Still no magic formula to understanding the markets
Uncertainties make it highly improbable for an algorithm to perform better than a fund manager
Had you asked John Bogle, the pioneer of index investing, about the merits of smart beta funds, he would have dismissed them as “just one more way of trying to do better”.
“Trying to do better” is what gets active fund managers up in the morning: the vain belief that their efforts to timely pick stocks will improve on the market return available at low cost from an index fund.
Smart beta investing — a form of active management dressed up as indexing — is just more of that.
The results show that sometimes they do better, but mostly they don’t. Invariably, it depends on whether their investment beliefs coincide with the prevailing market fashion or not. But there are no guarantees: notwithstanding the many strategies that would have reliably delivered “alpha” (market-beating returns) in the past, none promise to keep doing so in future.
That does not stop active managers from looking. Their quest for perpetual alpha brings to mind medieval alchemists who searched for an elixir of youth.
In the same way, active fund managers continue to search for a magic formula that will give them superiority over the market and the collective insight of all other investors combined. As far as we know, no one has yet succeeded. Even Warren Buffett, the greatest investor of the modern age, has lagged the S&P 500 over the past decade.
High fees, low success rates and low persistence levels — this is what ails the asset management industry and is driving an ever-growing number of investors to index funds.
Can artificial intelligence (AI) succeed where human intellect has met its match?
On the face of it, AI seems imminently suited to this pursuit. Professional investors must process huge amounts of data. It makes obvious sense to bring Big Data, and the science and analytical power of AI to this field.
Some forms of AI are already evident today in basic investment processes such as portfolio rebalancing, quantitative analysis and exploiting arbitrage opportunities. But these merely replicate human actions and mimic human intellect.
Picking stocks and timing markets — as a fund manager would do — requires more than just massive data crunching. It also requires imagination, intuition, the ability to draw on experience, to evolve and adapt to different market conditions, and to anticipate the behavioural aspects of investing.
Some such advanced AI operators are already active, but much like the alchemist of old, their activities are shrouded in mystery and secrecy. We don’t know how far they have moved beyond extreme quantitative analysis, and with what success.
But let’s imagine what is possible by looking at AI in another sphere altogether: the ancient Chinese board game of Go.
This is a game of abstract strategy. It’s considered to be the most complex in the world due to the vast number of potential variations. Estimates suggest the number of possible games far exceeds the number of atoms in the observable universe.
For years it was unimaginable that a computer program could match the skill of a professional Go player on a full-size board, without handicaps.
Computer-game algorithms, such as for chess, compute several moves in advance. In Go, players have an average of 200 available moves per turn. For a computer to weigh just the next two possible moves of both players, it would have to consider more than 320-billion possible combinations. Anticipating the next four moves would involve 512-quintillion combinations (5.12 x 10 to the power 20).
The required computing power is immense. In 2014, the world’s most powerful supercomputer, assessing those next four moves, would have required four hours to make just one play. It seemed years off before software would undo a professional Go player.
Yet one year later, the serendipitously named programme AlphaGo beat a 9-dan ranked professional and, in May 2017, its successor, AlphaGo Master, beat Ke Jie, the world’s number one, three games out of three.
This turnaround had nothing to do with increased computing power. What changed was that AlphaGo, developed by Google DeepMind, applied a fundamentally different paradigm than earlier Go programs. This included very little “direct” instruction and, instead, relied mainly on deep learning that involved AlphaGo playing itself in hundreds of millions of games such that it could measure positions more intuitively.
In October 2017, DeepMind introduced AlphaGo Zero, a version without any human data inputs at all and which learned simply by playing games against itself. It quickly surpassed the human level of play and defeated the previous version of AlphaGo by 100 games to 0.
This announcement suddenly made it far more credible that machines would eventually be able to develop their own artificial general intelligence.
The bottom line is, we don’t know, and we lack the imagination to appreciate what AI methods may one day achieve. If one algorithm can overcome the world’s best Go players, then machines outsmarting the market is surely not so far-fetched either.
But then again, perhaps it is. Go is a two-person game that provides perfect information and is rules-based. In game theory terms, it is a deterministic system, which means there is no randomness in future states of the system — a given starting point or initial state will always produce the same output.
The financial markets are none of these things. They accommodate an unlimited number of participants. There are no rules or laws that govern price moves, or how markets should behave at every turn. All we have is theories, such as the capital asset pricing model or arbitrage pricing theory, that attempt to explain what the markets are doing.
No individual investor ever has perfect information, not even with Big Data capabilities. Market movements are random. And the market’s initial state changes constantly and never repeats.
There is also no way to reduce human behaviour to equations. Here, too, all we have are economic theories such as the notion of free will and that we consider all available options and choose the best. But real life is very different. Our decisions are often irrational or based on an imperfect analysis of the potential outcome.
It seems highly improbable that an investment algorithm will learn to model all these uncertainties in order to predict future outcomes. Unlike Go, there is no evidence that human investors become more skillful (more likely to beat the market) as they gain experience, which suggests this is not a skill that can be learned.
We may one day hear of some algorithm that has beaten the market, but we won’t hear about the many others that didn’t. And one thing we can be sure of is that none of these AI strategies will guarantee their future outperformance.
Where AI investing — in whatever form — does have the edge, it is in the absence of emotion. That’s a huge advantage over our decision-making.
Of course, your index fund is also devoid of human emotion. If that proves to be the ultimate competitive edge of AI investing, then this seems like a highly convoluted way to find something that’s been around since the ’70s.
• Eddy is head of investments at 10X Investments, a money manager with more R11.5bn assets under management