Transcript
Speaker: This is HSBC Global Viewpoint, your window into the thinking, trends and issues shaping global banking and markets.
Speaker: Join us as we hear from industry leaders and HSBC experts on the latest insights and opportunities for your business.
Speaker: A heads up to our listeners that this episode has been recorded remotely, therefore the sound quality may vary.
Speaker: Thank you for listening.
Speaker: You're listening to the Markets and Security Services Outlook, a podcast miniseries exploring the critical topics that will shape our industry in the next decade, including sustainability, digitalization and emerging markets.
Speaker: Find out what's driving the global outlook for institutional investors and where the opportunities and challenges lie.
Speaker: Thank you for joining us.
Speaker: My name is Jeff Wertheimer.
Speaker: I run the global EFX sales team for HSBC and the
Speaker: Cross Asset Class E Execution Team for the Americas.
Speaker: And I'm proud to have with me my group of colleagues for this session called E-Panel, Artificial Intelligence and Machine Learning in Markets.
Speaker: I'll give a brief introduction to the team we have here today.
Speaker: So first we have Ed Duggan.
Speaker: Ed runs HSBC's global electronic trading business for equities.
Speaker: And Ed is based in London.
Speaker: He's a great person to have on the panel today.
Speaker: because he's been tasked with leading the build-out of our next-gen execution algorithms in the equity asset class.
Speaker: Next, we have Chris Ulf.
Speaker: Chris is the global head of equity execution quants, and that makes him the quantitative lead for the HSBC equity execution algos.
Speaker: And Chris has a PhD in econometrics, which I think could play into our conversation today.
Speaker: And last, but certainly not least, is Paris Panesey.
Speaker: Paris Panese is the head of systematic trading strategies for spot effects and commodities for HSBC and our secret weapon.
Speaker: So Paris holds a PhD in artificial intelligence and a master's degree in electronic engineering.
Speaker: So the first question that I wanted to ask, how would I explain AI to my young children?
Speaker: Paris, I'm going to start with you.
Speaker: We're going to give everybody a chance to chip in on every question.
Speaker: But Paris, if you could start us off, please.
Speaker: Yeah, thank you, Jeff, for the intro and for the questions.
Speaker: I think it's probably the most difficult question, right?
Speaker: Firstly, you know, when you have to explain something to a child and also to explain what is it, AI.
Speaker: I think it's interesting because you can take many angles.
Speaker: You can say, what is it right now in practice?
Speaker: How is it used?
Speaker: What people would like to be, what people think AI it is.
Speaker: I mean, they're very different.
Speaker: They might lead to different answers.
Speaker: So from my point of view, what AI is, if we remove all the magic around it, I think it's you can think about automating something, right?
Speaker: And what is it automating?
Speaker: You can think that robots can automate and help human performing actions, whereas AI automate and help human take decisions.
Speaker: I think that's the distinction.
Speaker: You have robotics, physical actions and AI is something that helps you to make decision.
Speaker: And then going what people would like AI to be or what's the vision might be is that the system that takes those decision right now is mostly dictated by what the designer of that software decided how those decision needs to be made.
Speaker: But then maybe going into the future, what we would like to have is something that that system learns by himself what's the best decision and what data does it need to take the decision.
Speaker: I think that's the way.
Speaker: I don't know if it's simple enough for a child, but I think that's the way I can see it.
Speaker: Yeah, I think, funnily enough, my son said to me, and like all children, they get their sort of content and their news or whatever they call it from lots and lots of different platforms.
Speaker: My son expressed this
Speaker: nervousness he's just on his teenage years cusp of his teenage years about being what are the what jobs are we going to do when we're older i don't know whether he had seen something on youtube or whatever uh some slightly hyperbolic uh description of of the role of ai and uh and so i sort of said no actually i don't i don't really think of it that way i don't think it's going to develop that way i think
Speaker: I think the way we use automation and computers which have an element of AI within them will be hopefully to take away tasks which are generally very data driven and quite repetitive, but won't no longer require a GUI which pops up decisions to be made by sort of human operators, you know.
Speaker: I tried to put him, to give him some ease on my view of the world is that that media story about AI, meaning that lots and lots of humans won't have jobs anymore.
Speaker: I don't believe in that.
Speaker: But one of the things I talk about just to my friends who aren't in the business and they talk about what does it mean, what do some of these things mean?
Speaker: I say, well, actually, in a way, often you'll interact with AI programs without even realizing.
Speaker: And this is not even the super sexy stuff.
Speaker: This is a really boring part of AI.
Speaker: So
Speaker: I say to them, think about the last time you did some internet shopping.
Speaker: Sometimes you're asked to confirm that you're not a robot.
Speaker: You have to type in a sort of quite funny sounding code.
Speaker: All you're doing is actually taking those few letters that they show you or the pictures where you've got to identify a traffic light or whatever, and you're just giving a human input into an optical vision that a computer has not been able to recognize.
Speaker: We've not been able to train a computer to recognize
Speaker: certain images.
Speaker: So they need humans to do that work for them.
Speaker: So when next time you go shopping and you think this is really annoying, I have to click through these squares and identify where the bicycle is and all that kind of thing, that's actually just data that we're helping clean so that these AI processes, machine learning processes can become more efficient and effective.
Speaker: Tell your son if he develops a program that can spot bicycles
Speaker: traffic lights and cars, he'll have a very bright future for himself.
Speaker: There's a very interesting sort of flip, which is we're not going to run out of work because eventually the computers will ask us to do their data cleansing for them, right?
Speaker: Exactly.
Speaker: All right.
Speaker: Thanks, guys, for that start.
Speaker: And Ed, you mentioned media, which is a good segue to my next question.
Speaker: The terms AI and machine learning, they're buzzwords at this point.
Speaker: You see them in the news a lot everywhere, including finance, but across disciplines really.
Speaker: And I just wanted to know in terms of the curve, the evolution, where are we?
Speaker: Is it hyperreality?
Speaker: Are people actively using AI and machine learning in capital markets today, or is it really more of a buzzword and
Speaker: you know, kind of a longer term vision and maybe Ed, you're a good person to start with.
Speaker: Yeah, I mean, I'll give my sort of take and then kick it over to Chris and Paris, who do a lot of work on this in the detail.
Speaker: It is real, right, guys?
Speaker: There is an awful amount of like all these new sort of
Speaker: developments in sort of computer processes and our use of computers.
Speaker: There was a lot of hype and inaccuracy and very broad brush terms used to sort of describe it.
Speaker: And then the media get hold of it and they want to paint pictures in a certain way.
Speaker: That's not a judgmental comment, but what it means in capital markets is
Speaker: For me, AI and really machine learning is all about how we think about data, interact with data and analyze data, right?
Speaker: And how we do that in a smart and efficient fashion.
Speaker: So for me, it's actually, it's relatively unsexy when you think about the practical application.
Speaker: So when you think about capital markets, you know, predominantly publicly traded capital markets is just an enormous amount of data.
Speaker: There's an enormous amount of data.
Speaker: That data used to be, as many clients will know, used to be relatively high level of asymmetry between the people who sort of were custodians of the data and how data was interpreted.
Speaker: And you can think of very obvious, you know, impacts of that, like price dislocation and what have you, is really a reflection of an asymmetry in information often.
Speaker: So we need to get better at how we, you know, how we assimilate huge amounts of data.
Speaker: And we use that data because Paris and Chris want to go and build algorithms which effectively can interact that data in a more efficient way than we can do today.
Speaker: So it's very real.
Speaker: Where we have to be a little bit careful is, it's like all computer applications, it isn't suitable for every area of life that we look at, right?
Speaker: So AI and machine learning are real operational today and extremely effective in capital markets which are very data rich.
Speaker: They are probably less effective in markets which are not particularly data rich.
Speaker: So I don't see them particularly as predictive toolkits.
Speaker: I see them very much as extremely strong analytical toolkits.
Speaker: But I'll let Paris and Chris sort of expand on that or take that in a different direction if they want to.
Speaker: Yeah, I think I think it pretty much.
Speaker: That's my view is in line with that in a sense.
Speaker: And what I think I would say is it used in capital markets?
Speaker: I think it's used in a sense that I wouldn't see this as a step change or a revolution on what people do in capital market when try to be
Speaker: quantitative and systematic on their decision is more a natural progression on what, you know, maybe a few years ago, people looking at more simple analytical tools or optimization tools to make those, you know, decisions and about where to allocate assets, for example.
Speaker: And I think the incremental step going to AI and where there has been a
Speaker: I would say a differentiation is where you start having availability of data that were too much and too diverse to be ingested by the traditional, I would say mathematical tools, more econometrics tools that were used in the past.
Speaker: So then
Speaker: At the same time, it emerged this new type of mathematical tools, then the discussion becomes a little bit technical, but effectively there's no new tools, but when I say new, it's 20, 30 years old, now are usable and useful because you have a lot of data, that many data are not really fit to be analyzed with the traditional mathematical tools.
Speaker: So now this new mathematical
Speaker: type of tools, machine learning, are the ones that allow you to analyze more efficiently this large quantity of data.
Speaker: And like I said, there is that trade-off where in certain areas of capital markets, you have a lot of data and no theory to base your decision on.
Speaker: In other type of area of markets, you have very few data points, but there is a lot more theory around it.
Speaker: Think about the ratio between interest rate and FX market, right?
Speaker: That's quite, so there is a rationale why things move together that are not entirely data driven, it's more theory.
Speaker: There is a theory supporting that.
Speaker: So then when you move away from that area where you have a well-founded theory, but not a lot of data,
Speaker: towards where something is more, less theory, particularly in a short time horizon when you want to execute a trade, when there's a lot of data, then that's where more those AI machine learning tools are used most.
Speaker: So I would say in the area of executions and people heard about maybe high frequency trading, that's another buzzword there.
Speaker: That's probably the areas where they're most used currently in capital markets, AI machine learning techniques.
Speaker: So the next one, I want to get a little bit more specific and talk about how HSBC is using AI and or machine learning in your areas.
Speaker: And maybe combine that with another question that's come in about client engagement and how should clients be thinking about it?
Speaker: Are the things we're doing impacting them directly yet?
Speaker: Is there anything they need to be thinking about?
Speaker: But maybe the first question is more just around a bit of specifics about how we're using it already today.
Speaker: And Paris, why don't we start with you for this one?
Speaker: Yeah, from my point of view, the way I particularly use it in my team is about how do we, so what do we do?
Speaker: We do market making in effects primarily and out of execution.
Speaker: And then the main thing we need to do our job is to come up with a price we offer to our clients.
Speaker: And then, so our belief and what we experienced is that using some of those techniques help us to make a better price.
Speaker: And then this better price gets immediately passed to the clients because we can be more competitive versus other
Speaker: the banks are the provider of this liquidity and then effective the advancement in technology, data and quantitative techniques.
Speaker: And in recent years, most of them have been around AI directly affect clients because they would be able to receive more accurate and
Speaker: efficient pricing.
Speaker: So that's the first benefit they get at least on my area.
Speaker: Then on a similar product is where we also offer algo execution tools for our clients, then all the advancement and all the improvements we can make on the decisions on how to execute a certain orders, they get embedded in these algo.
Speaker: And then the clients
Speaker: quite transparently, they will send the order in the same way they sent five years ago, but the way that the Argo takes decision and the improvement on the execution are driven, I would say, I'm not sure if primarily, but materially by the advancement in the way we use AI and machine learning.
Speaker: With the usage of AI within HSBC, it's absolutely growing and it's a huge part.
Speaker: In a similar sort of aspect, you know, in the equity side, we directly utilize AI as part of our execution algorithms, more driven by the fact that we are very rich in data with regard to that.
Speaker: And a lot of this is very much trying to
Speaker: enhance our decision process, being able to capture that, you know, lots of the aspects, particularly in trying to capture a lot more of the scenarios in the market and a lot more of the non-linearities in the market and lots of the interactions.
Speaker: the machine learning or AI very much is quite advantageous and we have lots of evidence that proves our performances as well.
Speaker: But I think the other aspect to think is also on an overall thing.
Speaker: AI is being broadly pushed right throughout HSBC in addition to our areas more in markets.
Speaker: And there's lots of research on machine learning, on NLP, looking at the language processing,
Speaker: Other aspects, you know, that we're very much looking is into research and advisory, being able to utilize the wealth of information, being able to focus decision points, very much enhancement.
Speaker: And on top of that, with the bank as well, there's also, you know, machine learning research in fraud detection as well.
Speaker: So it's very broadly used and, you know, certainly something that we utilize on a, you know, on a daily basis.
Speaker: And Chris, you mentioned their NLP, just in case anyone doesn't know, that's natural language processing.
Speaker: And we're using that all over the bank now where people can go in and type the way they speak.
Speaker: And our systems are able to parse that and be able to act on that.
Speaker: So you see that showing up a lot of places in your lives right now.
Speaker: And that's a big change.
Speaker: It really speaks, I think, to the accessibility of some of these tools.
Speaker: Right.
Speaker: You know, back in the day to build a website, you needed to know HTML.
Speaker: And now there's tools where you don't need to do that.
Speaker: So it makes, you know, it makes things more accessible to the masses.
Speaker: So it's in your opinion, what is the biggest misconception about AI?
Speaker: And again, I think that the term is used so much.
Speaker: There's probably quite a few.
Speaker: But, you know, Paris, why don't we why don't we start with you again on this one?
Speaker: Sure, yes.
Speaker: I think it's a bit linked up to when we started.
Speaker: I think there is the misconception, the way I see it, it comes from what people thought AI would be or should be versus what AI actually is and what you can actually do.
Speaker: So I think there is this vision historically, going back now, maybe Turing said, okay, what is AI?
Speaker: When you talk to a system or the robot you were talking about,
Speaker: and that you cannot distinguish them whether is it a real person or is it a computer.
Speaker: So that's the AI.
Speaker: I mean, we are quite far away from having a real conversation with a computer than where you cannot distinguish it from a machine.
Speaker: I mean, the misconception is that I think in the media and people think that we are very close to it, to this AI, the vision of AI and what actually
Speaker: the AI can deliver.
Speaker: So that misconception is that, in a sense, the AI can very closely replicate what a human does, because it's a bit misleading, because for certain tasks, it is very similar to what a person can do, for example, recognizing images, or even talk, or natural language processes.
Speaker: They're very similar to what a person can do, but those are like kind of basic cognitive, you know, function that a person can perform.
Speaker: And I think AI is at that level of basic cognitive function.
Speaker: You know, we talk about children the first time, maybe that's at the level of a four, five years old, you know, but it can do it very efficiently over a very large amount of data.
Speaker: So I think that is the misconception.
Speaker: It cannot have the higher level of reasoning that a person can do.
Speaker: I think that's...
Speaker: they say the way i see it yeah i think i think the i think the misconception really is about how far it's going to take us right and and i think paris has done a really good job of it of explaining that we're going to see an awful lot of incremental benefit and it's good but it's going to be incremental benefit in an awful lot of areas of society where where we apply that um
Speaker: ability to consume and learn from and train machines to analyze data for us right I think I think that's going to come across as much less of a leap forward for most of the areas that we apply it that may be the public things I mean I think where where it's going to be really really effective is where you're trying to scale up
Speaker: processes which are heavily dependent on human interaction, right?
Speaker: And so the fact that you can train machines to spot cancers and x-rays, right, means that you can do an awful lot more x-rays because they don't always need to be shown to an oncologist, right?
Speaker: So any area of what we do as a society which can be improved by the addition of scale is going to see quite a dramatic AI machine learning driven step forward.
Speaker: I think that's clear, that's clear with
Speaker: what we're seeing with the pandemic that need to separate at scale.
Speaker: We don't ask a human being to look at every single pandemic test.
Speaker: I'm sure that's not how we're analyzing the data.
Speaker: So if you think about it, those kind of areas are going to be incredibly effective.
Speaker: But for most areas of the economy, and certainly in capital markets, it's going to be a gradual improvement in the way that we process data and think about data and develop tools to interact with data-driven scenarios.
Speaker: That's really where we're going to, that's the kind of pace of progress that we're going to feel.
Speaker: It's going to feel quite slow and it's going to be heavily dependent on people like Chris and Paris and their teams, people to really keep inputting, refining the way that we do operate these tools and how we develop these tools.
Speaker: So it's, you know, in some areas, the misconception is AI is going to completely revolutionize the world.
Speaker: And I don't really believe that it might do in 50 years time.
Speaker: And you look back in some areas of the world where we need to operate at scale and we're unable to do that at the moment because we need very skilled human beings to Paris's point, not children, cognitive ability, but very, you know, very high level of cognitive ability in those areas.
Speaker: where some of those tasks can be taken away from those people and you can get benefits from just scaled processes very it's going to be very very effective but in an awful lot of areas it's gonna be quite slow progress we'll look back in 10 years and say there are things we're able to do now which we weren't able to do 10 years ago which are almost unimaginable but it won't feel like that on a day-to-day basis i don't think all right that makes sense thank you everyone
Speaker: I'm going to combine two questions here.
Speaker: I'm a little worried I'm going to be leading the witness.
Speaker: So feel free to take either question if you like.
Speaker: But the two I'm combining is, first, how do you safeguard AI deployment to ensure we don't run afoul of market conduct and compliance standards is the first part.
Speaker: And the second part is, are there issues and or implications from the opaque nature of AI?
Speaker: So please don't feel that.
Speaker: that's the only direction to take it.
Speaker: But Chris, why don't we start with you?
Speaker: You can answer one, both, or however you'd like to take it.
Speaker: Yeah, I'll get quite a long question to sort of answer.
Speaker: But I think one of the very big part of this and what's really essential in the development in the AI and how we involve control is very much making sure that it's part of the process, that we have lots of very much domain specialists involved in that.
Speaker: So our end, you know, we directly work with this all the time with executing in equity markets, having the understanding of the business, the regulatory aspects throughout that, you know, is extremely important.
Speaker: And that's where we very much combine the two in that.
Speaker: I think some of the other very clear things need to be made is that when we develop these models, we explicitly define the objective.
Speaker: We define what rewards that the machine is going to get to be able to optimize that.
Speaker: What are the permissible actions?
Speaker: that the machine is going to be able to do you know what constraints do we put on that and i think within that aspect it's very important that through the development process that we are thinking about this regulatory aspects that you know that may affect that and we are defining the problem you know to be able to satisfy those as some simple examples things that often come up and get asked about is in our execution algos we could
Speaker: Say for instance, set up a problem such that some sort of big deep neural network will learn how to say for instance, spoof in a market by basically sort of saying, well, there's lots of, you know, in the execution, there's lots of autobook imbalance and if I put a lot of volume on the other side, I might entice the price to do something.
Speaker: But very much part of that construction, that's a function of what are you defining as your problem?
Speaker: What are you defining as your potential actions?
Speaker: And very much within us, it's very important that you have a lot of domain expertise that defines that problem appropriately and applies it to make sure that we're very much within those.
Speaker: And, you know, in a general, you know, from HSBC, we, you know, we have guidelines being developed on how we should develop these AI machine learnings.
Speaker: You know, we have...
Speaker: independent reviews on what the behavior of this is.
Speaker: So there's also very much, very large structure about how we control and how we ensure that these machine learnings can provide the performance, but make sure that they very much satisfy regulatory requirements.
Speaker: Yeah, I think I take on the second part on the dark side or the opaque nature of it.
Speaker: I would say a couple of things that the one that worries me the least, but still might be a concern is the fact that we have these tools and now my take decision, I have past take decision.
Speaker: Now, what they might output is the decision, but then people want
Speaker: sometimes understand and tell it, why did you take that decision?
Speaker: Can you tell me, explain?
Speaker: So then those two don't come out of the box on why they took the decision.
Speaker: So, but then we have to, there is a stream of work and AI that try to also add that narrative on why that decision try to explain in a more human-tale.
Speaker: So I'm not too worried about it.
Speaker: I think it's not there yet, but we are working.
Speaker: The one that I'm a little bit more concerned because the more conceptual constraint is the fact that
Speaker: is related to what Ed's saying.
Speaker: I mean, I fully agree, AI is great when you want to scale up your ability to do things, your capacity.
Speaker: But then that scaling up comes also with a trade-off, then all the decision will be then very much concentrated on one person that has decided that decision will be taken in that way and systematize across many people.
Speaker: So you lose a little bit of that diversity on when,
Speaker: Some of the diversity is bad because there will be variation on the doctors looking at the image, but also you lose a little bit of that diversity and the difference of opinion.
Speaker: And when the problem becomes complex and there is some judgment involved, then that diversity sometimes is good.
Speaker: So there is some area of AI that try to solve that.
Speaker: They call it exploration, exploitation.
Speaker: But I think conceptually that might be a problem, that you have that concentration on who makes a decision on how those decisions are made.
Speaker: and then scaled up to a million or billion people.
Speaker: I think that's an excellent point.
Speaker: It's that, it's where, does the analysis
Speaker: become too concentrated as well as the problem finding, if you like.
Speaker: And Chris touched on it as well.
Speaker: I think that that's the issue, right?
Speaker: Who was in the echo chamber?
Speaker: Were you in an echo chamber when you decided you were going to train the AI and machine learning tools to look at a particular decision tree in the way that it did?
Speaker: And making sure that you have the right contributors to that potential analysis is super important.
Speaker: Yeah, it's a great point.
Speaker: Thanks guys.
Speaker: And we appreciate your participation, Chris, Paris, and Ed.
Speaker: Thank you for sharing your insights.
Speaker: This has been the Markets and Securities Services Outlook, a podcast mini series produced especially for HSBC Global Viewpoint.
Speaker: To learn more about HSBC's Markets and Securities Services offerings, visit gbm.hsbc.com forward slash solutions forward slash securities dash services.
Speaker: Thank you for listening today.
Speaker: This has been HSBC Global Viewpoint Banking and Markets.
Speaker: For more information about anything you heard in this podcast or to learn about HSBC's global services and offerings, please visit gbm.hsbc.com.
