Our readers may recall Bridgewater building an algorithmic Ray Dalio brain. If not, links after the jump.*
The writer, Aaron Brown is a former head of financial market research at AQR Capital Management.
From Bloomberg Opinion, December 6:
The AQR Capital co-founder says artificial intelligence is coming for his job.
Hedge fund executive Cliff Asness
says artificial intelligence is becoming “annoyingly better” at doing
parts of his job. AI deployed by his firm AQR Capital Management, where I
worked for a decade, is combining investment factors to build
market-beating portfolios, something that used to be Asness’ specialty.
“AI’s coming for me now,” he told Bloomberg Television in a recent interview.
Almost exactly seven years ago, Asness had expressed skepticism
to the Financial Times, saying that big data and machine learning were
dangerous because they found too many spurious patterns, and even
genuine patterns were quickly competed away in the markets. However,
like a good portfolio manager, he hedged his bets, saying, “We’re
feeling our way. If our first few experiments bear fruit, we’ll do more
of them. If we find out we’re good at this, it will become a bigger part
of AQR.”
Going back a further seven years to 2010, I recall the early enthusiasm
for AI in quantitative investing. Breakthroughs in AI algorithms and
improvements in computer processing caused a gold rush mentality among
many investment managers, and quants seemed well positioned to be the
first to the motherlode as they had the training and skills to
understand and apply AI. But initial results were disappointing — not
terrible, just not the kind of improvements that technophiles and
science fiction fans had hoped for.
But the last seven years — conventionally dated to the 2017 publication of “Attention Is All You Need”
by Alphabet Inc. researchers — have changed the picture dramatically,
and the dream of fully self-driving portfolios seems within reach.
There are three main steps in quant investing: identifying factors
such as value and momentum that predict future returns; combining
signals from those factors into optimal portfolios; and executing trades
to keep the actual portfolio optimally close to the optimal one. From
2010 to 2017, AI proved unsatisfactory at identifying factors or
combining signals. It was helpful at trading, but with the sort of
improvements we got from standard methods, not a quantum leap.
In
the last seven years, from 2017 to 2024, as Asness said, AI has finally
begun to pay off in factor identification and portfolio construction.
But that’s not the big story. The most exciting advance on the horizon,
the one that might finally realize the science fiction dream, is AI
researchers have figured out how different types of AI systems can
communicate. The synergy from combining three different AI systems
should be much greater than the individual advantages in each step.
For
example, the trading AI should be sending news about soft prices or
erratic volumes in securities back to the factor-identification and
portfolio-construction AIs, which in turn should be sending insights to
each other and the trading AI. This was one of the key advantages of
old-style hedge fund managers who did their own research, portfolio
construction and trading. But in large quant hedge funds today, those
tasks are done by different specialists.
We’re
not quite at the point of breaking the curse of the Tower of Babel.
Although different types of AI systems can communicate, it’s still more
of a pidgin sign language than a fluent and precise exchange of
information.
And
even the best AI systems cannot rid themselves of occasional gross
errors — the kind no human would make — and even one of these can
destroy the effectiveness of the entire process. That’s why we still
have humans examining the output of each AI system before feeding it as
input to the downstream system (so from the factor-identification to the
portfolio-management system) and why upstream communication (from the trading system, say, to the factor-identification system) is restricted or forbidden.
But
these problems now seem solvable with some algorithm tuning and more
computer power, issues for workaday quants rather than anything
requiring breakthroughs by geniuses.
In
addition to the three steps above for the core investment process, in
2010 we hoped for three indirect aids: AI coding, monitoring and
explaining. Coding was the first home-run success of AI in quant
finance; no one has to code anymore, AI does it much better, faster and
cheaper. But we had no success at monitoring tasks — such as having the
AI issue alerts that some factor was not behaving as expected, or some
data looked fishy; and none at explaining events — for example, why some
price relation in the market had broken down. I’m seeing progress in
these areas as well, although slower than in core quant investing.
If all this works, we may see the end of the investment management
business as we know it. Instead, each person could have a personal AI
that combined personal information — tax situation, financial goals,
income prospects — with superior market knowledge, and that never slept
or had its attention wander. These AIs could trade with each other with
no need for human intervention....
*And Ray?