Episode Transcript
Elizabeth : Hi. Welcome to the Hedge Fund Huddle. I'm Elizabeth Kuhr, director of news content strategy here at LSEG. In today's episode, we will talk about the value of news, the trust premium, reporting accuracy and speed, and the role AI plays as hedge funds build out their new strategy. On the episode today, I'm joined by Vik Bansal, systematic portfolio manager at Centiva Capital. David Tattan, business lead of LSEG EMEA Execution Solutions, and Alexandre Hardouin, head of equities at LSEG. Welcome to the Hedge Fund Huddle.
Vik, if you can tell us just a bit more about what you did at Centiva Capital.
Vik: Sure. So I lead a team of eight people. We are responsible for trading equities and futures strategies, which are all systematically based. And yeah, I've been there about five years and the teams based in London.
Elizabeth : Great. David, can I go to you?
David: I run the European business for execution solutions. We have a team focusing on working with asset managers and hedge funds. Basically looking at their front-to-back workflow and helping them select and improve how they work day-to-day.
Elizabeth : Great. And Alex?
Alex: I look after equity trading solutions at LSEG. So we deliver news, content, analytics to buy-side and sell-side customers to help them make faster and accurate trading decisions.
Elizabeth : Great. Well, thank you all again for joining me today. I'm really looking forward to this conversation. Alex, if I can start with you, how do you define or consider a trustworthy source?
Alex: So what is really important for a source to be considered as a trustable source is to know where does the data come from? How it has been created? Who has created it? So if I take an example, we have that partnership with Reuters News and Reuters News. They are very well known because they are transparent on the inputs. They have journalists on the ground. So they source the information locally, they verify it. There is a human element of it because we know that now, lots of the sources, they are scraping the news from websites. They have very strong editorial and trust principles that they apply to their news. So really, people can trust their news. And in some instances, if they make a mistake, they would report on that mistake. Of course, it's very rare, but when it happens, they do it. So it brings a lot of trust on the news they are reporting. So as we are selling solutions to people who trade, we are really selecting the sources. So we are also working with the Wall Street Journal and Dow Jones. We also have very strict and also not only strict but also transparent policies where they really tell people how they are sourcing the data, how they are publishing it. So that's really, really important. It doesn't only apply to news, it also applies to data sets and content. And also on top of that, analytics and AI. We are producing we want to be transparent, for instance, on AI, on the source that are used to produce the results and the models we use as LLM, also with our AI solutions.
Elizabeth : So Vik, Alex has spoken a bit about the ways that data providers and news providers build their trust. What factors are you considering when you're looking at bringing in new sources into your hedge fund?
Vik: I mean, we trial a lot of different data sets and news is part of that. There's many different factors. So the first one just relates to actually what is the kind of USP of that data? Like if it's, a news data set, for example, maybe that it's a model trained on a local language that we haven't got already or they're doing something that's faster. It's got to have something that grabs you qualitatively. It's got to make sense. And we generally prefer it if the data provider has done a little bit of work in terms of almost like just dangling a carrot and showing us some simulation results, that this is what you might be able to achieve with our data. Obviously, we're going to test it all very thoroughly ourselves. But when you're faced with 100 different data sets and people coming to you all the time, you do need something to say, oh, actually, that looks exciting. But then obviously has to live up to that. And then I'd say we find out a lot about the data and the data vendor through the trial process. Because especially when we trade stocks and they're very complicated instruments because companies do corporate actions and dividends and stock splits and they get taken over and so on. So how you map that back in history is it's almost like there's not a right answer. There's different ways of doing it and how a data providers accounted for that can often tell us a lot. So I guess it might be related to a, you know, we might talk about AI and humans, but like in my team, we've got two people who are dedicated to, they're very experienced data engineers and developers. And they've seen hundreds of data sets and they can almost smell it when it's bad.
Elizabeth : Very technical.
Vik: Yeah, I think that's what I call it. We've had cases where we've done trials, data looks okay, but the team members have said pretty much vetoed it because of the way that they've interacted with the vendor, the way that they have handled historical events, how hard it is for them to map everything as they need to. All goes towards trust, which is a big part of using data, especially when you're going to manage money off the back of it.
Elizabeth : I want to dig in a bit about how your team analyses that data. Do you always have a human in the loop? Or as my colleague says, human at the helm looking at these data sets, or is it often machine-led first?
Vik: No, it's definitely human-led. I'm not saying that's just, that's how we do it. I think that it's possible maybe some asset classes you could lead with a machine potentially. But as I said, with stocks, I don't think you can. Because there isn't, even one data vendor who has three different data sets might have handled certain corporate actions differently. I know of vendors who have prices and then they might have analyst estimates and they might have fundamentals and they all handle the same event slightly differently. So you need, I believe you need a good human to be involved with that. And yes, we use technology as much as we can to help make that a fast process. And it's a lot better than 15 years ago when I didn't have this same expertise in my team and I was doing a lot of things very manually myself. That's not a good way of doing things. It's not scalable. But I think you do need people who know what they're seeing and what they're doing.
Elizabeth : So, interesting to hear the human side of the validation. David, I want to go to you now about the technical side. Using technology, accuracy is incredibly important when it comes to trustworthiness. So is speed for some sources. How is technology helping verify sources, whether it's the accuracy of it or the ability to use it in your solution?
David: Yeah. I mean, firstly, I would agree with the comments from the guys here. Today I was walking outside the headquarters of LSEG and there was a huge TV screen there showing news and on the side of it says LSEG trusted AI. And the thing that jumped out to me in that phrase was not the AI bit, it was the trusted bit. And I think when we talk about technology and AI and trust, I think, it's important to remember that technology doesn't always mean truth. But it can help with trust. So I think what I'm seeing across the client base that I deal with on the hedge fund and asset manager side is, increasingly using AI to build cross-checking or cross-referencing data or information. It could be anything from news to sentiment, etc. So yeah, there's a huge push to try and use the technology we have at our disposal to try and improve the trustworthiness of the data.
Elizabeth : Does that come more in validation or even in ongoing usage of the data in the news?
David: I think it's both. I think it's not a single exercise and then you're done. I think the more successful hedge funds and asset managers that I work with are doing is turning it into part of their operating process. So it's not simply let's buy the data. And that looks good. It's a bit like what Vik was saying is making sure that the right oversight, whether it's people or additional technology, is in place to make sure that the machine continues to operate as it should.
Elizabeth : So of course, we want that technology to find good results in our data and news sources. Alex, what is the right balance between speed and accuracy? Because in some cases organisations may be seeing that as a trade-off.
Alex: Yeah. And it's also depending upon what's your job role. So for instance, if you are trading, of course you want speed. So you want real time market data tick by tick data even milliseconds if you have an algo trading. So here you will want really speed. And I'm not sure AI has been built for this real-time work where you need to react very, very quickly. If you lose two seconds, you won't be able to execute the trade where you should have been executed it because something has happened. So I think it really depends. On the other hand, if you are like more middle and back office risk management, then you would need accuracy. You need to have the right data to check the quality of the trade. For instance, like TCA kind of analysis you need really a very, very accurate data sets to run your analysis. So you cannot choose between speed and accuracy. It really depends on what you are doing and what you are doing in the day. So for a trader, if it's a pre-trade analysis, for instance, you may not need speed. You would need to have the right sources, run the right models, the right analysis to come to a trading decision that can take longer trading decision because you are exploring different stocks to buy, you are checking the peers, the competitors of that company.
Alex: You are checking also, I don't know, like the financials, the credit default swaps of that company run a long analysis or there is a news and you need to react and trade very quickly. So depending on job role, what you are doing when you are doing it, it really depends. You can't choose really. It's a combination of both and depending on what you want to do, you would either use full real time solutions, content sets, news and or use AI to help, which take a bit more time. You even have AI, like we have a partnership with Microsoft Copilot Researcher. So when you start answering a research in there, it takes time to generate the results. We also have a partnership with Reflexivity around a solution called Deep Research. So you can say, for instance, yesterday I made an analysis on show me what the volatility on CAC40 was. It takes time because you go through the 40 constituents, find the volatility surfaces and run the analysis which are the more volatile stocks. So this would take time. So it's really a mixed answer that I'm giving you depending on what you want to achieve.
Elizabeth : Interesting. It's something that data providers, I think, also consider and news organisations that balance, making sure that the accuracy remains throughout the output. But you want to also sometimes be first and get that out as well.
Alex: Yeah, it's a mix of both. You're right. You have to report first. But what you report as a news provider has to be right. This is critically important. And for anything AI-related, you are not only providing an answer, but you are also. We have to be transparent on the sources we use and the thinking process that the LLM is using. So if they are four steps to come to a results. We should provide our users with the ability to come back to step one and see what has been the first assumption, because maybe the first assumption was wrong and we are doing a lot of work to put skills on top of LLM, which make the results more deterministic, meaning that we want to rely on the LLM when we have to rely on the LLM. But when we have the data and we have the resources ourselves, we build the skills so that we use AI to produce the answer based on that skills and some rules that we have defined as the providers, because we know that to produce this answer, for instance, on Post-trade analytics, it has to use the reported trades. So we don't want the LLM to find the answer itself. We want really the AI to go and check on our data sets to find the reported trades.
David: Yeah, I was just going to add to that. I think obviously our clients come in all shapes and sizes, long only or systematic. But I think the speed versus accuracy thing is important throughout, but it depends on which stage of the investment workflow that we're talking about. Is it the initial insight or is it the portfolio management? Is it the execution or is it the oversight? And so they've all got slightly different speed, restrictions. But I think the key thing is transparency. Whether it's having the transparency using one of our APIs or various different processes that we run to give the results to our clients because I think when people understand where things come from, then you get the balance. The balance is basically not maximum speed, it's maximum confidence.
Vik: I do think when we talk about, okay, if you're trading, the speed is the most important thing. I think actually slightly depends on the time horizon that you trade over. So as an example, and this, you know, this kind of happened, we've had conflict recently and there's been you would get some kind of news item come out which said there's a peace deal been done between Iran and the US. And basically, if you get that immediately, you get that quickly oil price goes down and you would trade off that, right? And then an hour or two later, it turns out that wasn't actually the case. And then the oil price goes back up. So if you're someone who trades and has, very quickly and doesn't mind holding things for minutes or seconds or even two hours, then you need the speed, right? You need to have that story and react to it both when it goes down and when it goes up. But if I'm somebody who holds for two weeks, I'm better off not actually having the speed in that sense. Right? I need the accuracy and the accuracy is there wasn't a deal at that point. And then I'm not reacting to anything. So I think it depends on the time horizon. And then the other thing I'd say is I think for us, and we find that we always want speed for ongoing live data, but we want accuracy in historical data that we are going to build our strategies off. But that might also be, not just accuracy in the data itself, it's accuracy in the data vendor telling us exactly what time and what date that data would have been available five years ago, because that's something we often grapple with data vendors, because, they'll have a date and a data point and we'll say, yeah, but when was it actually available? And that's like, it's not accuracy about the data. It's accuracy about when it was available as well.
Elizabeth : Yes. And the metadata around when that point was created and was it edited and were there updates extremely important.
Vik: Yeah, exactly.
David: And also another angle here as well is our clients also value having consistent data through the trade cycle as well. So rather than sort of going from system to system or data source to data source, that's usually where things start going wrong. So having a unified layer basically behind all of the various different workflows is important as well.
Elizabeth: I would like to get into alternative data sources. So text-based news and kind of data sets are not the only ways that hedge funds are getting information anymore. Audio, visuals like video images, satellite images have become a really important part of some hedge fund strategies. David, can you tell me a bit more about the role these alternative sources are playing?
David: Yeah, I think what's changed in the last few years is, I think a few years ago the decision around data used to be a lot more linear. Now it's turned into a multifaceted problem or opportunity, I would say, for people who are working in this space. So now you know, the data or the alternative data comes in all forms. Obviously you mentioned satellite, shipping data, sentiment information, video data, transcripts. So our clients are using all of that, not all the time, but and not all of them, but they're using that's the opportunity, I think. So I think alternative data is becoming prevalent or has become prevalent. I think the trick is which are actually producing signals that you can act upon. Is it actually giving you a proper insight that lets you make a decision at the right time? And I think that's where technology can help.
Elizabeth: And verify.
David: Yeah and verify and surface. So it's not just clients trading via API, it could be screen based. And so, a lot of our client base have got many screens going on. And so they're only human, okay. How do they know what to, what to focus on? Alex mentioned our deep research tool in Workspace. That's an example where we're helping surface the right information at the right time. And there are plenty of examples of where we do that for our clients.
Alex: We have seen that a lot recently with the crisis in Iran around the vessel tracking localisation in Strait of Hormuz. So we have a map where you can see the boats queuing because they can't cross the Strait of Hormuz anymore. So that's an amazing data because this drives will then drive the oil prices, for instance, and the shipping company prices. So that's really, really important. So to your point, when they are saying, well, there is an agreement signed. Yeah, but the boats are still queuing so no one is crossing the Strait. So the agreement is not really concrete for anyone on that market. The other point I'd like to make is when you add the sources as a hedge fund, it comes with a cost as well. So the news source you are adding has to bring something for your value chain. Because when you add a source, you need to build your infrastructure. You need to have cloud. You have cloud costs to ingest. You may have AI models to fit in with this, connect an MCP, whatever to you to some of your AI solutions. So it comes with a cost. So there is always a balance to have a decision to be made between the quality of the news source you are adding to your workflow and your workload compared to the costs, because we know that data come with the cost. And the cost is not only how much it costs you to buy that data, it's also how technically you would ingest that data. Is it normalised? Everyone is talking about MCP now in AI. So how many MCP do you have? How do you connect them? How many tokens they are consuming? Because if you have an API and your staff, your portfolio manager, start asking thousands of questions, the token consumptions will be like millions and you will consume your AI budget within a quarter. So yeah, there are lots of these challenges that I'm currently seeing with lots of customers.
Vik: From my perspective, I don't really care if data is traditional or alternative and the way things move so fast. I think I've said before what was alternative yesterday is traditional today, right? So it's like you said, I just want it to add value and produce something interesting. And there's lots of the world, especially the high frequency world, they're just using prices. I mean, they might be using tic prices, but it's just prices. So there is information in lots of things. One interesting thing I thought I read recently. So if you look at news, and you consider news also include say things that you're seeing on social media channels and sentiment there. There's traditionally people, will use other NLP models or LLMs on what's being written. But actually on those kind of channels, you get more information or certainly a lot of information, not from the text, but from the emojis. It's a funny thing that, it's, you know, that's alternative in some sense because it's not what you think of immediately. Right.
David: A picture says a thousand words.
Vik: Right. Yeah. So it's kind of funny.
Elizabeth: Interesting as emojis go out of style.
Vik: Yeah. Right.
Vik: Yeah. And I also just wanted to agree about the cost of data. So again, going back to the guys who do the dirty work in my team. I mean, sometimes we can test data set it adds value in our, back tests when we think it makes sense, but if it's a high maintenance data set, is it worth it? You know, if you're going to get into a relationship, it's going to be too high maintenance. You might not decide it's worth it.
Elizabeth: It's interesting the discussion about grouping data sets in relevant categories like objective, authoritative news sources being one, social media being another, data points like satellite imagery and shipping ship movement data as another to provide also a fuller picture of what actually is happening on the ground and like in geopolitics, among politicians. I'd like to shift the conversation now toward AI. So, Vik, if I can start with you, what are some of the strategies that you're using to really bring out the power or the combination of news and data?
Vik: I guess we're using AI in different ways. So it's true that we have some sentiment models that use LLMs, and they might be quite, not just the kind of average. Here's some English news and give me some sentiment. They might, as I mentioned before, they might be looking more local language and so on. We aren't using AI to create signals on its own. We've actually started looking at that, but I know if I got an AI agent to create a model or a signal, I think I would have to treat it as a data provider. So, if it's produced me a signal and I'd need to see some out-of-sample performance on that. At the very least, I mean, there's a whole host of things there about overfitting and so on. But where we've seen really good use of AI is in things like so we use it for prototyping code. And again, you still need the person who knows what they're doing, who can then take that and then do something with it, right. We've managed to, I think we had a thing the other day where we managed to speed up one of our live trading processes by about 40%.
Elizabeth: With the use of AI?
Vik: Using AI. So it wasn't even that the thing we were doing was slow. We were just seeing what was possible. And using one of the AI systems, it was able to find a bottleneck we didn't even know existed. So that was very useful. And then kind of an odd use, but really useful one was we nearly had a trading outage at one point in the last couple of months. Which was really unusual because we don't usually have that kind of problem. And we kind of everyone was at a loss as to what had happened. And we got AI to inspect logs and all sorts of computer nerdy stuff. And it worked out that there'd been one particular program on the whole network that had been updating at exactly the same moment that we wanted to run one of our systems, which had never happened before. And we were able to then sort of fix that. And tell the firm what. And I don't think anybody would have found that out. I mean, so that was really, really useful. So I think it is good for productivity.
Elizabeth: David, what do you see as the future in this intersection of AI and information filtering? When certain firms are using AI to filter information, where do you think that's headed?
David: So I think maybe a two-part answer. So I think when our clients are using AI and tech and data, actually there's a really good quote I heard this morning on the BBC from one of the journalists there. And he was saying that you shouldn't just ask the question, you should question the answer. And I think that's really appropriate for this sort of general discussion because, the number of questions we're asking AI is shooting up. But the value of, credible outputs and trusted insights is also rising because of that. But I think the way I see the hedge fund and asset manager world sort of evolving is they're thinking much more in workloads these days. And it's about the machine of an asset manager from start to finish, how are they using AI to sort of improve the operational efficiency. So that's two good examples there. But there are others where it's looking at the full lifecycle of a trade and spotting inefficiencies in how things are done. Or it could be you know, acting as an agent on top of, you know, some of the decisions or sort of a, sitting on the shoulder of the PM, giving a second opinion on things. I think that's definitely happening. And it's also, over the next few years, it's going to increase.
Elizabeth: It's about that conversation back and forth questioning the output.
David: Yes, exactly. It's about that. And it's about the, looking at the holistic view of what the firm is doing day to day as well.
Elizabeth: So from back end, back office usage of AI to then I would like to talk a bit about LLM output with you, Alex. So we've all heard about hallucinations. We've probably seen them ourselves. If we're using AI tools that the output is not always accurate. How is AI being used to fact-check some of that output?
Alex: Yeah. So we use AI to produce results. But then we also use AI to check what AI is producing, which is sometimes a bit difficult. But to your point, I think what we do is it's very good for to automate some tasks that are very, very that are consuming a lot of manpower. I mean, like we have huge content sets. I was checking yesterday the equity derivatives world options. We have 40 million recs and we report back tech. So it's a huge database. And of course there are some spikes sometimes in that database. You may have a bit that is reported by the exchange which is too high or too low. And before that, it was all manual checks. At the end of the day, you would come to a close price. It needs to be manually checked. So now you can use AI to automate these tasks and fix those spikes much quicker than in the past. In the past, you could spot a spike that happened during one day, probably 1 or 2 days after, because the customer was calling. You said, this is wrong. That can't be right. Oh yeah, it's wrong. And you had a manual input to fix that. Now with AI, you can automate.
Alex: But to your point, I think you also have to have someone, a human, that totally validates what the AI is doing because maybe that spike was actually a trade that happened was a block trade or whatever something had happened on auction. So it's also very important to have a double check. You cannot fully rely on what AI is producing by itself. And you need to put a check on top of the LLM models as well. So I think the conclusion is yeah, you can trust AI trust the model, but you need validation as well on top of AI and the models. And I think what I'm seeing from customers is to create agents, people are a bit reluctant at the moment to create really end-to-end agentic workflows because you need to know what the agent is doing. Pre-trade and trade and post-trade. Maybe you would have an agent, which does some. Pre-trade work, but not pre-trade, trade and post-trade because you want to check each of the steps yourself or with your team to ensure that what the analysis that has been produced for pre-trade analysis is right before going out trade and post-trade.
Elizabeth: And David, how are hedge funds thinking about differentiating? There are only a certain number of data sources.
David: Yeah. I think, I mean, in five years time, ten years time, maybe three years time, it's not going to be about the model. That all the models, that's not going to be the differentiator. I think it's going to be how they orchestrate, I think all the different pieces of the jigsaw. So it's data, it's how they use AI, like we've just been talking about it's working with the people. So which people can work with agents. So it's the PMs, the traders. So it's going to be, I think the edge and the goal for the more successful hedge funds and asset managers out there will be who can orchestrate things in the most efficient way, and maybe AI can even help with that.
Elizabeth: So in my last question here, I want to ask all of you what investments you think that hedge funds should be making in order to be successful? If we look down the line five, ten years time. Vik, can I start with you?
Vik: I mean, obviously I'd say invest in my team if I was talking to my. No, I'm just kidding. I think, well, they have to keep doing what they're doing actually and have been doing. So investing in tech, the cutting edge technology, which as I say, has been happening for 20 years. It's just now it's in a different field. I actually think that one thing that might be a differentiator, which then relates to your question as well, is I think people are going to be more important. The reason being that, as we know, you can get some strange results from just blindly asking a question to an AI agent or whatever. I think there's a real danger that if you don't question the answer, you're going to lose something about curiosity. What's the truth? And just how to do research, right? And you might have more powerful tools to do it, but there's no point having an artificial intelligence if you've got a real dummy using it, right? So I think actually investing in people and making sure that, myself included, are using the tools in a good way. And you don't you lose that intellectual curiosity, which I kind of worry about for people generally because it's very easy to get a fact pack. And also the fact might not even be correct. So I think that's part of what we need to invest in technology. That's a given. But we also need to make sure that we're hiring people who have a particular mindset and who don't just take things at face value. So that's my 2 cents.
David: I think people is very important, people that understand the role that trusted data, AI and technology all fit together, I think is very important. I think obviously investing in a forward-looking way in technology is important. And obviously investing in the right types of data. Yeah, so really quite similar.
Alex: And selecting the right data sources, news sources is critically important because if you fail on that one, the results, whatever you use AI or not AI will be wrong and you will make the wrong decision. I think also probably something is, the skill sets required will be slightly different. I think there is a lot more work to do to test, validate, evaluate both the data sets and the LLM and the output of the AI. So lots more investment is probably required on the QA testing. It can even be manual testing, not necessarily automated testing, so that you are confident in the results that are produced by a data set, a content set, a news source, an AI, an agent. Because you have run tests, you have trained the model. As we say very often on AI, you build your own confidence on the tool you are using before you can actually add value with those tools and generate alpha, which is what people want to do.
Elizabeth: Vik, David and Alex, thank you so much for joining me today on the Hedge Fund Huddle.
Vik: Thank you,
David: Thank you.
Alex: Thanks, Elizabeth.
Elizabeth: Thank you for listening to this episode of the Hedge Fund Huddle. If you'd like to hear more, find us on Spotify, Apple Podcasts, or YouTube.
The information contained in this podcast does not constitute a recommendation from any LSEG entity to the listener. The views expressed in this podcast are not necessarily those of LSEG, and LSEG is not providing any investment, financial, economic, legal, accounting, or tax advice or recommendations in this podcast. Neither LSEG nor any of its affiliates make any representation or warranty as to the accuracy or completeness of the statements or any information contained in this podcast, and any and all liability thereafter, whether direct or indirect, is expressly disclaimed. For further information, visit the show notes of this podcast or LSEG.com.
