Earlier this year, LTX, a subsidiary of Broadridge Financial Solutions, launched BondGPT, an application driven by OpenAI’s GPT-4, designed to streamline corporate bond selection. Last month, the company announced the launch of BondGPT+, the enterprise version of the application, which contains a range of new capabilities.
In this Q&A, TradingTech Insight talks to Jim Kwiatkowski, CEO of LTX, about the feedback received from clients since the original launch, and how the product has evolved since then.
TTI: Jim, can you tell us about the journey from BondGPT to BondGPT+? What has been the client response?
JK: We released BondGPT in June, and launched BondGPT+ only four months later, in October. So the journey has been brief, but it’s been very exciting because we’ve received lots of client feedback, which has been overwhelmingly positive. Customers were challenging us to integrate some of the unique data that LTX has, such as data from our Liquidity Cloud and from our bond similarity model, with data that they can get elsewhere, like CUSIP number, issuer details, call features, basically bond description information and simple pricing and rating information.
So we pulled all that together and the reactions we received indicated that clients could now access this data a lot more swiftly than before, leading to all sorts of use cases where what would have taken them minutes previously, they could now do in seconds. That response time, that ability to access multiple data sets from different systems to achieve the insights that they’re looking for, really captured the imagination of our customers.
TTI: You mentioned receiving ‘overwhelmingly positive feedback’ from clients after BondGPT’s launch. Can you provide examples of how this feedback directly influenced features or improvements in BondGPT+?
JK: In speaking with some of the largest asset managers in the world, we found that many of them have projects underway to incorporate GPT for their own organisations, and they’ve been impressed with what we’ve been able to deliver so quickly. So the conversation then turns into one of how we can offer them the foundation and work with them together to build upon that. Clients suggested it would be really powerful if we could integrate their own proprietary data into our system. So as well as basic bond description and pricing information being pulled in from third parties, they wanted to include their own proprietary data, such as their own models, restrict lists, recommended lists, and analyst information that helps them to select bonds more carefully. For example, we have included our own bond similarity model but some firms have their own models as well.
Customers also wanted to make BondGPT available as an API. With BondGPT+, we have two different integration methods, one inbound and one outbound. And because customers are asking us to incorporate their own data, they’re also asking us about entitlements. Can we show them what are we going to do with their data? Is it being used to train models that might be used by other users, for example? We’re very conscious of security, so the answer to that question is of course no, but we needed to demonstrate these entitlement and security capabilities, particularly where customers want to entitle one set of their own users differently from another set. They also wanted to look at usage statistics, who’s logging on, who’s using what data set, i.e. information about their own user community.
TTI: The ‘Show Your Work’ feature in BondGPT+ seems fascinating. Can you tell us more about that?
JK: We’re really proud of the ‘Show Your Work’ feature and it is a great example of ‘explainable AI’. Along with ensuring accuracy through using curated sources of data and being careful not to use the generative aspect of GPT, it’s important to demonstrate how those accurate results have been achieved. That’s what the Show Your Work feature does. BondGPT+ is a conversational solution – you ask it a question and it gives you an answer. But how did it get to that answer? Well, by clicking on the ‘Show Your Work’ button, it will show every step it went through, whether it was going to a database to get some information, going to another database to get some other information, doing some bond calculations to combine that information, reformatting the information into a table, rounding, whatever it did.
TTI: How do you ensure that BondGPT and BondGPT+ adhere to regulatory guidelines, especially considering the role of broker-dealers, and what advancements in compliance have been made in BondGPT+ compared to its predecessor?
JK: From our first release of BondGPT, as a broker dealer, we were conscious of regulations because we have obligations to retain certain information. Also as a dealer, when we’re responding to questions, we have obligations that we don’t do things like provide investment advice. So from our first implementation of BondGPT, we had what we call an ‘adversarial AI agent’. To answer a question, the system gathers information, does some calculations, and at the point where it’s about to provide the answer to the end user, it will go through one more step where this other AI agent reads the answer and compares it to a set of rules. For example, if the result is a recommendation to buy a particular bond, the agent might say, well you’re not allowed to recommend bonds, and it will provide an alternative answer. One of the things that we’ve added to BondGPT+ is the ability for customers to use that adversarial agent and that rules capability for their own custom purposes. They might be in another jurisdiction and subject to different regulations, for example. So that compliance capability that we had from day one is now customisable for individual customers as well.
TTI: When you originally developed BondGPT, you must have had a roadmap of what you were expecting to add to it over time as the product grew. Have there been any particular surprises in terms of the client feedback that you’ve had, which have taken you in a different direction than you might have gone otherwise?
JK: We’ve received so much feedback because we have very imaginative clients who have been very excited about what we’ve delivered to date. Originally our roadmap was simply to add more and more data through that one front end, and to incorporate more and more training on the vernacular of the end users to deliver a better product. We hadn’t initially contemplated adding the customers’ own data, but in a way that was actually on the roadmap, because it’s just another data source. We hadn’t contemplated a lot of the customization features. Also, from our first version, we had some sample questions on the screen, which were there so that people understood the power of the product. They weren’t really part of the user interface, they just made it easier for someone to understand what the system was capable of. Customers have taken that in another direction and asked if they can augment those sample or default questions with their own, and organize them in their own way. What’s really interesting is that they can have those questions asked automatically at seven o’clock in the morning, or every hour, or every minute, or maybe based on what’s happening in the market, if Gold reaches a new high for example. We hadn’t really considered the power of what we’d built to be used that way. And in BondGPT+ the user interface allows for all of that.
TTI: Can you give us an idea of what will come next?
JK: We started out with a pretty basic problem we were trying to solve, which was to give customers a set of bonds that would meet their criteria. And ultimately, we’re a bond trading platform, so we want to help people select the most liquid bond that they might want to trade. Now that people are accustomed to what ChatGPT can do in everyday life, and they’re accustomed to the fact that BondGPT is delivering that level of accuracy with curated data, they are now asking more complicated questions around things like bond swaps, relative values, global macro indicators and so on, and wanting to use those factors in their criteria for bond selection.
TTI: What’s the commercial model? Is BondGPT+ standard functionality within the Broadridge/LTX platform? Are there additional costs if customers want to onboard their own data or other data sources, for example?
JK: BondGPT is a standard feature of the LTX trading platform. All of our trading customers have access to BondGPT. BondGPT+, because of its integration capabilities – whether it’s their data into our system or our system into their own delivery mechanism and user interface – is licensed as enterprise software. We charge differently for that, on a subscription pricing model. And with every new data source, we need to train the AI on what the data is, how is it used, what questions might be asked of it and how it combines with all the other data as well? That involves training the AI, which gets factored in to the BondGPT+ implementation fees.
Subscribe to our newsletter