Back in the 1990s, Paul Tudor Jones assigned a team of coders to a project dubbed "Paul in a Box." The effort sought to break down the DNA of the hedge fund manager's trading -- how he sizes up markets and generates ideas -- to train a computer to do the same.
The code created then was upgraded many times and is still used at his firm, Tudor Investment Corp. But it never took over.
Again and again, programmers had to feed in new types of data to mimic the changing price signals that Mr. Jones, famous for predicting the Black Monday market crash 30 years ago, zeroed in on, according to people with knowledge of the project. Even then, the machine couldn't capture intangibles like his gut instincts and conviction, as well as the market's uncertainties.
Ultimately, Mr. Jones remains the final decision-maker for trades -- not the box.
The limits of that model show why many jobs at the high end of finance are probably safe from automation a little longer. While machine-learning algorithms and other technologies are indeed encroaching on work performed by money managers, traders and analysts, many firms are still working out the kinks. Coders will be busy for years.
One not-so-well kept secret of Wall Street is that many companies rely on aging computers. Firms that weren't founded as algorithmic powerhouses often use a hodgepodge of trading platforms, Excel spreadsheets and data stored across servers that aren't in sync. Replacing staff with machines means automating roles that are idiosyncratic. And then there are executives, now in the prime of their careers, reluctant to usher in an era that no longer needs them.
In a sense, automation is ready to tackle Wall Street -- but not the other way around.
Even after coders overcome those technical issues, their software will need frequent fine-tuning. Michael Dubno, the chief architect of Goldman Sachs Group Inc.'s risk-management system known as SecDB, said that's one reason why salespeople and traders, at least for now, aren't obsolete.
"They have mental models of the world that are more complex perhaps than most of the computer systems," he said. In the short term, artificial intelligence isn't going to move as fast as people expect. "It will go through a number of fits and starts, where it will look like it's going to solve everything and then solve very little of it -- and then it's going to reset."
The finance industry is hiring coders to automate tasks. Algorithms are being taught to parse troves of data and identify patterns and relationships. The idea is to let machines calculate, for example, the odds that Apple Inc. stock will rise a certain amount in coming weeks, or suggest when to load up on gold amid geopolitical upheaval.
More than half of the 160 or so jobs that Goldman Sachs's securities division advertised online earlier this year were for tech workers. The bank told some recruits, for example, that they may build a chatbot that uses natural-language processing (which lets machines to comprehend human speech) and Bayesian inference (a statistical technique) to suggest trades, according to listings on its website.
Other firms have employees tinkering too. At billionaire Steven Cohen's family office, a secretive project has been experimenting with automating his best money managers.
But while money managers including Mr. Jones try to accelerate adoption of new tech, others are already far along. Renaissance Technologies, the quant fund founded by former military code-breaker Jim Simons, has been using machine-learning techniques for years, building a track record envied by the industry. Competitors relying on old-fashioned gut trading, meanwhile, have been suffering from lackluster returns and investor withdrawals.
The average age of software in finance is about 38 years, according to technology tracker CB Insights. And data, the bedrock upon which AI is built, are often fragmented or inaccurate.
"The No. 1 impediment is firms have legacy data systems, the wrong data and bad data," said Mr. Dubno, a former technology chief at Goldman Sachs and in Bank of America Corp.'s global markets business. "The standard plight for many firms is that they're stuck in a world where their data is stored all over the place."
Machine learning is valuable for tasks such as fraud detection and credit-risk analysis, but adapting it to trade in rapidly evolving markets is difficult, according to Alexey Loganchuk, who places data scientists at hedge funds. "By the time you have enough data to build a complex model, the dynamics you are trying to capture have often changed," he said.
It's hard to teach AI to take over sales, trading and investing roles, according to the Nomura Research Institute, which this year studied what it would take to apply natural-language processing to portfolio management.
"The workflow is irregular and ad hoc," said Yasuki Okai, president of NRI Holdings America. His company is partnering with tech firms to help solve problems. "Forecasting other market players' reactions to markets is an added complication."
Automating traders who handle illiquid securities, for example in credit markets, can be especially challenging because so much of their job hinges on judgment, the idiosyncrasies of each trade and human interaction. Traders use data that aren't standardized or work with clients to create bespoke contracts, such as for commercial mortgage-backed securities.
Billionaire Jones, 63, has experimented for decades with computers to automate parts of trading. One effort around 2011 called the Synergy Project brought his top money managers and groups of coders together over lunch to dissect how managers react to things like economic data and central bank announcements, according to the people with knowledge of the project, who asked not to be identified discussing internal initiatives. Patrick Clifford, a spokesman for Tudor, declined to comment.
To develop cutting-edge technologies, firms such as JPMorgan Chase & Co. and T. Rowe Price Group Inc. have created tech centres. But even when Wall Street innovates, implementation may not be the top priority, according to Adi Prakash, chief innovation officer at consulting firm Yerra Solutions. Firms are under constant pressure to reward their investors and keep up with regulations
"Inertia is the problem," said Mr. Prakash, who held various technology roles at Och-Ziff Capital Management Group and JPMorgan. "It's going to be a push to say that front-office processes are going to be automated within the next one to three years."
Not to mention the challenge of installing new systems on fast-paced trading floors.
"People are hugely time constrained," said Craig Butterworth, global head of Nomura Holdings Inc.'s client ecosystem, where he's leading technology initiatives across the trading business. "There is definitely an 'If it ain't broke, don't fix it' type of mindset." And sometimes, there isn't even space available on desktops for cutting-edge software.
Another obstacle: senior management.
Executives may try to preserve the status quo because they have too much at stake -- their income and status, former bank and hedge fund employees say. Bosses may be reluctant to displace large swaths of their staffs, reducing their authority, or to embrace technology that they themselves don't understand.
"Top management rarely want change, they want to keep intact a system that has worked for them for decades," said Mansi Singhal, a former trader at Brevan Howard Asset Management and Bank of America. "And it can also get political when you have lots of executives, silos and budgets, and there are managers who just don't want to cede control."
And then there are the current limitations of AI. Machine learning, according to Gartner Inc.'s hype cycle for new technologies, is at the "peak of inflated expectations" and heading to the "trough of disillusionment."
Nicholas Carr, author of The Glass Cage: How Our Computers Are Changing Us, predicts that while computers will assist people and help remove unwanted biases, they will stop short of replicating the elusive parts of human thinking -- like with Paul Tudor Jones.
"What's going on in your mind in many cases is taking place at a subconscious or intuitive level, and translating that into numbers, a perfect algorithm, will be hard to do," Mr. Carr said. "This dream of taking very subtle decision-making skills and turning them completely over to machines is at least premature and may not pan out at all."