of black boxes.” Embrace it.