Capital markets players, just like those in every other field, are head over heels infatuated with artificial intelligence. The prospect of using AI to speed up all kinds of data crunching and administrative tasks — and even to improve the results — is inspiring avid enthusiasm.
Bankers, investors and service providers relish the prospect of cutting costs — i.e. more junior employees’ jobs — apparently blithely unconcerned that their own skills could ever be bested by a bot.
This confidence seems ill founded, considering the breakneck speed at which generative AI is accelerating, and the reluctance of most governments to try and control its effects. The European Union with its AI Act is an honourable if not necessarily effective exception.
But can AI actually solve real problems facing markets, rather than just replacing human work with a glib, mushy but cheaper simulacrum?
Better call Hal
Specialists have argued that AI tools can beat humans at tasks central to investing, such as analysing company results and investor conference calls. They can spot patterns and predict the movements of stocks more reliably, some evidence suggests. Moreover, they are immune to the emotional biases that send human traders awry.
Financial professionals may recall with a sinking feeling that in high intellect fields from playing chess to flying aeroplanes to spotting abnormalities on X-rays, machines overtook humans many years ago.
But what separates those fields from financial markets is that they are closed systems. A computer analysing scientific data or the winds and engine performance that govern aircraft flight is looking in through a sealed window. The observation does not affect what is studied.
In chess the computer interacts with its opponent, but the game cannot stray outside its bounds. Both players follow the rules and one wins.
It would be very different if robots began to play football — as indeed, in experiments, they have. There would be nothing to prevent the machines winning 1,000-nil.
Free form
Financial markets are interactive and have open outcomes like that, but far more so. Not only are gains and losses theoretically unlimited. More importantly, all players adapt to how the others play.
In football, try as they might, other players cannot be Lionel Messi. In finance, all of them can copy Warren Buffett. But when they do, Buffett will try something else. The goalposts will move.
Ever since investors started to apply systematic maths to capital markets, the dream has existed: can someone devise a system to consistently beat the markets?
Some have done very well at it. But everyone knows that, as soon as this system became widely adopted, it would cease to work. The arbitrage it had found would disappear, or the cheap assets it found would become expensive.
Markets are inherently machines for generating uncertainty. As soon as certainty appears, it destroys itself.
This is even more clear when you remember that markets are not a game or mathematical system at all, but real life — part of society. Excess wins in one area — like the profits made from securitizing subprime mortgages pre-2008 — end up impacting real people until there is not only a market but an economic and social reaction.
January sales
With all this in mind, what could AI contribute to solving an apparently simple and basic problem familiar to all in European bond markets: the extreme concentration of issuance in the first weeks of the year?
When a dozen major public sector issuers and half a dozen banks try to cram deals into a single day, it seems about as beneficial as the crush at airports during school holidays. Airlines inevitably take advantage by putting up prices.
In the bond market, it’s true investors have built up a bit of cash during the quiet Christmas period when there was nothing to buy. But not that much. Much of the January rush is attributable to herd behaviour and other psychological factors.
Surely it would be better if deals were more evenly spaced through the year. Couldn't an AI app, like those that read X-rays, analyse data on investor cash balances better than bankers and issuers flying by the seats of their pants? Cities use computers, after all, to control traffic flows.
But as soon as one contemplates this idea, it crumbles. The main reason why issuers like to sell bonds early in the year is risk aversion. They don’t know what’s going to happen in markets and they don’t want to be caught out if there is a crash in February, with all of their funding still to do.
Funding officials are willing to stomach possibly paying a bit more than optimal, to take risk off the table.
Now suppose BNP Paribas or Deutsche Bank came to them saying “our AI can predict when it would be best for you to issue bonds”.
Would treasurers trust it? Hardly. However good the app, it can’t be prescient. The model may have been back-tested and work 99% of the time. But this year might be different. If that crash comes, no one is going to forgive the underfunded treasurer for following an AI.
Financial decisions are always going to be about weighing odds about the unknowable. You can use computers for that when your risk is well diversified, so you can afford to be wrong some of the time. Even then, the model will only stay right for as long as it is little used.
But for big, strategic decisions with binary outcomes, where your job is on the line and the human consequences for others must be reckoned, the emotional tilts that models try to eliminate are not a bug. They are the surface where markets connect with reality.