A Robot Wants Your Job

Picture a New York City street circa 1900, full of horse-drawn carts and carriages, carrying passengers and cargo. In the background, a few pedestrians are looking warily at a strange new contraption barely visible behind the crowds of equine transport.

It’s an automobile.

Imagine the same street, just 13 years later. Four lanes of cars traveling in either direction stretch off into the distance. To the right of the frame is a cart pulled by a single, perplexed-looking horse. The street—and the world—has changed forever.

Anders Hjælmsø Svennesen, CIO at the DKK 327 billion ($50 billion) Danica Pension in Denmark, uses these images in presentations to illustrate the speed at which technology can fundamentally change the way we live.

“We have computers giving us inputs today into the investment process, trying to figure out trends in economies or companies,” Svennesen tells me. “The question is: When will this technology be so good that it can be the decision maker? That’s when it can more or less substitute investment professionals.”

But Svennesen is not about to fire his investment team and replace them with cyborgs. That’s all science fiction stuff, anyway.

Isn’t it?

In March this year, a little-known, Google-owned computer program achieved what many people had thought impossible: It won a board game. To be clear, this is way beyond IBM’s purpose-built Deep Blue, which beat Grandmaster Garry Kasparov at chess back in 1997. Google DeepMind’s AlphaGo essentially taught itself to play the most complex board game humans have devised—Go—and destroyed world champion Lee Sedol, one of the best-ranked human players, four games to one.

Using the same technology—known as ‘neural networks’—DeepMind’s algorithms have learned to play several video games far better than any human. Several YouTube videos demonstrate its gradual development of strategies over the course of hundreds of attempts. DeepMind is not taught the rules; it is simply plugged in and told to get the highest score possible. It learns the rules as it goes, refines its technique, and beats the game. It is real-life artificial intelligence.

Demis Hassabis, co-founder of DeepMind, believes this kind of technology can have a similar impact on the 21st century as the car had on the 20th. Speaking at Oxford University prior to AlphaGo’s match with Lee, he told attendees that financial services were just one arena in which machine learning can have a profound impact. With financial data, Hassabis said, “if there is structure in them at all then the system should be able to find it, whether or not it is intuitive to the human brain.”

Bridgewater Associates was reported last year to have started work on an artificial intelligence (AI) unit, and founder Ray Dalio has spoken publicly about his use of AI tools. Perfect, I think. Without saying “no” outright, the PR agency expertly suggests another client.

Aidyia, based in Hong Kong, is led by Dr. Ben Goertzel and employs “artificial general intelligence”—a similar concept to DeepMind’s machine learning processes—“to identify patterns and predict price movements.” No response.

Two Sigma, Renaissance Technologies—similarly, nothing. Perhaps the investment industry’s experiments with AI have gone rogue and are refusing to cooperate? “I need your clothes, your boots, and your motorcycle. Oh, and your latest liability study results.”

Art by Jun CenThen, success. Rebellion Research CEO Alexander Fleiss answers his phone. I ask him if AI could really have a long-term impact on asset managers and owners. For Fleiss, there is no “if”: Machine learning is changing financial services now and, more importantly, is making money for his investors.

Rebellion Research, based in New York City, employs a learning algorithm to monitor hundreds of economic factors, recommend stock picks, and make economic forecasts that drive the decisions made by the firm’s trading team. Fleiss claims this system flagged up the Greek debt crisis and the 2014 oil price crash well before many humans saw it coming.

“Without a doubt our system is a better than the average stock picker, but it’s an amazing economic forecaster,” says a bullish Fleiss. “I think we’ve got the best economic forecaster on the planet.”

“Take palladium,” Fleiss continues. “It’s used for catalytic converters for gasoline engines. The US and China are dominant gasoline markets and will push those prices. We were very public about that in 2012 when the US and Chinese auto markets were heating up at the same time, and we thought palladium made very good sense. It was a very profitable trade for this firm.”

Many human fund managers claim to have seen similar patterns or links between two or three factors, leading them ultimately to a buy or sell decision. Rebellion’s system, Fleiss explains, considers more than 150 factors and crunches data on performance and correlations across different markets and geographies. If a decision or forecast doesn’t work out, the system analyzes it and attempts to learn why.

All of a sudden, the idea of Anders Hjælmsø Svennesen—or any CIO for that matter—hiring an algorithm to his investment team seems a step closer.

“This type of innovation will affect the way we operate our business,” Svennesen says. “With big data, you have the power to use a lot of statistical models, and you can get so much information that you hadn’t been able to get before. It’s just which direction it might affect.”

By this point my inner nerd is far too excited. “Isn’t the job of a CIO basically all numbers anyway?” I ask myself. Let’s start with liabilities. Big data on population health trends could easily result in far more accurate longevity predictions. Parameters agreed on by an investment committee could set the risk and investment budget. These inputs could then be fed into an algorithm, which selects the optimum asset allocation based on volatility, correlations, costs, long-term megatrends, demographic shifts, and economic indicators. Manager performance could be monitored according to pre-set performance guidelines… Hold on. Could a robot CIO be a thing?

“I think that’s even feasible now,” says Professor Andrew Lo of the Massachusetts Institute of Technology. “I don’t think we’re that far away from it.”

A few years ago, Lo’s team was asked to explore the benefits of machine-learning tools within a commercial bank’s credit card business. The team’s “recognition algorithms” processed 10 terabytes of data from half a million customers over five years. “We were able to identify consumers who were likely to default or be delinquent on their credit card payments much more accurately than traditional tools,” Lo says. “By addressing those particular hotspots in their credit card business, the bank was able to reduce the risk of the overall business but only by affecting 3% of its customers, as opposed to making broad credit reductions for the entire customer base.”

So will future credit analysts come in software boxes?

“Absolutely,” Lo says. “I think there are a number of large data sets that could be mined to provide that kind of information, and I’m sure that it’s being done even as we speak.”

“The only way an automated system would work is with full buy-in from all stakeholders.”In 2013, Carl Benedikt Frey and Michael Osborne of Oxford University published an analysis of more than 700 jobs across all sectors to assess the likelihood of automation displacing humans in each role. For credit analysts, the probability is 98%.

Fortunately, Chetan Ghosh, CIO of Centrica’s pension schemes in the UK, is on hand to provide a voice of sanity over a cup of coffee. Would he hire an algorithm?

“It’s dangerous to rely on assumptions” from backtested data, Ghosh warns. Even if such assumptions are accurate, many do not last over the long term, he argues. “There would be times when it will work well for you, and times when it won’t.”

As Professor Lo points out, for all the advantages of Moore’s Law—the observation that computer processing power roughly doubles every two years—we must not forget Murphy’s Law: Anything that can go wrong, will go wrong. Regulators across the world need to be on top of the “important systemic implications” of the latest technological developments, Lo says.

“The real question is whether or not we are willing to relinquish control to an algorithm,” he adds. “It’s very much the same situation with driverless cars. According to Google, we have the technology today to have a driverless car. The question is, do we have the legal, social, and political will to be able to implement that on a large societal basis? I think the same issues apply to all kinds of automation, including automated financial decision making. I think it’s going to be a long time before humans delegate all of their control to algorithms and machines.”

Centrica’s Ghosh agrees. “The only way an automated system would work is with full buy-in from all stakeholders: The investment committee, the trustees, the sponsor, and the members.”

There is another human element to counter the rise of the machines, and that is the importance of trust. “We spend a lot of time trying to optimize the work of our asset managers,” Ghosh says. The Centrica team conducts regular discussions with managers about “how best to use the ‘first principles’ thesis for investing”—i.e. not making assumptions based on rigid models.

The UK corporate bond market, for example, is “very inefficient,” so Ghosh will discuss with the relevant manager the removal of securities in the index and the possible inclusion of off-benchmark, less-liquid securities. “These are real value-add discussions,” the CIO says. “They solve problems for the investment manager as well as enhance the potential outcome.”

Social interaction is one obstacle with which even the most advanced algorithms will struggle. [Insert your own joke about an actuary here.]

Developers of machine-learning technology, however, are convinced that they are on the verge of the next major leap forward in our society. DeepMind’s Hassabis likens his firm’s work to NASA’s Apollo missions, and states his intention that humans and artificially intelligent computers can “work together to make the world a better place.”

“I know where machine learning will go but it will take so long to get there. From a technological, mathematical standpoint it’s depressing how early we are.”That’s rather sweet, but manned moon missions only lasted three years. Rebellion Research’s Fleiss is far more outspoken.

“It’s like 1497: We know the New World exists, we know we can get there, but it’s not consistent,” says Fleiss. “Five years after Christopher Columbus sailed the Atlantic, you had lots of people going over. If you were a Spaniard in the year 1500 I’m sure life was nice, but imagine how much more exciting the New World is now, 500 years later. That’s how I feel about machine learning. I know where it will go but it will take so long to get there. From a technological, mathematical standpoint it’s depressing how early we are.”

Robotics and artificial intelligence technology in financial and legal services will likely affect 25 million workers in the next decade, according to the McKinsey Global Institute. Such advances could generate productivity gains of between 45% and 55%, the research group claims.

Frey and Osborne’s paper doesn’t give a specific timeline, but indicates that several other elements of an investment team may be subject to automation in the years and decades ahead. Accountants (94% probability), auditors (94%), and real estate managers (81%) should all be concerned, the research shows.

(Looking on the bright side, Frey and Osborne also calculated that economists have a 43% probability of being replaced by computers; tax collectors, 93%. Every cloud has a silver lining.)

Time to polish your résumés, bond analysts. Perhaps you can make use of that engineering degree after all—unless the robots have got there first, of course.

“I don’t think in 10 years from now I will have the same business model as I have today,” concludes Danica’s Svennesen. “I think our customers will ask for other products because there will be a big change in the available technology.”

In his acclaimed 2011 novel The Fear Index, British author Robert Harris introduces the algorithm-driven, artificially intelligent hedge fund VIXAL-4. Created by the book’s protagonist, Dr. Alex Hoffmann, the fund makes hundreds of millions of dollars for its elite group of clients. It then becomes self-aware and kills people. “Preposterous,” as one reader succinctly described it to me.

Early on in the book, Dr. Hoffmann relays to his investors the power of his creation:

“Over the past couple of years a whole new galaxy of information has come within our reach. Pretty soon all the information in the world—every tiny scrap of knowledge that humans possess, every little thought we’ve ever had that’s been considered worth preserving over thousands of years—all of it will be available digitally. Every road on earth has been mapped. Every building photographed. Everywhere we humans go, whatever we buy, whatever websites we look at, we leave a digital trail as clear as slug slime. And this data can be read, searched, and analyzed by computers and value extracted from it in ways we cannot even begin to conceive.”

Suddenly, this doesn’t seem so preposterous—except, hopefully, for the murderous computer.