Growing market fragmentation and the rise of off-book trading are challenging traditional buy-side trading analytics, says Henry Yegerman, TCA product lead at ISS LiquidMetrix. Speaking at TradeTech Amsterdam, Yegerman discusses where lit market trading volumes have fallen while off-book liquidity has, complicating price discovery and performance measurement. He looks at where transaction cost analysis is evolving beyond averages toward predictive analytics and shares his views on how deep learning models are playing a growing role in the future of data analytics, trading decisions, and optimizing execution strategies in increasingly complex markets.
Full Interview
Josephine Gallagher – So, let’s start off with some context. Where are you seeing big market-structure changes and changes to liquidity, and how are those dynamics changing the way buy sides need data?
Henry Yegerman – Well, buy-side trading analytics, where you want to actually provide them with actionable information, is built on a foundation of reliable data. Now we’ve actually been really fortunate in the markets over the last 20 years, because of the electronification of the markets, we’ve had a large increase in available data. We’ve also seen a lot of fragmentation in trading venues. There are 50 to 60 trading venues in both the US and in EMEA, but fragmentation has changed over the last five years or so. It used to be that you had these fragmented markets, but they were all what we would now call addressable liquidity. They were largely lit markets, and they were on the order book. Now we’ve seen a migration to more of what we call off-book non-addressable liquidity, and to give you a sense of the magnitude of the issue, a few years ago, lit markets were 42% of volume; now they’re down to 27%. Some of that’s gone to auctions, and you know that’s not during the regular trading day as well. On the other hand, off-book venues have increased from 21% to 36% of trading volume. Now that creates a big problem in terms of price discovery, and if you don’t have price discovery and volume discovery, that calls into question a number of key buy-side metrics, like the percent of daily volume, participation rates, the ability to identify adverse price momentum in your trading, so really the whole foundation of buy-side trading analytics is called into question due to this change in market structure.
Josephine Gallagher – So TCA (transaction cost analysis) is all about measurements, it’s just like your average cost versus your benchmark. So, how can TCA just go beyond measurement and toward you know, actionable decision-making and alerts?
Henry Yegerman – Well, that’s a great question. Because of the, let’s say, the massive increase in computing power, we can do things that we’ve never done before, in terms of recommendations and providing alerts, you know, to clients to help them trade better and avoid particularly bad trades. We look at a few things. First thing, I hate averages. Averages lie, you know. What we really need to look at our distributions. If you have a cost, let’s say a cost of negative 15 basis points on an order, what is that? Is it a, you know, tightly converging around negative 15 bps, or is it a barbell where you have some executions which are really good and other executions which are really poor. You don’t know, but that’s key to understanding the risk associated in your trade, and how you should trade. You know, actually, it makes a lot more sense to say, oh, I have a 70% chance of getting a cost between 10 and 20 bps than it does to say I have an average cost of 15. It gives you a real sense of the likelihood of where your costs are going to be. Secondly, we like to look at synthetic benchmarks, so we use a lot of scorecards, we use broker scorecards, algo scorecards, even algo wheel scorecards, where you really can take an overall look and bring many metrics together for a single benchmark. The problem that people have is that very often when they look at their trading, they’re looking at mutually inconsistent metrics. You want to have a high spread capture, but having a high spread capture lends itself to having a high adverse selection, which you don’t want. So, what we do is we allow you to create a benchmark, which you incorporate all of these metrics, which you can weigh the different constituents accordingly, so you can get an overall number that takes into account your risk appetite for spread capture versus adverse selection. Finally, we’re doing a lot, and this is really cute. You know, I wish you could see it. We do online A/B testing, horse races. We will set up in EMEA for stocks which have five to 10 basis points of spread, and an ADV from 1-3%. How does JP Morgan’s Aqua compare to Goldman Sachs’s Sonar? And you can actually do an A/B test to see right there which of them performs better for you under those circumstances. So that all leads to recommendations.
Josephine Gallagher – Really interesting, the horse-racing element, that’s really funny. I’d love to see that.
Henry Yegerman – It’s cute.
Josephine Gallagher – So we know that, like, obviously, there’s no one-size-fits-all, though, when it comes to TCA, it’s one of the biggest, I guess, challenges out there as well. Like, how can TCA providers like yourself, you know, tailor analytics and models specific to different types?
Henry Yegerman – Well, after you get past cleaning the data, because I can’t stress enough how cleaning the data and mapping it is the prerequisite. We spend a lot of time talking to the client up front when we on board them about what is their investment strategy, what is their philosophy, what are they going to do? If you’re a small hedge fund that’s a bottom up with a long alpha horizon, you might be looking for the next great consumer goods company in Indonesia, and you’re willing to wait two, three years for that investment to ripen. You do not, you’re not cost sensitive. Your focus is on accessing the shares, getting the liquidity, which might be particularly challenging, given that a lot of it still might be owned, say, by the founding family of that company. That’s a very different set of metrics than a large mutual fund complex, where a lot of their trading is what we would call uninformed trading; it might actually come from people who are contributing to their retirement plans or taking money out to finance a mortgage or something like that. So they don’t have a view, they don’t have a conviction on those stocks, they’re just taking out liquidity. So the firm is very focused on controlling costs there, and that’s an entirely different set of metrics that we can imply. You really want to tailor the reports and the analysis to the nature of the firm. It is not one size fits all.
Josephine Gallagher – Really interesting customization element of it. So, looking ahead now, what is next for trading analytics in itself, like particularly in the world where we’re seeing more and more involvement in AI and also, you know, more advanced data modeling. How do you see trading analytics becoming more involved in day-to-day decision-making?
Henry Yegerman – That’s a great question. The first thing I’ll say for any doubters who are out there is AI is real and everybody’s working on it, and we use AI, we use machine learning, right now, you know, for a number of our analytics, but the next step is deep learning, and the basic difference is machine learning. A human being sets out what you’re actually looking for. In the case of deep learning, the machine, the model itself, begins to think creatively to find its own patterns, and then provide that and suggest that to you to see what you want to consume. That involves an enormous amount of computing power, but more than that, it involves very large data sets. Even large mutual fund complexes are only looking at their own history, their own data, that’s not a big enough data set to train these kinds of models. If you’re going to use deep learning, which is really what you should be doing if you want to look at tick data or order book data. You know, we bring in, you know, Dow Jones Global News to also try to identify how news fits into these patterns. You know that requires a global data set. We have a peer universe of over 250 clients, which is thoroughly anonymized, but we are using that to actually build our deep AI models, and we’ll make that available to clients because they’re going to need that to have a broader set of data than just their own experience. I’ll give you an example of what I mean. You, as a fund manager, might know how Morgan Stanley’s VWAP algo in Asia compares to UBS and Goldman and JP Morgan, but all you know is how yours performs in Asia. You know, what we can provide you is, how does the whole market, you know, how does Morgan Stanley compare to JPMorgan, UBS, Goldman across the entire universe of trades. Because we have the venue statistics, and that to use that to power AI models will allow them to be that much more powerful. And also we are creating unique data sets, so for example, with some of our sister companies that have market-specific kinds of market intelligence, we know a lot about ETFs, which ETFs are active, which ETFs are passive, whether they benchmark to, so we can then apply our trading statistics and venue statistics to specific kinds of ETFs, so people can trade that specific asset class more effectively. So we think that the future is in large, unique data sets, you know, to empower everyone’s AI models, ours and our clients.
Josephine Gallagher – Really, really interesting. I got to say, thank you very much, Henry, for joining us.
