Transcript
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.
Speaker: Make sure you're subscribed to stay up to date with new episodes.
Speaker: Thanks for listening.
Speaker: And now onto today's show.
Speaker: This podcast was recorded for publication on the 9th of January, 2025 by HSPC Global Research.
Speaker: All the disclosures and disclaimers associated with it must be viewed on the link attached to your media player.
Speaker: And don't forget that you can follow us wherever you get your podcasts.
Speaker: Just search for The Macrobrief.
Speaker: Hello, and a very happy new year to all our listeners.
Speaker: I'm Piers Butler, and welcome to The Macrobrief.
Speaker: The hype around artificial intelligence was one of the big stories in financial markets in 2024.
Speaker: And it looks like it will be more of the same this year as the technology continues to take big leaps forward.
Speaker: So on today's podcast, we're putting AI through its paces.
Speaker: I'm joined in the studio by Mark McDonald, head of data science and analytics, who set AI a series of tasks, pitting it against humans, including me.
Speaker: We're also going to take a look at how the rapid growth of AI could affect the outlook for venture capital this year.
Speaker: So Mark, welcome back to the Macrobrief.
Speaker: Thank you very much for having me back on.
Speaker: So yes, pitting AI against the humans.
Speaker: We talked about a recent piece of work where you actually set AI against one of your team members.
Speaker: little bit in competition in a way.
Speaker: And there were some interesting conclusions on that in a way that it was quite positive, but AI could go in a complete tangent and so one had to be careful about some of the results.
Speaker: And actually it was the combination of AI and the human that produced the best results.
Speaker: Yes.
Speaker: So what have you done this time?
Speaker: Yes, this one's a bit different.
Speaker: Again, we're trying to see to what degree you can use AI practically in financial market research.
Speaker: This one is different.
Speaker: The previous case, we were testing out AI's ability to do data science.
Speaker: is it coming for my job?
Speaker: This time we kind of broadened out the scope and so is it coming for other people's jobs perhaps?
Speaker: And we set it the task of following the Fed.
Speaker: The piece was called Can AI Follow the Fed?
Speaker: Now obviously to some degree the answer is definitely yes.
Speaker: It would only take a fairly small amount of experimenting with feeding a Federal Reserve speech into ChatGPT to see that it can often do a very good job of this.
Speaker: But similarly, it would only take a little bit further experimentation to see that sometimes it does a much poorer job and sometimes you give it the same speech twice and it gives you completely different answers.
Speaker: And so really for something to be useful from a business perspective, you need to be able to engineer it into a sort of reliable and repeatable process.
Speaker: And so that's what we were looking at here.
Speaker: So we were trying to build an AI tool that we would give it a Federal Reserve speech and it would eventually analyze it along the directions of is it hawkish, dovish or neutral.
Speaker: It would give a summary of the monetary policy discussion that was in there.
Speaker: It would assess how relevant it was to the monetary policy outlook and it would explain its thinking.
Speaker: But rather than having kind of speech in and analysis out and trying to do it in one step,
Speaker: We instead split this into two steps and would rely on AI's ability to turn text into structured data.
Speaker: And this helps you to kind of corral the output, corral the output of the AI.
Speaker: in a more reliable manner, and it enables you to sort of engineer the task a bit better.
Speaker: So that's kind of why we went for this slightly more complex approach.
Speaker: So I guess, Mark, the big question is, did it work?
Speaker: Yes, I mean, at a high level, I would say the results are pretty good.
Speaker: We got Ryan Wang, our US economist, to do a compare and contrast and to analyze some of the same speeches himself manually.
Speaker: and to assess the outputs of the AI tool on the same task.
Speaker: And his view was that it was already performing at the level of a hard-working junior employee.
Speaker: And that is quite impressive.
Speaker: Despite this, he's not concerned that it's coming for his job just yet.
Speaker: And part of that is a function of the way that AI works.
Speaker: When you compare the AI performance to the human performance, what you see is, as I alluded to earlier, you can ask AI a question, then ask it the same question and get a difference.
Speaker: And it won't be the same.
Speaker: Yeah, exactly.
Speaker: So is this this issue of consistency?
Speaker: Consistency, yes.
Speaker: And it's one of the things that when people see an AI tool give what look like inconsistent answers, that worries people because we're used to the idea of computers being deterministic.
Speaker: Just so you interrupt, I was sort of listening to other podcasts about AI, in particular, self-driving cars.
Speaker: And everybody's terrified about self-driving cars because they can occasionally have accidents.
Speaker: Yeah.
Speaker: In a way, are we judging the AI too harshly because humans have lots of accidents in cars and actually statistically speaking, self-driving cars using AI are actually quite safe.
Speaker: So similarly, I mean, humans give inconsistent answers to data.
Speaker: So are we a bit too obsessed about this consistency point?
Speaker: Precisely.
Speaker: Obviously, the AI would give different responses on multiple runs of the AI on the same question.
Speaker: But we asked five of our colleagues to do a sample of the same analysis that AI was doing and it showed even worse variability from one human to the other than from one run of the AI to the other.
Speaker: And so I think what AI is doing a good job of is it's replicating human performance.
Speaker: The average output of the human and the average output of the AI were very similar.
Speaker: when aggregated over several instances, and the variability was comfortable from one to the other.
Speaker: However, we're fine with the idea that you can ask one human to analyze a piece of information and ask another human to analyze the same information and get different answers.
Speaker: And we're fine with that.
Speaker: But not with AI?
Speaker: Well, because if you ask the humans again tomorrow, you get the same answer.
Speaker: Right.
Speaker: And that's the key.
Speaker: It's the same person.
Speaker: And at some point down the line, if one of them capitulates on their view and changes, that capitulation is informative in and of itself.
Speaker: Whereas with asking AI to give you advice, that's very tricky because it will give you a sample of plausible advice from humans.
Speaker: And unless you're in the situation where you'd be happy to phone up 1-800-Strategist and get matched with a random strategist,
Speaker: then you're probably not going to be happy with an AI-generated opinion and AI-generated advice.
Speaker: And so I think that sort of use case for AI, it seems like these tools are particularly ill-suited to that.
Speaker: And that's not something that's going to change as the AI tools get better.
Speaker: It's not a performance issue.
Speaker: They're actually already performing very well at doing, on average, what good humans do.
Speaker: And also, strikingly, it's the speed at which they do it now.
Speaker: You're not going to answer this question because somebody else is.
Speaker: And you sort of did something that's a little unsettled to me ahead of this podcast.
Speaker: But maybe we should just let people listen to what you created by, again, using AI, which is an AI-generated podcast.
Speaker: Yes, this is a fantastic and free tool created by Google, which allows you to upload a document or even a series of documents and on the fly generate a fully AI generated podcast about the topic.
Speaker: Okay, so let's hear a quick clip.
Speaker: Think about it.
Speaker: You could use this AI to quickly analyze all the Fed speeches, all the reports, and highlight the most relevant points for you.
Speaker: I see where you're going with this.
Speaker: Right.
Speaker: That would be super helpful for anyone trying to stay ahead of the curve.
Speaker: Exactly.
Speaker: On monetary policy.
Speaker: Exactly.
Speaker: Imagine having this AI analyze every Fed speech in real time.
Speaker: Wow.
Speaker: Flagging potential shifts in policy before they even make headlines.
Speaker: Right.
Speaker: That kind of advantage.
Speaker: Yeah.
Speaker: Could revolutionize how we approach investing.
Speaker: Wow.
Speaker: I've listened to it already once before, but it's still striking as to how good it is.
Speaker: And also how enthusiastic they sound.
Speaker: Enthusiastic, clear, picks up on this point about real-time analysis of Fed statements.
Speaker: So very powerful when you think there's a whole army of people listening to the Fed.
Speaker: that potentially using an AI tool, you could do it quickly, you know, fast results that people could trade on.
Speaker: And in fact, I think there's another piece of work that you've done that takes that advantage.
Speaker: But before we get to that, is this the end of me and you as podcasters?
Speaker: That is a great question, because I think often when people see things like this technology with these AI-generated podcasts,
Speaker: An obvious question is, what's the point of that?
Speaker: Is the idea that eventually these podcasts that are generated by AI get good enough that you don't need humans doing it anymore?
Speaker: Maybe that's possible, but really I think the bigger change here is there is some minimum threshold below which there's a number of listeners below which you wouldn't bother doing a podcast.
Speaker: We'd be very upset if we only got one listener to this podcast.
Speaker: That would be seen as a massive failure.
Speaker: Whereas, you know, with an AI-generated podcast, it allows you to generate a podcast that literally only one person in the world cares about.
Speaker: You can upload a document that's only really of relevance to you.
Speaker: You're the only person who wants to hear that podcast.
Speaker: And that is fine.
Speaker: And this, in the same way that when, you know, Amazon was able to have this incredibly long tail of books that were available for people to
Speaker: to buy, that completely changed the structure and the economics of the book selling market.
Speaker: And there's a good chance that developments like this will completely change the economics and the setup of the podcasting market.
Speaker: Because it's literally a question of loading up the content and saying, produce a podcast.
Speaker: Yes, exactly.
Speaker: And at the moment, it's always the same two speakers that are generated by AI, and it only does it in English, but clearly all of that is going to change.
Speaker: So let's come back to following the Fed.
Speaker: So we can see it's a powerful tool.
Speaker: There are some issues regarding consistency, but in terms of the speed of analysis, your other piece of work seems to demonstrate that it can be used for trading strategies?
Speaker: Yes, definitely.
Speaker: So we did some analysis with our FX team, and they were looking at can they take the output generated by this AI tool and use it to help with investing in either the rates or the FX markets.
Speaker: And what they found is that, you know, there are signs that if you aggregate over a short-term window the signals from this tool on recent speeches, then it does a good job of tracking the aggregate hawkishness and dovishness of the Fed.
Speaker: And in some instances, that has led important changes in rates.
Speaker: However, there are many situations where
Speaker: The Fed and the market might disagree on the likely direction in which rates are going to go, and sometimes when that happens the Fed turns out to be right, and sometimes the market turns out to be right.
Speaker: So in and of itself that doesn't feel like a natural use case for a trading strategy.
Speaker: You need more than that if you want to use this at low frequency.
Speaker: what is reliable is looking at short-term price moves after the speech happens.
Speaker: So what they found was that looking at very short-term moves over the minutes following a speech in the FX upmarket, then if you follow the direction in which the AI tool said, whether it was hawkish or dovish, then that led to a reliable price move in the FX market.
Speaker: And I think speed
Speaker: is really of the essence there.
Speaker: These tools can process these speeches much quicker than a human, and that can be helpful from a trading perspective.
Speaker: Well, we'll see where that takes us, but I wanted to just come back to another piece of work.
Speaker: I mean, this is the regular survey that we do, the Funding the Future survey.
Speaker: I guess not surprisingly, in the latest edition that came out just before Christmas, AI is still a big factor there.
Speaker: There's a lot of enthusiasm.
Speaker: But the survey generally is,
Speaker: is unusual in a way that it's all the respondents, be it private equity and venture capital respondents or institutional investors, which we also survey, are all enthusiastic about the market.
Speaker: Yes, the overall sentiment in that survey in this most recent one seemed very positive.
Speaker: I believe it was 83% of respondents saw an increase in venture capital and PE activity.
Speaker: over the next year.
Speaker: So overall it was very positive.
Speaker: One of the big drivers in here is clearly AI.
Speaker: The most commonly cited tailwind given by survey respondents was the rapidly changing technological abilities, which is clearly related to AI.
Speaker: And AI was by far and away the most commonly cited sector in terms of likelihood of an exit and the areas which are most likely to see continued investment in coming quarters.
Speaker: What's interesting about AI is, in a way, one has to think about it as a little bit as a sort of cross-sectional type of influence, because one of the favorite sectors in terms of response for respondents was health care.
Speaker: But I guess AI is also influencing that because it's helping to be more effective in terms of research.
Speaker: Yes, AI is a critically important area in health care.
Speaker: partly for the drug discovery process, but also there are many other use cases where people are creating startups that apply AI technologies to other areas within healthcare.
Speaker: So, for instance, automated note-taking systems that can enable doctors to spend less time writing up notes and more time seeing patients.
Speaker: And there are automated generative AI tools that can help with putting the documentation together for the clinical trials that you're running.
Speaker: And there are thousands of use cases of AI within the healthcare sector too.
Speaker: It's an important sector there as well.
Speaker: Sometimes we ask a one-off question and we ask them the respondents what they thought about our nine themes.
Speaker: Which were the themes out of the nine that sort of came to the top?
Speaker: Well, the one that really stood out was disruptive technology.
Speaker: Not surprisingly.
Speaker: Exactly.
Speaker: I mean, the survey is called Funding the Future Survey, so that's a natural focus for them.
Speaker: Other of our key themes which were regularly cited were the energy transition and digital finance.
Speaker: Every time we've had a discussion on AI on this podcast, I've always been surprised by the speed of development.
Speaker: Do you think that momentum is maintained or does it start to plateau?
Speaker: I mean, there is clearly a lot of hype.
Speaker: That's unambiguous.
Speaker: But there's a lot of real meat there as well.
Speaker: And so I think even if the capabilities of AI improved no further today, even if the models that we have today, that turns out we can't build models better than that.
Speaker: I don't believe that's going to be the outcome, but let's, for the sake of argument, say that that were to happen.
Speaker: I believe there's still five to ten years of real productivity benefits that come as people work out how to really embed these new technologies into business practices and make the efficiencies that will result.
Speaker: So I'm very bullish on the impact of AI.
Speaker: I think the technology has accelerated much further ahead of the ability for people to use it.
Speaker: And do you think maybe to finish on, is this a ubiquitous benefit for society or some people are going to have an advantage that they really have for a while?
Speaker: So I think as we discussed in our earlier work about AI, the good, the bad and the ugly, there are definitely ugly sides to AI.
Speaker: And one of the ugliest is probably the distributional aspect.
Speaker: It would
Speaker: AI is leading to rapid change.
Speaker: It's impossible for the change to be beneficial to everybody.
Speaker: There will be some winners and there will be some losers.
Speaker: And the more rapid the change and the harder it then is for society to adjust to those, the more disruptive that could be.
Speaker: The U.S. is clearly the area that's dominating at the moment.
Speaker: They do seem to have a clear advantage there, not just in Europe.
Speaker: having the tech but also when you look at the companies that are reporting that they're practically using it a lot of those are concentrated in the US as well so it's not just on the tech side it's also on the implementation side that they're doing well but there is a massive open source AI movement it's not just the models that are
Speaker: offered as APIs by the big tech companies, there are a lot of very powerful sort of open source AI models, not just from the US, some coming from say China as well.
Speaker: And so I think the technology is there for everybody to use.
Speaker: So I think it could be very beneficial.
Speaker: Brave new world.
Speaker: Indeed.
Speaker: Well, Mark, I'm sure we'll have you back on the podcast.
Speaker: I hope it's me interviewing you.
Speaker: Me too.
Speaker: I hope we're not both replaced by avatars.
Speaker: But for now, thank you very much.
Speaker: Thanks for having me back on.
Speaker: If you'd like to catch up with the latest views from the team here at Global Research, then check out our new 2025 HSBC House Views report.
Speaker: This multi-asset report outlines our thinking about global economics, currencies, fixed income, emerging markets, multi-asset strategy, equity strategy, global trade, sustainability, and commodities for the year ahead.
Speaker: And with the world bracing for the impact of proposed changes to U.S. policy, our chief global economist, Janet Henry, will be taking your questions at our next live insights.
Speaker: The event takes place on the 14th of January at 9.30 UK time.
Speaker: Anyone can join.
Speaker: You don't have to be an HSBC client.
Speaker: Full details on how to sign up are on our LinkedIn page.
Speaker: Just search hashtag HSBC research or email us at askresearch at hsbc.com.
Speaker: So that's all from us today.
Speaker: Don't forget to follow the macro brief wherever you get your podcasts.
Speaker: Thanks for listening and we'll be back next week.
Speaker: Thank you for joining us at HSBC Global Viewpoint.
Speaker: We hope you enjoyed the discussion.
Speaker: Make sure you're subscribed to stay up to date with new episodes.
