Six Takeaways on the State of GenAI in Fixed Income
Last month, LTX CEO Jim Kwiatkowski sat down with Sean Burton, Director, Fixed Income Trading at Liberty Mutual Investments, to discuss how GenAI is already being used in fixed income. They covered a range of topics from the basics of GPT technology, to the adoption spectrum and overcoming barriers to adoption, to how market participants are benefiting from this technology today, to what’s next. Here are six takeaways from their conversation.
1. The democratization of data with natural language processing (00:59):In recent years, pre-trade data has become increasingly available in the historically opaque fixed income market. While this abundance of data has been welcomed, participants now face the challenge of grappling with vast amounts of disparate data in different formats, from across multiple systems, and in different user interfaces. Despite data being everywhere, it is often not usable or easily integrated into trading workflows, and it’s proven difficult to extract value from the data. GenAI has arrived at exactly the right time to solve this problem.
View Transcript Sean: Everybody has been focusing on collating data and centralizing it in places like Snowflake, cleaning the data up, making sure that it's fit for purpose, and accessing it intermittently. Because most of the platforms and ability to access the data have been spotty at best. Right? You're doing these mini-apps trying to extract value from the data. And this GPT/AI has arrived at exactly the right point in time, I think, because, you know, using AWS as an example, I think our data is in a very good place. And now we need to democratize the ability to access that data properly and access all of it at once. Right. Because we're accessing a patch here, a patch there. But now this allows us to, using natural language, query that data. And that natural language part is the most important part because, like I said, it democratizes it. It means more people are able to question the data and get very creative.
2. The three categories of GenAI’s potential (02:36): Sean sees GenAI’s potential in three categories, all of which can help users gain competitive advantage. The first and second are Efficiency and Insights, which are where we are now in the journey. Much further down the line will be the potential for Tasking.
View Transcript Sean: I think, you know, the potential of AI in our market. I divide it into two main segments and maybe a third smaller segment. But at first, and I think the first way people will look at it and use it is for efficiency, right? The most obvious is the ability to conduct quite complex searches on data to find what you need. You can look for exposures, game loss as a trader, curve anomalies, and things like that very quickly. You can look at counterparties and a number of different things. Right. And it gives you the ability to quickly search empirical observable data. The second part, which by the way is a massive advantage, is, I think, the most exciting part. And that's what I would, I guess, call insight. And that's where people can get really creative. And you know, this is where you're going to see people really separate themselves, using that data to gain insight on the market. For example, if you imagine you've got a portfolio and you're trying to take a position where you're very overweight in Triple B's, and you really think this space is going to work, or you've got a certain duration strategy on, there may be inadvertent cannibalization of your progress that you haven't noticed. Right. So you can use the system to look at correlations, for example, and spot correlations across sectors that you may not have normally noticed and use those correlations for positive or negative advantage. In the sense that, you know, the system might be saying to you, you're using, you think long Triple B’s but it's not performing as well as you think because one of the positions has a very low correlation that you think should have a high correlation, and you shouldn't have that position, or vice versa. You know, it might say your focus is Triple B, but really you should be looking at these names that are highly correlated if you really believe in that strategy. And that's just the tip of the iceberg with this. But I think portfolio construction can be honed in a way that we reduce cannibalization, we're much more efficient, and we can get some truly interesting insights and use the system in both efficiency play and insight play to gain advantage. And the third, I think, is tasking. I call it tasking, which I think is further down the road. But this is where, for example, you can get that discipline to balance your portfolio duration multiple times a day or asking it to actually perform tasks. And those tasks may not come down to trading, but they certainly might be able to construct orders that have to be approved and to take a look at them.
3. Efficiency now: Separating the signal from the noise (01:22): While there’s unlimited potential in terms of how GenAI can provide value, the reality is that early trading and portfolio management adopters are already benefiting from efficiencies today. The ability to identify bonds within seconds based on a user’s specific criteria is one example.
View Transcript Sean: Because I think the big thing we need to do, Jim, is to separate the signal from the noise. And, you know, we've got so many disparate systems. Aggregators are a big thing, right, where we're trying to get market color and aggregators and pull all of this data into a central place that gives us the central, I think, signals, should I say. And I think that's really where this is exciting because. You know, like I said, data protection is a very big thing now, right? And it's very difficult to get individual data sources and to use individual data sources in your platforms. And I think AI allows you to search at a larger scale, and that's a very, very clear advantage on the efficiency front. Right. I can get a portfolio and I can say or even the market and I can say, you know, show me bond pairings with two standard deviations of diversion right across the entire market. Then I can say narrow that down to 5 to 7 years. Then I can say narrow that down to autos. Then I can say, take out Ford. Right. And I'm left with very specific bond pairings now that have an anomaly in that and that are maybe a couple standard deviations away from their mean, and then it's up to me to figure out the rest. Right. But I've been able to go through now and spot something and spot a signal or spot anomalies that I would never have been able to find, or certainly would have taken hours to find.
4. Addressing client concerns with BondGPT (02:16): As excited as clients are about GenAI and despite all the benefits it offers, there are many clients who are either not permitted to use it at their firms or are reluctant to use it. Often this reluctance comes from suspicion about the technology or concerns about privacy and accuracy. From the start, our GenAI application BondGPT’s architectural design has been developed with accuracy and data security in mind.
View Transcript Jim: It's a brand-new technology. And based on what we hear from most of our clients, the concerns when they're raised come down to a fear of the unknown, the fear of what the technology might do because it's new, a fear of the accuracy of the answers because clients haven't used this technology before, and they need to grow comfortable with it and with the accuracy it can provide. And then from a data security perspective, fear of how the questions might be used. And, you know, we took all this on board when we created BondGPT. And from the start, the architectural design was developed with accuracy and data security in mind. BondGPT avoids hallucinations from day one. We said there can be no hallucinations. We don't use the large language model for bond data. Instead, we use the large language model exclusively to understand the question. And then we return the data from verified, curated, high-quality sources. But what we found is, it's not enough to be accurate and to know we’re accurate. Customers, particularly traders in the moment, need to be confident with the accuracy. So explainable AI is also a key element of our solution, and something that we really focus on, allowing a user to just click on an answer and know how it was arrived at. BondGPT also has a Special Agent, which systematically reviews the responses to ensure they meet our compliance rules. And in some cases, our clients have looked at that. So that's very interesting. We'd like to use that to make sure that what we're providing our end users meets our own compliance rules. We've worked to educate our customers on how the technology is being used so that they can have the same confidence in data security that they do in its accuracy. Data security is always top of mind. What we receive, what we retain, how we safeguard it, and how we deliver it to the customer, for example. We never use client data to train our models. With BondGPT, no identifying information is sent to our large language model and any client data that we have access to is never stored or cached for later use. It's exclusively used to answer the question that the customer asked and then move on.
5. Today’s use cases and how use cases will evolve (03:38): Beyond assessing market color like new issue volumes or inflows, other use cases include the ability to quickly understand exposures to a particular credit at a given time, find curve anomalies, and search for correlations between credits or segments of the market. For those market participants interested in incorporating their own internal datasets and models into GenAI tools like BondGPT+, the possibilities are endless.
View Transcript Sean: This connects so many disparate points of data, right? And so many different things come at once. Normally when we've got data, we're looking at a feed, maybe 2 or 3 feeds in a platform and hoping that gives us some insight. But if I'm using a platform like this to analyze credit or analyze what's best to buy, I'm just thinking from a trader’s perspective. Right. And so, we've been told, I'm told by a salesperson to look at a particular credit, say Ford, not only can I pull in trade, buy or show me the Ford bonds that are the most traded bonds, show me the most traded bonds in a particular space, show me the most liquid bonds, etc. Right. Great. But I can actually query Ford's results. I can query if we've got connected. I can query our analyst opinions on it. Right? It just saves me so much time because I get a holistic picture now using natural language, right? I don't have to be a coder to get all of this. And I can get it pretty simple and pretty easily. I can create these sorts of insights. I can immediately look up what my gain or loss is on my holding position. Connect those to, you know, the most liquid bonds. Show me my holdings. Show me the ones that have traded more than 10 million on trace in the last 24 hours, right. So, you can really narrow things down really quickly, but I can look at so many different disparate sources and I think, really, people are going to be playing with that for a long, long time. You know, the efficiency play. I think that's the first stage. And I think where the bulk of the action will be for a long time also because it's easier to get approval for, right, particularly on external data. Once teams incorporate their own data, then we're going to get real separation because you're going to have companies that have very, very smart quants and people who know how to analyze data like that. That's really what you're getting down to, that it's people who they may not even be from the financial world, but if you've got people who know statistically how best to look at data and how to read data, working with traders, then you're going to get the best answers out of this and you're going to get the most interesting things like correlation and, you know, things that I'm not smart enough to think about. And that's the kind of people that we will look for, people who really understand how to use it and how to understand statistics and models, because it's all, like we say, observable data. So, I think, that separation, you know, there's lots and lots of people can easily understand the separation of creative thinkers, who ask creative questions. You know, I think that's the key to adoption. And that's the key to this thing really turbocharging these markets and finding efficiencies and finding anomalies. But that's going to take more time and it's going to be very disparate between different firms.
Jim: Yes at LTX we really started working on BondGPT a little bit more than a year ago, and we're just seeing the use cases evolving very rapidly. I think one of the most exciting things about gen AI technology is that every time we look ahead at the next use case, at that next requirement that a customer gave us, we find ourselves getting there very quickly, much more quickly than traditional software development ever allowed us to. You pointed out, incorporating client data and how interesting the use cases can be when your own data is there. We're working with several of our own clients on implementations of BondGPT+; that's our enterprise version of BondGPT. And what these clients incorporate in their own data sets, their models, and workflows. Really, the possibilities are endless. I think we're just beginning to scratch the surface of how the technology can be used.
6. When is the time to embrace GenAI? (01:58): While most agree that GenAI is not going away anytime, ultimately, it’s human beings who make investment decisions – that’s not changing any time soon either. Informed human beings make better investment decisions, and we hear from clients who haven’t yet begun that they expect to use AI tools more in the coming year. For those who hesitate to embrace it now, they may be at a disadvantage.
View Transcript Jim: I was looking at a market survey that we just ran with Markets Media, the publisher, and 67% of respondents said that they expect to use AI tools more in the coming year. And from our conversations we hear a lot is if, if we don't embrace this now, we feel like we might be at a disadvantage in the future. What are your thoughts on that Sean?
Sean: You know, our CIO has a saying: You won't be replaced by AI, but you might be replaced by somebody who knows how to use it. And I think that's a very apt thing. And look, you know, I'm not someone who bangs the drum and thinks the most hyperbolic things about AI. I am pretty sanguine and realistic about it. And market adoption, particularly in fixed income, tends to be slower than in other markets, partly because of the opacity of the market and how people run it. Right? But look, if you can't ignore it, it's not going away. And I think embracing it, particularly with empirical observable data, is a fundamental shift in how we examine these markets. Because, you know, you look at a market like we've got right now where everything is completely squeezed and the technicals are incredibly tight, spotting individual value is very difficult. And so, you know, we can use AI to find the most severe anomalies in those markets and position ourselves in the future. Right. When everything kind of looks the same and everything's so compressed, the values are very difficult to spot. Using this kind of platform to scan the data and find the anomalies is incredibly valuable. Right. So, I think you have to embrace the fact that it's coming. And at the very simple level, you can (1:39) embrace the efficiencies that will give you life. And that at the more complex level, you can embrace the insights that it will give you and the ability to, you know, to really open up how you view portfolio construction and the liquidity of your portfolio.