Giuseppe Nuti, managing director and global head of the central risk book and data analysis at UBS explains how an AI system can work even on sparse bond market data, while Jim Switzer, global head of fixed income trading at AllianceBernstein, reveals his firms plans for AI on the buy-side desk. Market data provided by MTS.
Our open trading platform links to 1300 global credit institutions and provides 8 billion dollars in daily tradable liquidity. Join us at MarketAxess.
Dan Barnes Welcome to Trader TV Fixed Income – your insights into the trading climate for professional investors. I’m Dan Barnes. In September’s show we’re speaking with firms who see artificial intelligence as the future for trading bonds. We’re talking with Jim Switzer, global head of fixed income trading at AllianceBernstein, about automating the fixed income trading desk. And Giuseppe Nuti, managing director and global head of the central risk book and data analysis at UBS, about using artificial intelligence to enhance bond trading. Jim, welcome back to Trader TV.
Jim Switzer Thanks for having me, Dan.
Dan Barnes So tell me, why should investors care about efficiency on the trading desk?
Jim Switzer Well, I think before we talk about efficiency on the trading desks, we need to step back and talk about efficiency in the investment process, because that’s what investors should care about. And efficiency on the trading desk is just one part of that. So, when you think about an investment process and the connectivity between portfolio management decisions and portfolio construction, the bottom up construction of those portfolios at the research and trading level, it’s very important that we think about how we systemize our processes and bring all those processes together. Then you can ultimately get down to the trading desk and say, what are the efficiencies on the trading desk? And ultimately, look, the holy grail of of of trading is pre-trade transparency. It is having all that information at your fingertips before everybody else and in a more ordered manner, and that’s ultimately what we’ve done and we’ve spoken about it before. With our Algome Alpha tool, which is just an aggregator that’s bringing all that information in, and if we talk about how many messages we get, it’s 3.5 million messages a day, and how to put some kind of order and structure around that. And that’s where Alpha comes in, which is our liquidity tool, which is digitized liquidity, our view into liquidity, our Prism tool, which is the basically digitized, fundamental view. We have our quant views, which is already digital. And where all those intersect, that’s where our bot-technology, Abbie, takes over. And that’s a big part of the efficiency, whether it’s workflow for the APM’s creating billing orders and getting it to the trading desk, or Abbie actually highlighting to us trends in the market or things that we’re supposed to be aware of. That’s all part of it, so, a client should care about that because in the end, that’s all driving best execution, that is all driving optimization of portfolios, and hopefully actually generating real alpha into the portfolio.
Dan Barnes So how much of the bond trading process at the moment do you think you can automate?
Jim Switzer Well, I’m probably going to annoy a lot of people when I say this, but it’s 100% and basically you have to take a step back. ‘Are we talking about electronic execution or are we talking about electronic connectivity?’ Because if we’re talking about connectivity, then we should be able to do everything automated in some kind of electronic way. So 80% of our notional bonds that we trade are still done bilaterally over the phone, but we do not process them manually anymore. Everything is sent to some kind of platform where we can have it, auto-spot, auto-executed or whatever it might be. We break everything into high touch and low touch. We are building our low touch capabilities right now as we speak, trying to build sensitivity to our algos to be able to increase the bands or tighten the bands, depending on what we’re trying to do. And we’re trying to auto-execute as much of this stuff away. We might have a beta-trading desk and an alpha-trading desk. At the alpha desk, the ones that are actually looking for Alpha for our clients portfolios, and the beta where it’s just process and moving data as automatically and systematically as we possibly can.
Dan Barnes When we’re talking about the automation, are we talking about artificial intelligence or are we talking about more robotic process automation? Whereabouts does this sit?
Jim Switzer I would say it’s more of a robotic process, automation. It’s very algo heavy, and fixed algos, if you will, where there are specific rules, and that’s what our robots are looking for. I could see down the road – our robot is called Abbie, we have Abbie 1.0 and we’re working on Abbie 2.0.. Abbie 1.0 was there just to systematize and streamline our order billing process to allow our associate portfolio managers to get orders, process, and run compliance and get it to the trading desk quicker. Abbie 2.0 is something where we can actually have Abbie recommending certain bonds or certain sectors to us, based on what she sees in our algos in the rules that we’ve created. Down the road, we could see machine learning, and that’s something that we’re thinking about in Abbie 3.0, where she’ll have the ability to have flexible algos, based on the feedback she’s getting either from a PM or a trader on the success, so, ‘yes, I want to do the trader or no, I don’t.’ Basically, what we’re trying to do is identify, ‘are there bids filling in? Are the offers filling in? Is the market leaning one direction?’ And what that does is that it tells a trader sometimes to hurry up or to slow down. If the market is coming your way or the market has suddenly flipped. A sector specifically has flipped. And we see that right away with our tools. So when you think about best execution, and this is probably Pandora’s box here, best execution used to be a very two dimensional process for us. It used to be point of entry and then measure, ‘how did we do relative to our print, relative to the quotes at that time?’ And what we’re trying to do is say, ‘OK, if we systematize all our processes and we can come up with ideas quicker, because we see them in the market, Abbie can build us orders faster and get them to the trading desk quicker.’ And we have tools that are telling us the liquidity score’s not just a liquidity score on a particular bond, but a liquidity score in a particular bond on the bid-side or on the offer-side, because we believe there’s two liquidity scores for every bond. And then you have a momentum indicator that’s telling you that, that bond, that particular CUSIP, that ticker, or that sector, bids are fading, offers are filling in, and that will help you address speed. And to us, when we take all that together and put that into the best execution process, we start to look at something that’s much more three-dimensional or four-dimensional. What you didn’t add on to that is the strategies you have to go trade. You know, the typical ways of RF-paying bonds or bid-wanted in comps, whatever it might be, are changing now. And a lot of that’s driven by the ETF ecosystem that we’re seeing develop out there.
Dan Barnes Do you mean the fact that there’s more passive investment in markets?
Jim Switzer I’m saying we can use, you know, the sell-side right now is managing a lot of their risk through the create-redeem process, through the ETFs. So now, if we go to them to manage our beta risk in portfolios, just a slice of beta. We want to take some beta out, we might want to liquefy it into a different type of asset class or a different security. Maybe we want to sell cash bonds and take total return swaps, or we want to take index. Just to have a different certainty of liquidity, if you will. We can go to those to those dealers and say, here’s the risk, this is a slice of beta. Now, if that’s not truly a diversified basket, they can augment that basket with risk off of their desk and turn around and go back to the ETF, deliver that in and move it out that way. So there’s a lot of really interesting things that are going on in the market in terms of how we’re going to move risk, how we’re going to trade going forward.
Dan Barnes That’s great. Jim, thank you very much.
Jim Switzer Thank you.
Dan Barnes Now we’re going to speak with Giuseppe Nuti, global head of the central risk book and data analysis at UBS, to hear how it is employing machine learning to enhance bond trading for its clients. What are the quantifiable advantages in using machine learning to enhance bond trading?
Giuseppe Nuti Well, this particular initiative is really about the corporate bond space, where liquidity is fundamentally the premium commodity. This is a recommendation engine outing, so, like Netflix or Google is, at its core, a recommendation engine, in the sense that you put something into the Google search engine and it recommends at least a website. It’s very much like Netflix recommends movies for you to watch or series. And this is exactly the same thing, it recommends our salespeople a list of clients to call to the trader vs our axes. So here, what we’re trying to do is build a system that enables us to say, ‘well, these are the things that we want to do. This is the patterns that we’ve seen in the past from from our clients. This is the public data that we can we can scan.’ And through that, we can say, ‘well, we should really call these 10 clients, because they’re most likely to respond.’ So in terms of the advantages, first of all, you’re more likely to do the trade you want to do, and that’s a pretty obvious advantage. Secondly, you cover your clients more effectively. They will get a a more a more pinpointed phone call, and that allows our salespeople to cover their clients more efficiently.
Dan Barnes So what do we mean by machine learning in this context?
Giuseppe Nuti Where I think for us at least, the machine learning comes in, is trying to augment both the data set and the way that we analyze it with more features. A perfect example is, we have this big list, it’s an Excel spreadsheet, nothing fancy, and we have a bond that we’re trying to buy. We say, ‘hey, there’s this fund here, it’s traded in the past, so we should give them a call.’ But maybe that fund is a passive fund that just does index rebalancing and that happens only at month-end. The month-end is just a feature. Some features are obviously much more complex. Maybe there’s a relative value fund that wants to trade a particular US-dominated, corporate bond vs its equivalent in euros when the spread, and that’s quite a complex feature. I think where machine learning really does come into play is in being able to throw a large number of possibly interdependent features and analyzing all of the data together and say, ‘well, this particular fund is quite likely to respond when the spread crosses,’ but it has to also mean that the treasuries has to be within this level. And so combining rules in a way that it would take a human an inordinate amount of time to do it by hand. The problem, I think there is that there isn’t actually that much data. Most of it is price data and price data, If you want to measure the amount of information there vs data, that’s just not that much information.
Dan Barnes How do you train a machine learning system on that data set if it’s so sparse?
Giuseppe Nuti It’s a great question. So there are two challenges there. One is the more obvious one, which is just the statistical significance. In other words, like, yes, we’ve seen these particular two funds trade this bond before, and then we can start making all sorts of theories about what triggers that and the spread here. But if we’ve seen two data points, well, it’s kind of irrelevant. Or more interestingly, what about a new bond that’s just been issued? There’s zero data points by definition. So I think from the statistical significance point of view, it’s is a reasonably well-studied aspect in academia, and I think you just have to approach it that way. So the other, more challenging aspect is actually trying to squeeze more information out of the data, and I think this is probably where most of the work is for us. And the best way to sort of illustrate that is I’d say, imagine that a particular client trades a financial name quite actively in the five-year sector. Well, if a new bond comes in and has just been issued and this is day two of the bond, giving them a call is probably not a terrible bet. And so here, linking the information in such a way that it makes sense, is fundamentally the challenge of making the most out of what is a sparse data set.
Dan Barnes So it strikes me then that there’s a certain amount of time that you need to be in this to build that data set to learn those lessons?
Giuseppe Nuti Well, yes, and I think probably the point that you’re bringing up is the single most important point of why we’re focusing on this so much. I think, in in our industry in general, there are certain situations which present a clear first mover advantage, but they’re not that common. In this particular case, I think there is a bit of a first mover advantage in the sense that it’s quite likely that more and more firms are going to go this way, and more of these algorithms are going to become more and more sophisticated, both on the buy-side and on the sell-side, which basically means that the more data you have, the better your recommendation.
Dan Barnes Giuseppe, thank you very much.
Giuseppe Nuti Thank you. Thank you for having me here.
Dan Barnes I’d like to thank Jim Switzer of AllianceBernstein and Giuseppe Nuti of UBS, for their insights into the use of artificial intelligence in bond trading, and of course you are watching. To catch our monthly reports on other markets or to subscribe to our newsletter, go to TraderTV.NET.