AI Conquered Chess. Investing Could Be Next.

view original post

The deans of wealth management theory, Nobel Prize winners Harry Markowitz and Robert Merton, decades ago defined the challenge of investment planning as one of dealing with uncertainty. 

Especially relevant for financial advisors today, Markowitz wrote in 1959 of securities analysis: “Only the clairvoyant could hope to predict with certainty” what may happen with a given investment. His advice was to incorporate the understanding of risk into building a portfolio. Similarly, Merton wrote in 1971 that the “investor does not know the true value of the [expected return]” for any investment and can only choose their individual appetite for such uncertainty. 

An AI approach called reinforcement learning says if you can specify the end goal of a problem, such as getting an opponent in checkmate, you can work backward to calculate the series of moves that will most likely lead to the goal. 

Fizkes/Dreamstime

But what if you were less uncertain? What if you could more precisely identify where uncertainty lies and work around it creatively? Would you make different choices?

A modern cohort of scholars, using artificial intelligence (AI) techniques, aim to redefine uncertainty. And they’re flipping the old portfolio reasoning on its head. Rather than merely setting a level of risk tolerance for investment, these AI theorists argue one should set a goal and then work backward, calculating with some precision which steps along the path to that goal are more or less certain.

It’s an approach that likely won’t gain wide favor for years to come, but in a decade it could reshape investment planning. Advisors would do well to keep their eye on this AI movement. 

The new generation of portfolio theory is a bit like playing chess. It borrows, in fact, from machine-learning approaches that have conquered chess. 

In 2017, scholars at Google’s DeepMind division showed they could beat the human grandmasters of chess as well as the masters of the ancient strategy game Go, with a neural network program that took only hours to advance from novice to unparalleled mastery by playing thousands of games.

At the heart of the DeepMind program was a broad AI approach called reinforcement learning, known as “RL.” The RL approach says if you can specify the end goal of a problem, such as getting an opponent in checkmate, you can work backward to calculate the series of moves that will most likely lead to the goal. 

The key is that modern computer horsepower can calculate probabilities at every turn in a game of chess or Go with far greater precision than could a person or even previous statistical computer models. Now turned to the world of investing, the same calculation of uncertainty can be applied to every moment of investment choice in a path to retirement, a level of calculation that was unthinkable in Merton’s time.

The new AI efforts begin by simply improving a bit on the standard approaches formulated in portfolio theory based on Markowitz’s and Merton’s work. For example, University of Illinois Associate Professor of Applied Mathematics Matthew Dixon in 2020 introduced an RL approach with collaborator Igor Halperin of New York University that sets a goal, such as target wealth at retirement. 

The program then decides, for something like a defined-benefit pension plan, at each moment in time what the optimal cash contribution and the optimal asset allocation are based on how that moment in time will contribute down the road to the ultimate payout. Unlike Merton’s approach of defining a single “utility function” that is supposed to maximize the payouts over the entire lifetime of investments, the RL program chooses new strategies, and tactics, at each moment in time in relation to perceived uncertainty at that stage.

The program has implications for robo-advisors because Dixon and Halperin are able to invert the RL approach and ask: If certain steps are taken today, what future financial rewards, previously unknown, might  result? 

They write that the program, called GIRL (for G-Learning for Inverse Reinforcement Learning) “would then be able to imitate the best human investors, and thus could be offered as a robo-advising service to clients that would allow them to perform on par with best performers among all investors.”

The Dixon and Halperin approach makes use of relatively simple mathematical tools that are easy for computers to run, even for very large portfolios. However, more recent work taps much more ambitious AI techniques. 

In research published in March, Wing Fung Chong and colleagues at Heriot-Watt University in Edinburgh, Scotland, take a novel approach to variable annuities using what’s called deep learning, a form of machine learning within AI that builds much larger combinations of artificial neurons. 

The challenge Chong and colleagues confront is that RL, by its nature, experiments with choices, seeing which ones lead to better or worse outcomes. For an insurer writing a variable annuity policy, such experiments could produce catastrophic losses.

Their solution is a two-stage neural network. The program first practices on simulated markets based on the insurer’s historical data. Once the program can hedge as well as established hedging strategies, it’s let loose to make choices in a live market, where the program refines its hedging strategy with each new choice. 

What results is some automation of investment choices. “The further trained RL agent,” write Chong and colleagues, “is indeed able to self-revise the hedging strategy.”

These RL programs for annuities are still at the R&D stage for a couple of reasons. For one thing, they have yet to be developed for the broad swathe of investment considerations pertaining to the age range of clients, the range of survival possibilities, the diversity of portfolios, and the variety of contracts that a given wealth manager has to construct. 

More important, it will take time to figure out how AI programs square with humans’ intuitive sense of risk. Programs that use AI are the proverbial black box, which means they can be dazzling and disturbing.

That led world chess champion Gary Kasparov to write of AlphaZero that its style of chess play “reflects the truth” of the game, but also that the program “[prefers] positions that to my eye looked risky and aggressive,” moves he wouldn’t have made himself.

Hence, advisors in years to come will have to find a way to talk with clients about such programs so that the alien approach to investment, however efficient and effective, doesn’t itself become a new source of uncertainty that confuses and puts off clients. 

Write to advisor.editors@barrons.com