When Canadian Bank ATB Financial deployed a robot in bank branches to greet customers, it was maybe the most visible sign that artificial intelligence is becoming an important part of banking. But this emerging technology has even more potential behind the scenes where it will transform all manner of processes, including institutional bond trading.
AI can revolutionize bond trading desks, which have been struggling since the 2008 financial crisis. Previously, bond desks would buy bonds from an important customer, taking the liquidity and market risk of holding those securities until a suitable buyer could be found. Under the Volcker rule of the Dodd-Frank Wall Street Reform and Consumer Protection Act, banks were prohibited from engaging in proprietary trading and that activity mostly stopped. Only firms that don’t fall under the Volcker rule can still take up some of that liquidity, but the end result has been less principal trading, especially of corporate bonds and more illiquid securities.
Now, AI holds the promise of moving bond markets closer to how they operated pre-Volcker, leveraging the information of past trading patterns and customer preferences to facilitate bond desks transacting more trades. Banks will be able to do that without taking the risk of holding bonds on their balance sheet because AI will add insight into who might take the other side of the trade.
Here’s how it works: Historical bond trading data is used to map out the transaction flow among a network of counterparties, revealing which customers banks trade with, what securities they trade, and who they trade with most frequently. Using advanced data analytics, that information does two things — identifies bonds that clients consider as substitutes for each other, and, by learning from the trading activities of customers, produces coincidence patterns. AI tools can then discern how often a certain group of customers consider trading similar bonds and, ultimately, analytics software will recommend who might take the buy and sell side of any given trade. Armed with those recommendations, brokers should be able to increase their ability to satisfy the liquidity demands of their clients’.
It’s the same technology and analytics that helps Netflix make movie recommendations based on a combination of your prior picks and the viewing history of other people who have similar habits. However, AI can do much more than recommend the latest Will Ferrell comedy. It can tell a bond trader that a customer that bought bond X may also buy bond Y. Using analytics to understand a customer’s investment process will take much of the guesswork out of fixed income trading. Now a trader will know not only what a certain customer bought last week, but also that some of the same investors who bought that bond also purchased another particular security.
That could amount to a sea-change in bond trading, potentially making trading desks more willing to fulfill orders, which would improve market liquidity. Dutch bank ING is trying this approach with an AI tool called “Katana.” A six-month trial of the technology produced quicker pricing decisions for 90% of trades, and cut trading costs by 25%. JPMorgan Chase is using AI tools in the bank’s fixed-income sales and trading operations to help sales people and traders anticipate the direction of markets. Other banks are experimenting, too.
In the coming years, the technology will be able do even more. Beyond leveraging data from one bank’s trading desk to predict investor behavior in buying and selling, the technology will leverage a broader pool of data (from many primary dealers) to discern what factors influence behavior and what attributes of a bond change that behavior. Ultimately, AI could facilitate trading across entire fixed-income market segments, matching buyers and sellers and providing dealers with a data-driven view of what impacts behavior.
Getting there, however, will require banks to clean up the massive amount of data that exists today, which does not follow any standards. Getting the most from AI will require industry cooperation in setting standards and best practices for how data is normalized.
For example, today, a structured note means different things to different dealers. One may view it as a payment stream linked to the performance of an underlying instrument, another as a bond with cash flows that require modelling. Similarly, counterparty names are recorded differently in databases even within the same firm. In addition, data outliers and human error must be scrubbed before data can be used to create a feedback loop. The ever-growing pool of that data then would need to be stored on a cloud-based architecture capable of scaling. Ultimately, all that may be too much work for many dealers, especially at smaller banks, who will turn to outside firms to undertake that massive task. Partnering with outside firms on such work will give dealers a network effect of learning from an even larger pool of community data.
At this year’s Consumer Electronics Show, The New York Times declared that, “the clear darling of this year’s show was not a gadget but the growing amount of artificial intelligence software” powering devices. Soon, the impact this powerful tool has on bond trading could be just as great.