By Mark Haraburda, CEO, Barchart
JP Morgan has been testing whether an AI agent can be trusted to move money between stocks and bonds on its own. In two decades of backtests, all eight of its agents beat the classic 60/40 portfolio on a risk-adjusted basis. The bank warns the results are simulations, not proof. True: a backtest is a wager on the recorded past, and it is only as good as the record. What strikes me is that the record exists at all. When I entered this business, you could not see every trade in a market unless you were standing in the pit where it traded. Access to data was geography, and the geography was one trading floor.
I have been in market data since 1999: first at the Chicago Board of Trade, licensing exchange data to distributors, and since 2007 at Barchart. Part of the Board of Trade job was replacing the Zip drives and CD-ROMs we shipped price history on with a web service. Within living memory, the market’s past traveled by mail. And I worked in the data department of an exchange whose own floor kept no complete record: no report of every bid and ask, no view of the order book up and down the price ladder, no record of every trade. The information existed for the people standing in the pit, and then most of it evaporated.
Starting around 2000, exchanges demutualized and went public, and electronic trading took off. The transition took about a decade, and it multiplied the data: every quote and every fill now left a record. All of it demanded bigger telecommunication lines, bigger servers, better software to analyze it, and that is really when cloud computing came into the picture. A decade ago the story everywhere was big data, big data, big data. Today, it’s AI. And the software layer assembled over a quarter century — the APIs and the delivery infrastructure — turned out to be exactly what AI models needed to work at all. It has been an evolution, each stage feeding the next.
This is where I am supposed to say AI changed everything. And, indeed, much has changed, and it keeps changing faster. But AI did not change the thing at the center. A data point is a data point; a price is a price. My analogy is Uber. It did not invent the car or the fare; it changed the delivery and the packaging, and suddenly anyone could summon a ride. AI is data’s version of that event. It changes how much you can get out of facts you already had: quicker, deeper, cheaper. In roughly the past year, those gains have run beyond our wildest dreams.
What did change is access. Analysis that once required a quant desk now runs off a question you type out for an AI to run with. We watch customers write their own dashboards and custom analytics against data APIs, requests that used to mean an Excel integration. Retail widened it further. Smartphones put brokerage apps like Robinhood in every pocket, commissions went to zero, at least in equities, and a wave of new participants generated data sets that did not exist before. What the floor once rationed by geography is now rationed by little more than curiosity.
Appetite followed. AI is great, but if you don’t have the data, you can’t do anything with it. Look at the exchanges’ earnings reports: data revenue keeps rising, a bigger share of the business and a bigger goal of theirs. The answer has not changed: stay a data company and layer the new technology on top. The institutions in this market — banks, hedge funds, brokers, energy and commodity houses — can ask questions of the data in plain English or pull it raw through an API; underneath, the inventory keeps growing, because nobody serving this appetite survives on one asset class or one geography. The unglamorous half is compliance: everything a data provider hands a trading desk answers to the policies of dozens of exchanges and other information providers, and the good data companies get very good at managing those requirements. The more decisions get delegated to agents like JPMorgan’s, the more that dependence compounds.
JP Morgan’s own strategists, for their part, caution against uncritically accepting what amounts to “overly confident answers” from an AI. They are right, and the problem is older than the technology. An agent will answer anything you ask; it will not pause to wonder whether the numbers underneath it are wrong. Somebody still has to. From the trading floor on, every stage of this evolution has arrived at the same place, and this one does too. You’ve got the world at your fingertips, or at your keystroke, to ask anything of that data. But the data’s got to be good.
