Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
The Macro Brief – Where next for AI?
155 Plays1 hour ago
Can AI trade FX? Is the future 'agentic'? And are we seeing a bubble? Mark McDonald, Head of AI and Data Science, and Yuning Bai, Data Scientist, discuss recent developments in artificial intelligence.
Click here for appropriate Disclosures, including analyst certifications, and Disclaimers that must be viewed with this podcast: https://www.research.hsbc.com/R/101/DkgSvdV
Stay connected and access free to view reports and videos from HSBC Global Investment Research follow us on LinkedIn https://www.linkedin.com/feed/hashtag/hsbcresearch/ or click here: https://www.gbm.hsbc.com/insights/global-research.
Recommended
Transcript
Podcast Introduction
00:00:01
Speaker
Welcome to HSBC Global Viewpoint, the podcast series that brings together business leaders and industry experts to explore the latest global insights, trends, and opportunities.
00:00:13
Speaker
Make sure you're subscribed to stay up to date with new episodes. Thanks for listening, and now onto to today's show.
Welcome to The Macrobrief
00:00:33
Speaker
Hello, I'm Piers Butler in London and welcome to The Macrobrief from HSBC Global Investment Research, where we look at the issues driving financial markets across the globe.
AI's Impact on Financial Markets: 2025 & Beyond
00:00:42
Speaker
Now, the hype around artificial intelligence has been one of the big stories in financial markets in 2025, and it looks like we're set for more of the same in 2026. So on today's podcast, we'll be catching up with the latest developments around the technology and asking, could it be capable of making accurate investment decisions?
Meet the Experts: Mark McDonald & Yuning Bai
00:01:01
Speaker
To help us find out, I'm joined in the studio by Mark McDonald, head of and data science, and making her first appearance on the podcast is data scientist Yuning Bai. Mark and Yuning, welcome.
AI Applications in US vs. European Markets
00:01:10
Speaker
Thanks, Pais.
00:01:11
Speaker
So Mark, back in March, we had you on the podcast talking about AI, and the title of the episode was, How Well Do You Understand AI? I guess given the continued pace of development, the question is still relevant.
00:01:22
Speaker
Is it fair to say that practical applications of AI are amongst us? I mean, I'm using AI models in my day-to-day life now. So maybe let's recap on how far we have actually come in 2025. Yes, of course. I mean, I think the question of how well do you understand AI is going to be ah an evergreen question um in that the technology is developing so quickly that I think a lot of people's intuitions for know how it works and what it can do well are sort of often playing catch up with reality. In terms of the you know the degree to which the sort of implementation of AI is really here, I think it's very much coming of age.
00:01:57
Speaker
When we look at what companies are saying on earnings calls, you do see many companies actually talking about practically implementing AI. We see this particularly in the US. the The adoption rate there is much higher than, say, their peers um in Europe, and that's not a sector
AI & Stock Performance in the US vs. Europe
00:02:13
Speaker
composition thing. This is the same when we look sector by sector and do the comparison. there's just much more active implementation of AI into business processes happening in the US compared to Europe. And as we heard, was it two, three weeks ago on this podcast from your colleague Shiva Jun, companies, quoted companies are being rewarded for this, aren't they, in terms of their share price performance? Yes, absolutely. So if you look at a basket of, say, the S&P 500 companies who are actively implementing AI into their operations, and compare this to to their peers, you can see that they have been outperforming for for some time now. And this suggests that the market is giving these companies the benefit of the doubt. Very few of these companies are saying you can see this in our reported numbers. You can't measure this in the traditional financial analysis. But the market is giving these companies the benefit of the doubt in the US.
00:03:04
Speaker
In Europe, we're not seeing the same outperformance. At least not yet. Not yet.
AI in FX Trade Recommendations
00:03:08
Speaker
so I suggest that the market is much more s skeptical of what's happening in Europe. Anyway, so your team continues to test the boundaries of what AI can do, and recently you put various AI models through their paces to answer the question, can AI trade FX? How did you approach this? What was the methodology?
00:03:24
Speaker
Yeah, it's quite an interesting piece, this, because you know from the title, can AI trade FX, it might might sound a bit silly because we've had you know electronic market-making systems in the foreign exchange market for years now, but that's not what we're talking about here. We're not talking about automating the market-making process. We're saying can AI reasoning models reason coherently about financial news and then come up with directional trade recommendations for the foreign exchange market. And in terms of all the AI experiments that we've done so far, this is by far and away ah the hardest task that we've set AI.
00:03:58
Speaker
Because what we're really asking AI to do here is to independently generate alpha. It's a very hard challenge. So Mark, so how in practical terms did you do that? So what we did is we took the um HSBC research publication, the America's FX Morning Bullets, which contains sort of an overview of all the news that's happened overnight for the US market that's relevant for the FX market. And we chose this because it's in the same form every day and it's already been pre-selected by our human experts in the FX team as being relevant for the FX market. And we would feed in this news information into these reasoning models and ask them to reason about the macro implications and then suggest directional trade recommendations. This is a very hard challenge. And in some ways I think that's fair.
00:04:43
Speaker
um Humans would struggle at this task too. Even FX specialists would struggle to reliably generate directional trade recs purely based on overnight news. And so I think probably a fairer assessment of of what we did was when we got our human specialists in the FX research team to do a sort of subjective analysis of the quality of the reasoning. How good was the trade rationale? And then on the basis of this, we're able to you know score how well the different models did on this subjective basis.
Performance & Biases in AI Models
00:05:12
Speaker
So, Yuling, how did we do that in practice? So in terms of this, ah we ask our different models to spit out um different outcomes from the FX morning bullets. And then we ask our FX strategist to choose which one that they prefer. And all of these answers are grouped in pairs, so they choose between the two. um The reason why we're doing this instead of asking them to just score it um objectively is that a human can do a terrible task at trying to score objectively. It is almost always subjective. And this is not some new and wacky thing that we do.
00:05:44
Speaker
This is actually at the heart of modern AI training process. This technique is called direct preference optimization. Essentially, it is coming up with the ranking that match perfectly with the observed human preferences. And based on that ranking, we can observe that GPT-5 is everyone's favorite model. And meanwhile, other smaller and cheaper models perform less well.
00:06:06
Speaker
But the bottom line conclusion of all this work is it can't really predict effects movements or AI can't really predict ai movement Or you need a human. Thankfully, still.
00:06:18
Speaker
Yes, because like with this pairwise ranking process and looking at the human preferences, you can see that there are some models that did much better than others. So as Youning mentioned, GPT-5 was by far and away everyone's favorite, and Claude also did very well here.
00:06:31
Speaker
And so what this suggests is that from the perspective of like pure and reasoning quality as judged by a human, then you kind of get what you pay for with these models. But what you find if you look more deeply at some of these models is they have some quite strange and quite worrying and dangerous behavior. Yes, that's right. There are some really super strange behaviors going on here. like There are some totally mad behaviors. Give us an example. Okay, so the strangest one has to be something we saw with Glock, where we would see instances where it would suggest both buying and and selling the same currency pair on the same day. so you're hedged against all profits. Good to volume I guess but not necessarily for the investor. But what we would what we would find here is um although it looks really strange was actually quite a positive development for for AI models in that you would see something where it would say okay I expect to see a Swedish krona weakness so I'm going to sell Eurostocky and they would say, oh no, wait, oh my god, I've got it the wrong way round. And then it would go on a long tirade, berating itself for getting it the FX quotation the wrong way round. And then the immediate next trade recommendation was the right way round with a sensible rationale.
00:07:38
Speaker
What we find here is that you know if you go back even just a year, if you went to an AI model and asked it to do something and it went down the wrong path, it was almost impossible to correct. You'd say, no, no, that's not what I meant, and it would say terribly sorry, and then do the same thing again. um And now here you've got a model that spotted its own mistake and corrected itself. So i think this is actually quite a positive development.
00:08:00
Speaker
But in terms of the the sort of more worrying behavior, it's really to do with the sort of the biases that the models have learned. And these models kind of bake in biases, which mean that they're They're very dangerous to rely on using them for high stakes decisions. Perhaps not surprisingly, one of the clear biases was the US dollar?
00:08:18
Speaker
Yes. And so this is one of the the weird things that, you know, we only really noticed it this at the end when Yuning was doing an analysis of all the trade recommendations in aggregate. And what you find is that for all the models in all the time periods, there's this massive imbalance between the number of times it suggests buying the dollar versus selling the dollar. And so GPT-5, which was the one that our human experts thought worked best, and had the worst imbalance. It was over four and a half dollar buys for each dollar sell in each time period, including 2025, when the dollar had been weakening throughout the um the first half. So it suggests that what's happened here is that these models have learned this very significant bias, which is also very subtle, because when our experts were looking at the trade recommendations, they were never saying, oh, this trade recommendation is really biased, it's cherry-picking all the evidence, it's' it's a bad recommendation. The recommendation appears coherent. Yeah, on a standalone basis, it looked all right. yeah when you sort of aggregate the whole thing, suddenly you're discovering this overall bias. Exactly. And this is why it's pernicious and dangerous, because you know you might think, well, I'm an expert in this market, I can... i can see when it's making bad decisions.
00:09:24
Speaker
But you probably can't. And these models will have learned similar biases towards and away from all sorts of assets and
Agentic AI: Potential and Limitations
00:09:30
Speaker
asset classes. And if you rely on them to make your decisions for you, then you get those biases baked into your trading results. So we think it's much preferable to have humans in charge of the high-stakes decisions and then have AI as tools to help you with your research. And they're amazing at that. These tools are fantastic at at this.
00:09:46
Speaker
Now, there is a new buzzword in town on AI, which is agentic. AI. What's that? So agentic AI is clearly the huge new buzzword, not just in the AI sphere, but also from company management.
00:09:59
Speaker
These are ai models that can autonomously make their own decisions and plan when given a big complex task. And the reason that companies are excited about this is that for most people's roles, it's not just a random collection of tasks. A lot of the role is actually in that messy glue in the interplay between all the different tasks that you do. With agentic AI, if it delivers on its promise, then you know you could have agents that you give big tasks to, and they will go and independently work towards achieving those aims. And this could be a a big productivity boost.
00:10:32
Speaker
And you say in one of your recent reports that agentic AI is in a similar position to gen AI was two years ago, which sounds really promising, but what do you base that conclusion on?
00:10:45
Speaker
So it's really based on, there's been some interesting new research looking at how agents perform. And the the way that they did the analysis was they were comparing how agents did a task and how humans did tasks and then you know comparing and contrasting which bits agents did well and which bits humans did well. um The thing which seems to work best at the moment is the human-agent combination, um where humans can sanity check things and do the things that the agents find challenging, which is often using sort of graphical information and visual information, but also stopping the agents from lying. Because the agents, when you autonomously let off and go do their own thing, that the truly agentic systems, which have just been trained by reinforcement learning, um they very heavily penalize an agent for not making it to the end of a process. And as a result, to avoid that huge penalty, if it gets to a step where it's blocked, let's say it needs to analyze some data for some files, if for some reason there's an operating system problem and it can't access the files, instead of saying, terribly sorry, I can't do this, it will say, okay, um here's a JSON object containing some information I've extracted from these files and pass it to the next step, just completely fabricated. And if you include humans in the process... So that be quite dangerous. So this is quite dangerous. And i think this is very akin to me to the like hallucination problem that we had early on with generative AI, which has got a lot better. And in the early days of making this work, companies would spend a lot of time scaffolding ah code processes around generative AI and breaking big tasks into smaller tasks and guiding it along a fairly carefully controlled path in order to make it safer. And I think we probably need either companies to do that sort of work with agents and sort of constrain them, um or the agentic tool calling needs to be better and safer and more reliable. But I think you know it's very much at that stage where companies were two years ago with generative AI, where it's like a very exciting technology with huge promise, but still not really quite sure how to bake it into existing business processes. Nevertheless, it sounds like agentic commerce
00:12:52
Speaker
has the most promise and is already to some extent in practical applications. I mean, I feel like I've already experienced in going to certain retail websites where there is an AI model that's trying to help you with your questions and what have you. And often I find it goes round and round in circles because my question is quite complicated. But in a way, that's not the point because in reality, it's already dealing with a lot of the very basic queries.
00:13:14
Speaker
Yes, so I think it does have a lot of promise. I think it's still very early days though, if we're being honest. But most of it at the moment is in companies that are setting up the sort of fairly embryonic agentic payment rails. So you've got companies like Google who've come up with their agentic process. So what's a payment rail?
00:13:32
Speaker
So the way of safely guiding um the way that payments can can go without them going off the rails and sending money to the wrong place. Oh, I see, okay. um And so it's really protocols to make sure that your agent um that you're asking to buy something on your behalf actually does buy the right thing and doesn't come back with 72 top hats and a bowl of soup. You know it' so you need to make sure that the processes can be reliable. And given the challenges I mentioned earlier with agents having a tendency to make things up if they're under pressure to get to the end of the end of the process, you don't want your agent that you've asked to go off and buy one thing, sort of getting stuck in a loop and buying it many, many times or buying the wrong thing or paying way too much money for something. And so having these protocols that allow agents to negotiate safely with ah with a retailer and make sure that the the transaction that happens is one that the user is happy for to happen, even though the human's not in the loop, this is like a big safety issue that people are working to solve now. And so I think they're doing this based on the promise of what they're
Debate: AI Investment Returns
00:14:30
Speaker
expecting to see. So is the holy grail that as a consumer I could have an AI agent where say, well, I'm looking for a pair of trainers and the AI agent will basically do the work for you in terms of finding what you're looking for and the best price and what have you?
00:14:45
Speaker
Yes. I think it's probably most likely in the consumer discretionary space yeah like that, where you've got kind of things where, you know, it's quite challenging to know what the vis what the right trainers are that you might want to buy and there are lots of different options. is It feels less likely that you're going to have agents go off and do your supermarket shop for you because you've you've already got an app at your supermarket that yes that can quite a good Although if you can tell me where I can find a particular item in in the supermarket when I'm doing the shopping because they keep changing it. They keep changing the layout, yes.
00:15:14
Speaker
ah but Maybe the model will keep up to date on that. So promising but... still But I guess if our experience of AI, if we've learned anything from it, is that a lot of the challenges that you currently highlighted are likely to be overcome quite quickly, possibly more quickly than people imagine.
00:15:33
Speaker
I suspect so, yes. it's yeah we've We've solved similar problems to this before in some of the earlier rounds of generative AI, so I'm quite confident that these these challenges will be solved.
00:15:45
Speaker
So let's finish on an update to this report that was entitled Breaking Bad Narrative and which was referring to looking at the AI bubble story versus the causal evidence. unit You've just published an update on that, I think, called Beyond the Bubble. Are we beyond the bubble? Yes, yes, it did. So in terms of this report, um I think it's also focused on the fear of the AI bubble, the growing amount investors. And um there's been this report that's published earlier this year, I think it's around July, um that is published by MIT Nanda. and This report stated that 95% of the organizations have seen like zero returns among their Gen AI investment. So this is definitely a very astonishing number.
00:16:25
Speaker
Ever since this report was published and this ah constant there's been constant debate of is there an AI bubble, and often this 95% number is constantly getting quoted. um But we do think that we need to take this result with a pinch of salt. And in this specific report, we refer to this recent study published by Wharton and GBK, is that they focus on also the return of investment of Gen AI. And they observe that 74% of the companies have seen measurable benefits from AI implementations. And this is definitely a very strong counterpoint to the bubble narrative. And of course it is very clear that it is impossible that this success rate of AI implementation can just rise from 5% to nearly 75% in such a short time, three months since the MIT report has been
The Evolution of AI
00:17:20
Speaker
published. So that kind of tells us it's it still very difficult to kind of define and accurately measure what is the ah ROI of this Gen AI implementations in general.
00:17:33
Speaker
ah But we do believe that this Wharton report is more reliable because this is a study they've been carrying on for three years ever since GPT has been published. And they've been tracking this different performance from different companies about this amp implementation. And this is the third year they've been conveyed this survey in the same methodology.
00:17:53
Speaker
So we believe that the levels of AI success is still open to debate. But in terms of um this survey the changes in surveys comparing to the next one, it's definitely more reliable.
00:18:05
Speaker
So Mark, you and I have been around long enough to remember the internet period and all the excitement. And the reality was that a lot of what was said at the time about the potential benefits of the internet did come through, but it took a very long time. Why is it that this time around it seems to be happening almost as we observe it, you know, that some of the some of the promise of the eye that we talked about, I mean, we've had many conversations, including on this podcast about this.
00:18:31
Speaker
And, you know, we've often had this sort of conversation about the speed of development. And every time we speak again, it seems that that speed has just been maintained or and indeed accelerated.
00:18:42
Speaker
Yes, but I think what's going on here is actually this AI success has been a very long time in the making. So there have been multiple waves historically of AI excitement that then didn't quite achieve what people were expecting. And then you'd have this sort of AI winter where everyone was disappointed for some time afterwards. and i would say this current wave really started back in 2012 when there was an image recognition competition that was won by a sort of neural network based model. And it did so much better than everybody everybody else's model that this reignited the sort of excitement about neural networks. And neural networks are at the heart of modern modern AI models. And so really, it's been this slow and steady progress since that point. But it's the. It's just that most of us were unaware of it. Yeah, and it's the nature of exponentials that you know they they go very, very quickly the later you go. Yeah. And so you have this what looks now like very slow progress from 2012 to 2022, setting the groundwork for all these language models and all these image models that we now take for granted. And they appear to be like, oh my God, they're like developing so quickly, but they're only possible because of this like decade or more of preparation work that happened
Closing Remarks & Staying Updated
00:19:58
Speaker
behind the scenes. So it's not a flash in the pan. It was a really strong buildup but to sort of lead to where we are now. So it does feel like we're going to be talking again in 2026. I suspect we will be back on the podcast, yes. But for now, Yuning and Mark, thank you very much for joining me. Thank you very Thanks having Great to be on.
00:20:18
Speaker
That was Mark McDonald and Yuning Bai on the latest developments in AI. If you're an HSBC client, you can keep up to date on our latest research by downloading our mobile app. The app features all our key reports, videos, and podcasts, and can be downloaded from Apple's App Store or Google Play.
00:20:35
Speaker
Also, don't forget to check out our sister podcast, Under the Banyan Tree, where hosts Fred Newman and Harold van der Linde put the region's markets and economics into context. And if you've got any questions or comments, then you can get in touch with us at askresearch at hsbc.com.
00:20:51
Speaker
So that's it from us today. This episode of The Macro Brief was hosted by me, Piers Butler, and produced by Tom Barton. Thanks for listening, and please join us again next week.
00:21:27
Speaker
Thank you for joining us at HSBC Global Viewpoint. We hope you enjoyed the discussion. Make sure you're subscribed to stay up to date with new episodes.