What Happens When AI Starts Building the Portfolio ft. Nick Baltas

[00:00:01] Intro

Welcome to Top Traders Unplugged. In markets success doesn’t come from predicting what happens next, it comes from being prepared for what you can’t predict.

In each episode we go deep with some of the world’s most thoughtful minds in investing, economics, and beyond to understand how they think, how they prepare, and how they decide, and the experiences that shaped how they see the world. No noise, no short-cuts, just real conversations to help you think better and invest with confidence.

[00:00:32] Alan

Welcome back to Top Traders Unplugged where each week we take the pulse of the markets from the perspective of a rules-based investor. My name is Alan Dunne and this week I’m sitting in for Niels who’s away on his travels and I’m delighted to be joined by Nick Baltas. Nick, how are you doing?

[00:00:50] Nick

I am doing well, Alan. I’m doing well. How are you?

[00:00:53] Alan

Good. I know you’ve been traveling as well, so there was a bit of tension earlier, whether you going to make it or not, but you’re here in good shape.

[00:01:00] Nick

But I’m here. It’s been quite a good September, obviously back to school, back to work, it’s been very busy – the kids going back to school and my son starts nursing, all that lot. So, it’s been a busy month, I must say, but all very good, all very good.

[00:01:14] Alan

Good. Yeah, well, it’s busy and a positive month for trend following, so that always helps as well. So, everybody in the industry, I’m sure, is in good spirits. But we do always like to kick off with asking our guests what has been on their radar? What has been on your radar in the last number of weeks?

[00:01:31] Nick

So, I don’t know that it’s something in my radar, but certainly something that just caught my attention, and I don’t think it caught just my attention, I think it caught various people’s attention and it was part of a small article on Nifty, a couple of days ago. This comes from GS, from Christian Mueller-Glissmann. He’s running our kind of multi-asset strategy on the research side, and he put out a chart which is very, very simple but actually quite striking.

So, what it shows is that, sorry, it shows basically the percentage of S&P 500 constituents (so, 500 stocks in the index) that have a negative beta to the index that they belong. So, naturally you would think if I belong to an index, then more likely I have a positive beta to it, partly because I’m part of the index, obviously at smaller sizes you would expect that this will become less direct, but also because there’s a significant beta factor that is driving assets and stocks in this particular case. So, using like a three-month calculation point in time, it’s 45%. So, about half of the stocks in S&P 500 have a negative beta to the index which I think is just mind-blowing. Last time that we had this situation was in.com, but at the time the number was just a tad less than 20%. It’s now almost 45%.

And obviously the background here is the fact that S&P is a cap-weighted index. The AI dominance is effectively building its own kind of mini index within the index and all the rest are becoming less correlated to it. So, this clustering and concentration is basically causing this effect. But again, that in itself is quite a striking image once you see it. I know you’ve seen the pictures.

[00:03:21] Alan

I’ve seen the pictures.

[00:03:21] Nick

So, you can relate, right?

[00:03:22] Alan

Yes, I mean, it shows a dramatic… well, what looks like an acceleration on the chart. I mean people can check it out. It was in the FT, so you’ll see it online or check out Goldman’s research. I mean is energy probably is driving it to an extent as well, I guess?

[00:03:41] Nick

It could be. I mean there is not too much kind of discussion in the actual paper from Grishman on that part aside from the fact that the market is getting concentrated and there’s a lot of discussion as to whether bonds are defensive in this current environment and obviously this decoupling in the equity space is creating opportunities for dispersion. Some people talk about going long correlation, and so on, and so forth.

So, that’s kind of the background under which this was kind of projected. But that, in itself, even out of context, is quite a chart. So, not necessarily on my radar but let’s say something that caught my attention in the last couple of days.

[00:04:11] Alan

So, noteworthy for sure? Yes.

Wel,l on my radar, I mean, probably similar in the sense of more noteworthy I would say… I’ve just been following the debate in markets and amongst analysts about the rise in yields. Obviously, everybody’s been talking about the bond market, the bond markets around the world for the last number of weeks, but it just really strikes me it’s quite heated debate, at times, between analysts as to what’s driving this rise in yields, as to whether it’s… Is it economic growth, is it fiscal concerns, is it that people expect rates to be higher on average over time? Is it the competition for capital and is that being driven more by budget deficits or the AI spend? Is it the debasement credibility issue?

So, all of these things, and everybody has their own view. I mean it is striking from my perspective. I mean, I do think the fiscal story is there in the background but at the same time what’s really struck me, in the last couple of days, just looking at it, is the big repricing at the short end. Like the two-year yield has gone up 75 bips, or maybe a percent, at least 75 bips, I would say, since the middle of August. So, there’s been this… And also, at the short end, around the world has repriced. So, it’s as if everybody now has taken a much more bullish view on global growth or else sees the conflict in Iran staying on much longer.

But I suppose my perspective on this is this kind of tendency amongst analysts to view the fundamentals and then the market reaction as being very much linear and proportionate as opposed to markets are complex adaptive systems. A given change in the fundamentals at one time can have a big impact, other times can have a small impact. Marketing position matters (as we all know), the market dynamics, do people have barriers being hit, hedges, etc., all of that seems to get removed and people are trying to find this very linear cause and effect relationship which I don’t think exists.

But I’m curious to hear what’s your perspective on yields? Do you lie in bed at night wondering why yields are rising or do you just take it as face value that they are trending higher?

[00:06:30] Nick

I mean I’m looking into my mortgage at the minimum and I’m trying to project what’s going to happen the next time we have to do the mortgage. But I guess jokes aside, I mean some of them also have been quite substantial.

To your point we’ve seen some kind of multi-year or perhaps decade records in the US in the UK as well. And even six months back… Is it six months? Maybe six months back analysts were projecting rate cuts. For those that have forgotten, that this was the case that was back in Feb or January. So, it’s not that long ago that the projection was that we’re entering into kind of this normalization stage whereby central banks will start kind of easing and all we’re seeing now is actually yields reaching those kind of levels that we haven’t seen for some time.

I don’t have a strong view, frankly speaking, and I wouldn’t claim any particular expertise in that space. The one thing I would say, and that was like a quote I heard recently ,and I cannot recall exactly where this came from, but it was more about, it is very hard to predict not only because prediction is hard, but sometimes you cannot even tell today’s situation. To your point, it is a struggle to understand, point in time, what is actually happening in the market, let alone predicting the markets.

Obviously, that was not a quote that quants would be extremely happy about, on prediction, and so on, and so forth. But I think it’s painting the picture of how complex markets are today. Whether we see geopolitical tensions kind of running ahead of us, whether that would have inflation implications, whether the inflation is more of a reaction to it or the cause of it. At the minimum, I think rates have been the strongest contributor to trend following strategies.

[00:08:20] Alan

Absolutely. Yeah.

[00:08:20] Nick

I would imagine most programs will be short and benefiting quite substantially month-to-date and year-to-date.

[00:08:27] Alan

Yeah, well that’s a good segue. We do always like to give our performance update and as you say it is a good month for trend following.

SocGen Trend up about 3.7% and SocGen CTA Index up about 3.7% as well. And year to date both SocGen Trend and SocGen CTA up about 15.2%, 15.25%. They’re pretty much neck-and-neck. And as we were chatting, just before we started recording, that monthly number probably understates it a little bit because it’s up to the 23rd and yesterday was also a pretty good day.

So, [it’s been a] very good month, very good year. And as you say certainly this month looking, certainly, across the managers that we allocate to it very much, I’d say 80% of the cases concentrated in fixed income short positions. And at the long end as well, obviously, as I mentioned, there has been a very big repricing. So, interest rate futures, 2-year yields have had a big move as well as I suppose the more publicized move in bond markets.

And I think this is often really where people see huge value in trend following as an element in a portfolio. Because if you think about it, if you’ve got a 60/40 portfolio and you’ve taken out some of the allocation and added it to trend following, that trend following exposure now is short bonds. So that bond short is basically taking down your core long position in bonds that you would have had. And by adding trend to your portfolio you’d probably be net short bonds in this move up.

So, that’s the kind of environment where you really see the value of it as well as also, in August, we had more long positions in commodities that were contributing. But certainly, this is where you’re seeing that kind of dynamic element of trend following exposure; what it can do to a multi-asset portfolio, which I think is very valuable.

It’s been a good month for you guys as well. I think you were saying performance wise?

[00:10:36] Nick

I mean it’s been a very good month, as you say. I think it’s one of the stronger months that we’ve seen in a number of years probably.

I think January was very strong this year and obviously the month that we went past, August. I mean the rates have significantly outsized the contribution from anything else. Commodities and equities, obviously, come second and third but at some margin, which is basically driven by this month’s activity. So, said differently, the beginning of the month rates, commodity and equities were basically kind of equal contributors to performance. And now that gap has widened by some amount precisely because rates have doubled their contribution. So, it’s been quite substantial.

Currency is a bit more of a mixed bag, I would say, year to date, nothing very extreme in either direction. I think with the dollar moves we’ve seen there’s a bit of a lack of trendiness, which was the case by the way for rates until maybe what, Q2? We hadn’t seen, actually, rates kind of trending. It’s been more of the last couple of months activity which, I guess, that’s the other benefit of the diversification. Different contributors in the asset class at different points in time.

But that’s a very good month and a very good year so far. I think it’s the third best, at least from some of the stats I have, I think it’s the third best so far year in the last 15. So, 2014 obviously with a dollar…

[00:12:03] Alan

Yeah, 2022.

[00:12:04] Nick

2022, and then this year, so far so good.

[00:12:05] Alan

So far so good. There’s always the last day of the month or the last day of the year to navigate. But as you say it has…

[00:12:17] Nick

It has been a good month.

[00:12:18] Alan

It has been very good. And I think you make the point as well. I mean, yields had been in a range. If you look at the 10-year yield, it had been a 3.5% to 5% range since October 2022, whipsawing, it had been very difficult. Obviously, the market had priced in a recession probably three times since 2022 and now has flipped back towards a much more positive economic growth scenario. And now we’re starting to see that move accelerate. So, interesting to see how that evolves.

Did an interesting thing that struck me, I just looked at it a few days ago, so I’m not sure if it’s moved much with the more recent run up. But volatility in the bond market has been relatively contained. It’s been very stable. I mean the magnitude move has been sizable and, obviously, I think when you breach a psychological level like 5%, people tend to get very excited. But at the same time, from a volatility perspective, we’re not seeing rising vol in the bond market.

[00:13:21] Nick

No. I mean and I’m seeing that also in the strength of the signals which are typically volatility scaled. I know they haven’t basically shrunk or let’s say they haven’t actually moved… sorry… they haven’t stayed flat.

You could argue that the yield move you see, if it will come at a higher vol the net effect vol adjusted is like flat. We haven’t seen that the case. What we’re seeing is strengthening of those signals which attests to the point that the move has been greater than the volatility at which it comes along.

[00:13:52] Alan

Yes, yeah, obviously, and as people maintain those positions, the P&L attribution is that much greater than if you were seeing a very choppy and volatile run up. So that’s favorable too.

[00:14:07] Nick

Less equities because equities have been a bit more challenging for the month. Right? Because you came into the month long and then, obviously, most of the markets have kind of retracted. Not by a significant amount but that’s not like a positive month for equities, not too negative either. I think like less than a percentage, for example, I’m seeing some stats now.

[00:14:26] Alan

And commodities had been a standout in August?

[00:14:29] Nick

Mixed bag, a mixed bag for the month. So, you have some, obviously, some of the energy markets kind of performing but then some of the kind of more ags and softs have not actually delivered any performance or maybe detracted some performance, like flat for the month. It’s actually quite striking that the month was like, flat, flat, flat except for ags, I should say.

[00:14:49] Alan

Yeah, well that’s pretty much…

[00:14:50] Nick

That’s the story by the way. That’s the only story. It’s a rates story.

[00:14:55] Alan

That’s about it. 2022, I guess was a rates story, but it was also a dollar story as well. So, that accentuated.

[00:15:00] Nick

It was a bit of commodity story in Q1 as well with inflation moves and, obviously, the geopolitical tensions as well.

[00:15:13] Alan

I want to move on to some of the topics that we had prepared. And, I suppose, on the theme of that mix of volatility and trend and how to condition trend with volatility, there’s a paper, a research note that came from the guys at Alpha Architect recently around the whole area of whether you can enhance a trend system or a momentum model by looking at the volatility environment. So, basically, the idea of trading faster in higher vol markets or trading slower in lower vol markets. You had a chance to go through the paper?

[00:15:56] Nick

Yeah, I did take a look and thanks for pointing it out. I mean the results are actually quite straightforward and the methodology, I have some kind of question marks around it. But, as you say, the punchline is should you adjust your kind of lookback window as a function of vol? If vol increases, should you go faster? If it drops, should you go slower? Can I reduce some of the false positives? Does that bring any value? That’s kind of the high level question.

But if you like the approach, it is actually very simple and there could be nuances related to the results that come from the design choices, and it’s unclear to me whether we can generalize, but to kind of go a bit step by step, they’re looking into… And by the way, the whole idea is that this was a research paper done in September 2017. So, they wanted to now extend with nine new years, if you like, in the sample, and see whether this is kind of VIX dependent. They use VIX as a measure of all VIX dependent lookback windows bringing more value vis-à-vis a static window.

[00:17:06] Alan

Yes.

[00:17:06] Nick

So, it’s this extension of this kind of research nine years later. So, the whole idea is about allocating into nine… sorry… into five ETFs.

So, they have (and I’m kind of checking the list) S&P, US equities outside of S&P, international equities, global IG bonds, and then short-term US treasuries, those five ETFs. And the way that they design this, let’s call it trend following strategy, is that they choose either based on a dynamic window or on a fixed window, either the best or the two best assets out of the five based on their momentum.

Now a few details here. This is, to a certain extent, more of a cross-sectional momentum strategy. Because if you’re picking the best, it is not clear that this is a long or a short position, but it just happens to be the better of five. And obviously the same applies for the two. There is no shorting, so it’s a long-only portfolio. You’re always long the outperformer of the five or the two outperformers of the five.

So, we can just draw the parallels or maybe the differences to how trend followers would look at it from a directionality, not just the relative performance perspective. That’s point number one.

Point number two, the fixed window uses 10 months. I don’t know what is the 10 months. I don’t really personally use it, but I guess it’s the 200 days, that’s the closest I can get to. It’s like a 200 kind of moving average or some sort of a 200 days of a signal, that’s the fixed. And then they have kind of three windows depending on the level of VIX. So, if VIX is too high they go very short-term. I cannot recall exactly the numbers, but let’s say short-term. If VIX is in an intermediate range they go to an intermediate window and then if VIX is low they go to like a longer window.

Now, obviously, as soon as you use kind of VIX triggers, everything becomes very parameterized and quite sensitive to the choices you’re making. As an example, the periods of high VIX in those nine years of extension of the research are just three periods, three months actually.

By the way, they do monthly rebalancing, which is another, perhaps, more of a simplified assumption. I mean, in all fairness, and I guess being fair to the approach here, possibly the audience is a bit more retail oriented, so they’re looking at let’s say five ETFs. How can I allocate between the five ETFs should I use kind of momentum signals and VIX?

So, with three data points in what they call the red zone, I don’t know how much statistical confidence we can build which, by the way, is fully disclaimed in the article. And then the last point that I would flag in this whole kind of design is that either you choose the top of the five or you choose the first two out of the five. And if you happen to choose the two, you would allocate 50/50 notional. So, there’s no volatility scaling, there’s nothing as such. So, it’s a very simple design – top one or top two, equal weights, in dollar notional terms, monthly rebalancing. So, there is no follow the trend, no risk scaling, no optimizations, no risk allocations. There’s none of that.

[00:20:34] Alan

So, as you say, it’s very much aimed at an RIA type audience. Managing long-only portfolios and then assessing the merits of adding a momentum overlay.

[00:20:44] Nick

Perhaps, but there are times, for example, that you might find one variation outperforming the other because one is in S&P and the other one in Treasuries. But US$100 in S&P is much more risky than US$100 in treasuries. So, but that is the mindset here. We’re talking about fixed notional. It’s a fixed notional allocation.

So, with all those caveats in place, what they find, broadly speaking, is no significant value above and beyond a 10-month static lookback. There is an argument that when you choose the best out of the five, you’re kind of outperforming. But frankly speaking, it’s a period between 2020 and 2021 that you happen to be a bit more kind of moderate, that volatility wasn’t too high but it wasn’t too low at the same time. But if you look through the details, the rest of the periods, there is no statistical outperformance, neither underperformance.

Obviously, the good question is what happens after cost? And they have some analysis on this one. When they use the top two, there is no significant difference at all versus using the same two but coming from a kind of fixed window.

So, all that to say that there is no contradictory information to what I guess was the case back in 2017. Perhaps there’s an argument that, I know without statistical confidence, there is some value. But as we know, with these type of studies and those triggers, it is hard to make a statistical case. And triggers themselves by nature lend themselves to very low statistical power because you need the trigger to be activated. And guess what? The trigger is not going to be activated that many times because if it’s too many times, then it’s not even a trigger. So, the whole trigger rabbit hole (and I’m not personally like a big fan of triggers in a strategy) doesn’t seem, in that case, to add significant value. That’s the whole story.

[00:22:44] Alan

Did selecting one condition on the VIX, it did outperform, but as you say, it was very much linked to, I think it was November 2020.

[00:22:57] Nick

2020, exactly.

[00:22:57] Alan

Which was the month when the COVID vaccine was announced. And so, the small cap stocks, I think they were up 18% (I checked). I think up 10% or 11%. Yeah.

[00:23:10] Nick

I think I was looking into this one just a second ago. I think S&P at the time was like (exactly to your point) in double digits. And then the more kind of static strategy had like no treasuries at 1%.

[00:23:20] Alan

Yeah. I mean as you said, it’s concentration and small number of instances driving some of the changes. But I mean there is, I mean it is certainly, I suppose, an idea that I know some managers do model and do integrate into their trend strategies this idea of trading faster when market vol is higher and slower when it’s not or when vol is lower. Have you looked at it yourself?

[00:23:55] Nick

So, I mean, that’s the point I wanted to make as, I guess, as the next remark above and beyond what this article discusses. Volatility for us is a very important indicator and it’s utilized in other places. So, the way that we approach portfolio design is not dissimilar in that mindset to utilizing volatility to be more dynamic. But instead of us utilizing a different kind of window, depending on volatility, we scale our signals by volatility.

So, we look into signal strength that, by nature, more volatile regimes would bring your strength down. And if that is the case, depending as to how you do volatility targeting at the top level, it could have implications as to how much risk is eventually deployed. When those moves happen, I guess in a coordinated fashion, not only do you have volatility spikes, but you have correlation spiking, which also has implications on how much dollar notional is required for a specific volatility target.

So, I’m not in disagreement with utilizing volatility, it’s just that the way or the lens through which I utilize, personally, volatility in our research philosophy comes through the portfolio design, comes through the scaling of the signal, comes through the risk budgeting between the assets. So, it’s a bit more convoluted discussion without necessarily having to use a trigger, or rather a continuous function because you’re scaling by some that is continuous your estimate of vol, to scale your portfolio and to assess the relative strength of the trend signals. Obviously, we’re not using VIX because every asset will have a different volatility, but in the context of utilizing market volatility, we use parts thereof and the correlation structure as well. So that’s how I would look at it.

[00:25:44] Alan

I mean, essentially it’s kind of a regime classification.

[00:25:49] Nick

Yes. If I come to that point, then the whole discussion becomes, can we identify a regime whereby a correction is more prevalent and is going to be more extreme? And if we could, then we could time trend following. So, now we’re pivoting into the camp that says maybe we can identify trend regimes and maybe in that context we can utilize the volatility of the market. I mean, the reality is that this is a very hard gig, and I wouldn’t necessarily argue that I can prove statistically, at least using data, that this is a pursuit worth pursuing, if I may say.

[00:26:34] Alan

I think these are all areas where people do look at from the perspective of trying… I mean, obviously, you mentioned using volatility as an input in terms of market sizing and scaling. But then I suppose it’s then taking vol as another impulse as either a regime classification indicator or saying something about the likelihood of maybe an equity correction.

[00:27:03] Nick

I think my principles, and it’s something I’ve discussed here, maybe with you, maybe with Niels, is that there are periods that we can accept the fact that predicting a reversion is hard or that there are no real trends in the market. And as you said earlier on, we’ve been in the right space now, in the last few years, with no specific direction whatsoever.

So perhaps the prudent thing to do is to say, well look, I just succumb to the temptation of identifying any trend on something that is looking almost like white noise and I make either an active call to get out of some of those markets, because I’m lacking conviction, or I’m looking for systematic profiles that benefit from those regimes that would be complementary to trend following.

So, rather than trying to make the model much more parameterized just to make it work, I think, to me, it’s fairer to argue that hey, this is supposed to be doing what it’s supposed to be doing when trends exist. If they do not exist, I’m not going to try to recreate them with a more complicated structure. I’d rather find other components that would allow me to navigate through, and perhaps complementing my trend allocation is the prudent thing to do for long term viability of the allocation. I mean that, that’s my mantra, if you like, in this regard, from that perspective.

[00:28:33] Alan

It probably is fair to say though that I mean if you look at the performance of short-term managers and short-term momentum, they do tend to be stronger when you get probably higher realized volatility.

[00:28:45] Nick

For sure, for sure, but I guess my point is that is that a feature somebody should introduce to a more dynamic model? Because I would not disagree that being faster will get you better around turning points. But my question then is do you time the turning point or do you happen to be fast and when this arises you happen to benefit because you’re there?

So, you see my point? Calling the reversion is very hard, but if I were to be utilizing a fast model, I would have gotten it on time. The question is, can I deploy that speed dynamically, and will it be fast enough to cut my exposures? That’s a big question.

[00:29:29] Alan

I hear you, yeah.

[00:29:30] Nick

That’s a big question.

[00:29:31] Alan

Okay, fair enough, interesting. It sounds like you’re not going to be making any immediate changes on the back of these ideas. You’re pretty stable in your assessment.

[00:29:45] Nick

I think, I mean, I have some views, that’s a fair point. I mean, making changes using triggers, specifically in that world of QIS, is not an easy path because proving statistical value is impossible, just because you don’t have that many observations. When this particular article talks about three months…

I was in a conversation recently, somebody was discussing about a specific indicator, or achieving two timed specific market events, and I was like, okay, that’s great, tell me your false positives and then we’re going to have a conversation. Because it’s great if you can pick the turning points. but if you pick another 30 turning points out of the 33 in total, then I’m not so sure that those that you picked right are actually valid.

But jokes aside, I think the way that the markets have been operating recently, in the last couple of years (and we’ve discussed that quite at length), with very short live vol spikes that moderate very quickly (these are all the V shapes), that we have seen, basically, one per year for the last four years. I think these can cause significant disappointment if somebody were to follow an unconstrained trend follower.

So, in some respect, I think they do create an invitation for some thought to be put into how we act and interact with those V shaped markets that we start seeing more often. I don’t know if that is related to the economic regime or to the current presidential status in global markets, but the reality is that we have seen them, we have seen trend following going through those V shapes with a bit of a scar. Thankfully, we’re now here, a year and a bit since the last V shape or at least the last more aggressive V shape. I think the March event for this year was not as contagious as Liberation Day was, with trend following having done really, really well, and I think keeps on kind of getting attention from serious investors.

[00:32:02] Alan

Very good. Well, switching gears, there is a second paper, a paper that you highlighted. It’s very interesting and very topical and it’s in the whole area of applying agentic AI to an asset allocation portfolio construction process. And I mean, certainly reading it at a high level, it’s very much an eye-opener as to the way things may go looking forward in terms of institutional asset allocation. But the paper is from Andrew Ang. Do you want to give us, I suppose, the overview of the paper and then get into your thoughts on it?

[00:32:43] Nick

Yes. I mean, obviously, as you said, we’re shifting gears in a way, but I don’t think we’re shifting gears too much from how institutions kind of manage their assets and how asset allocation is perhaps getting reshaped in the world of agentic AI.

So, this is from Andrew Ang, I would imagine he’s quite well known. He used to be an academic at Columbia. Then he used to be kind of the head of factor investing at BlackRock. And now he set up Altbridge with another fund, I guess of his co-founders. So, that’s basically one of their first papers, I would imagine. He basically talks about agentic architecture for asset management.

And I mean those that haven’t actually seen it, and I’ve read it the last few days with a bit more detail, it’s actually quite fascinating. I personally kind of found myself kind of drawn into it just because you start seeing how the transition into agentic AI could have implications in many things that historically we used to do using people, and processes, and time, and various parts of that process could be now abstracted. Obviously, human in the loop is extremely important, but then that human in the loop becomes more of a supervisor, more of the fiduciary deploying, if you like, those recommendations and having, if you like, a sense of the truth versus the false positives in that respect.

So, how the whole story plays out… I’ll try to summarize the paper if I manage. It’s more of a nice story to go through because above and beyond it’s almost like a framework. Let me just start by saying, what is the overall process?

The overall process is that, let’s assume you have a big institution that has to put together a dynamic portfolio of multiple asset classes given a certain level of risk, or maybe appetite, or kind of guidelines. And it all starts with, I guess, a policy statement. What are the risks we can run at? What is the volatility? What is the expected return? What are the asset classes we can allow ourselves to invest in, and so on, and so forth?

So, then what happens, and this is now the pipeline of the agents that they’re kind of proposing to put forward, goes as follows, like six steps. The first one you have, let’s say a macro agent. So, the macro agent will do what? They will basically look into data, the economic regime, macro indicators, market indicators, and all this agent will come about is what is the economic regime, if it’s like expansion, late cycle, a recession, a recovery, associated with some confidence. And I think that’s important because it’s not about identifying a regime, but it’s also about producing an assessment of the identification risk of that regime.

Then you have asset class agents. So, now what you end up having is to say, well, I have 20 asset classes or maybe 20 assets, I don’t know, US equities, global equities or European equities, and then treasuries, and corporate bonds, and real estate, and so on, and so forth. I think they run an example of that model for March 2026. So, I think they have 18 assets to allocate. But let’s say you have a number of agents, each of whom is given a specific task to focus on a specific asset class. And the main purpose is to come up with an expected return of the asset class of a volatility estimate and some confidence around that, and then put forward an investment case, like a memo, for that asset class.

Then comes a, so-called covariance agent, which is, I would think, less of an agent and more of a kind of an econometric or a statistical tool that produces historical covariance estimation to bring this thing together. So, now here we are having, for all the asset classes, some projection of return, capital market assumptions, volatilities and covariances. And you need to put this thing into a portfolio.

What typically happens is that we have one or two portfolio allocation methodologies. Maybe we do some sort of a mean variance or some sort of a risk budgeting, and that’s about it. Well, the suggestion here is to say, why don’t you actually allow the scale that agents can give you to run 20 portfolios? So, there are now actually there are 19, each of whom builds a portfolio based on the previous inputs using different methodologies. And if I were to go through the methodologies (just for kind of naming a few), there’s like five categories.

The heuristics are more like no equal weighting, cap weighting, inverse volatility, and so on, and so forth. Some are return oriented, like mean variance in Black-Litterman. Some are more risk-based; the risk parity, max diversification, minimum volatility. Some are, as they call them, non-traditional, like a CVaR optimization or like a TPA, or drawdown constraint.

And then the last category of all the portfolio agents has two agents which are actually quite special. The first one continues doing research. So, for the first 19, everything is almost like formulaic. Later on there are some tasks that they would have to undergo. But at least as a principle, they build portfolios, different portfolios that they will make available to us.

So, the last kind of group or the last two are special. One of them is a researcher. So, this researcher will just do some active research, perhaps look into the latest innovation in the portfolio construction space and come up with new ideas to design a portfolio that perhaps down the line can become like a permanent kind of agent. So, for example, in this application they have is maximum entropy. So, this researcher is now finding maximum entropy as another possibility of building a portfolio.

The last one, and I found it very, very interesting, is called an adversarial diversifier. So, this is now an agent that observes the rest and effectively builds a principal component of the rest and tries to do opposite stuff. Now, why is that important? Because the whole aim is to build debate between them and get them to a place whereby the adversarial diversifier is trying to find pockets of correlations that the rest might have not actually identified.

So, now those 21 agents, what they end up doing, post designing the portfolios, is that they assess the rest and they start voting between them. They make some point about, obviously, whether that is computationally very efficient or expensive. And they say, well, we’re going to force each agent to only vote for a peer that is in the same group and one outside of the same peer group. So eventually out of 21 you have 42 votes. The voting happens in… sorry… these are the reviews. Then they all vote to each other and the voting is even more interesting. So, every agent has to give six votes, a five, a four, a three, a two, a one and a minus two.

Okay, and then there’s another phase whereby all those portfolios, because they have now a backtest based on today’s composition as inferred by the specific agents, also have a composite score of Sharpe ratio, drawdown and so on, and so forth. So, they try to basically score every agent now, with the voting that comes from all the rest, and an historical aggregate measure of performance. Those two combined on a 60/40 kind of scaling basis, give you the composite score of all the agents. So, suddenly you have now all the portfolio agents being informed by the asset class agents who talk about the capital market assumptions and come up with arguments in favor or against, who have already accounted for the macro regime. So, you can see how one brings that next.

And then you have kind of the strategy review, which is this point of bringing together all those data points. You have the CRO that is basically the covariance agent, and then everything goes to the CIO. So, CIO is a new agent here, and the CIO has to produce now a memo for the board that is going to condense all this information into a recommendation and some reasoning. The interesting bit, and they show into some analysis, and maybe I should pause in a second for you to kind of start reflecting, because I know I’m going to keep on speaking now, in the example they show, if they were to run their model in March 2026, is that the CIO is actually getting a very different view to the voting mechanism between the managers… sorry… between the agents. So, the most popular allocation scheme in their example is not the most popular in the selection of the CIO. And this is primarily driven by the macro uncertainty, and so on, and so forth.

All that to say that eventually what they’re bringing forward is that (I’m kind of scrolling through my notes because this is so many things I could actually talk about), all they try to bring forward is to say, with GenAI… sorry… with agentic AI, we cannot have scale. We can have some, as they call it, productive dissent, which is about debating and peer review. So, between them, using natural language, they can interact. So, the third one is a natural language audit trail. So, if you were to go back and check how fiduciary requirements are actually satisfied and deployed, there is, I guess, an audit trail of those decisions.

Then fourth kind of composable expertise. You put the specialist there, the CIO, some sort of forensic accounting module, and so on, and so forth. And then it’s called governable delegation. So, you have the kind of the policy statement governing all the decisions, how much risk you should take, and so on, and so forth.

So, to summarize, through all that process of agents, their point is, human beings will do the analysis historically, human beings will build their portfolios, and so on, and so forth, but the scale at which they can operate and the attention they can have is limited by the fact that we operate in real time and in human being capacities. Their point is that there’s an element of supervision rather than replacement now that should allow us to elevate the level of abstraction, guide the process with a policy statement that a human being would be writing and getting us the output from the CIO agent, the board memo for the actual board to then decide upon, with an audit trail through to be fully auditable.

Anyway, I’m pausing here. There are so many things that we can go about, but that’s the overall summary.

[00:44:14] Alan

Yeah, it’s very interesting. I scan through it myself and I mean, as you say, there’s a lot of rigor to it and there’s, in the sense of applying multiple portfolio construction techniques that wouldn’t be typical, I’m thinking in most investment committee processes people maybe would use one or two or maybe a couple more. But I mean you’re not going through… And even like in the Allocator series, which I host, and speaking to allocators who go through their process, they might say… I spoke to one recently hwo said they produced, I think it was three different portfolios, but here they’re producing, I don’t know, 15 or 20, effectively.

I suppose what I’m curious about is what elements of the process is the outcome most sensitive to from the perspective of…? Obviously, in this example, as you say, they ran it in March of 2026 and from a macro perspective, they concluded, it was late cycle with some stagflationary risks which I guess reflected, obviously, what was going on in energy markets at that time. So that made sense. I mean, so that’s one thing.

And then obviously you have the behavior of the CIO agent. As you say, he tilted the allocation towards portfolio construction techniques that the agents didn’t really vote for. But they were also relatively plain vanilla techniques, from memory, the ones that had the highest allocation. So, I didn’t get enough detail or I didn’t delve in to understand what drove his decision. So, I’m kind of wondering at that. Obviously there’s not a rigor there, but ultimately what really drives the tilting?

And in this situation the benchmark was a 60/40 portfolio, and I think they came out at 44% equity. So, they came out with meaningfully, relatively meaningfully underweight equities. Maybe the macro regime was part of it, I don’t know. Obviously, portfolio construction was a bit because they come up with capital market assumptions and they have a mechanism. I suppose that ties to the macro regime. So, obviously, they put less emphasis on historic capital market returns and more on, I guess, forward-looking estimates. So yeah, that’s a bit…

Any insights on that or how do you think, ultimately that 44%, that underweight equities, what did that stem from, would you say, reading the paper?

[00:46:57] Nick

I mean it’s not very obvious how it comes about because it is ultimately a combination of weighting schemes of different portfolio mechanisms. I mean, my gut feeling goes as follows, and that’s abstracting a bit from an already abstract context here. I mean, ultimately the way we do asset allocation is not just informed by risks, and covariances, and capital market assumptions. I think the other statement I would make is our ability to nowcast (if I were to use a term we used earlier on), to nowcast the economic environment in a very accurate manner, then I think we are more prone to build some concentration in our portfolios. When the economic regime is more mixed, in terms of evidence and data points to categorize it with some confidence, we typically vouch more for diversification.

Perhaps it’s a very similar comment, but in a very different context to the one I made earlier on about trend following, reframing a model to give you a better estimate of what is ultimately unobservable, the current regime ultimately still bears the risk of the model being wrong, and the allocation being incorrect, and ultimately leading to a loss. And I posed a question here. Should we instead be more prone to diversify at the times that we cannot tell how the macro outcome is clear? So, if it’s more unclear then we should vouch more towards diversification.

So, kind of going back now to the paper and your question, my gut feeling is that the mix of kind of contradicting macro signals at the time, and the fact that it was like a stagflation and a late cycle type of a period, perhaps suggests allocation that is less driven by capital market assumptions. So, I think the allocation that the CIO goes for is less informed by return-oriented allocations (so like cap-weighted, volatility targeted, equal weight, inverse variance), these are some of the higher ranked portfolios that the CIO chose which do not contain any capital market assumptions. These are no risk-based allocations.

So, in my mind it’s more like I’m not trusting my capital market assumptions because the regime is not very clear and because it looks a bit fluffy. I should then rather tilt my allocation towards risk-oriented and more heuristic allocation frameworks rather than trusting the data as much. That’s my kind of gut feeling or maybe my interpretation.

[00:49:45] Alan

Yeah, yeah, that makes sense. And they end up with a more conservative portfolio, 44% equities, which is interesting. I mean, another thing that was interesting that was commented upon in the paper is how it’s very hard to do a backtest.

[00:50:04] Nick

I mean that’s the most interesting for me.

[00:50:07] Alan

Backtest LLMs because they all currently incorporate information now that they wouldn’t have had back running it in real time. I mean they do point out that you could… Some researchers have tried to limit the LLM models to use models that were only available at a point in time, but it sounds like that’s not… Well, maybe those ones weren’t even computationally strong enough at that time to run these types of analysis. So yeah, I mean, you have to kind of take it maybe at face value the merit a bit. I mean, as a researcher then how do you deal with that if you can’t backtest, how do you evaluate?

[00:50:44] Nick

So that’s an amazing question. That’s why kind of reading it through put me into a lot of thoughts because there’s a lot of forward-looking biases and they fully disclaim. So, the discussion is actually very nice.

There is look-ahead bias by the way, in a variety of ways. The first one is the obvious, that you’re using a model that was built today, perhaps to build estimates for yesterday. So, it has all the information baked in the model and even if you were to contain the data as of when, the model still has the ability to, or has the knowledge in its parameterization of the future, even if you contain, I guess, the input.

The second part is, at the time that they start voting between them which one is the better model, and so on, and so forth. Then we have the typical look-ahead bias, or maybe data mining bias that we always have in backtesting. If I look into backtests, most likely I can choose the one that looks better. So, I would imagine the agents themselves, having 20 portfolio allocation schemes between them, perhaps they start kind of voting more in favor of the better performing one. So, it’s almost inherent now in the process what historically has been human beings kind of assessing backtests. So, these are two considerations.

The other consideration which I find even more interesting and, I guess, challenging, from your perspective, is LLM models are not deterministic, and even with the same model with the same prompt, you might end up having a different answer. So, reproducibility is nowhere to be seen with these type of models, at least as it stands, which in itself creates a significantly harder task for someone not only to build a backtest, but suggesting that this is representative of the methodology because it is not.

Like, even with all the rest being solved, let’s say today we start building a backtest as of today. So, tomorrow we’re going to have a new model and we’re going to do the same thing. So, we come here in a year from today, you and me, we have the same conversation with a year of a backtest. Had we rerun history or had we asked the model twice at the same time, a similar question with a slightly different prompt, we would have gotten two different backtests. And I think that’s the other challenge.

Setting that aside, I mean a few other points that they make, by the way, just looking through my notes, you might also have issues with regards to the fact that using the same LLM model, they call them monoculture risk. So, if all agents follow the same model from one of the big kind of GenAI kind of companies, you might find yourself being exposed to some correlation in the model. So, that’s another interesting one.

But overall, I think what I find interesting is the fact that there is here some suggestion that I would imagine can inform more accurately and more holistically asset owners. We cannot negate the fact that having the power of parallelizing all those tasks and creating better estimates of maybe the macro regime, or the capital market assumptions, or portfolio construction details, at the minimum should get us and investment boards more informed. So, in this regard, I think there is value. In the architecture of it, I think there is value. And above and beyond, I think the perspective is quite interesting.

[00:54:41] Alan

Just if you were to take this framework, and, obviously, it’s applied from the perspective of longer-term asset allocation, multi-asset portfolio construction, but at the end of the day it’s still building portfolios. If you were to apply it to building a trend portfolio, is there anything that you learned from the paper? Anything that you thought, okay, that’s something that would be implementable in the near term?

[00:55:07] Nick

Yes. So, I think some parts are actually quite interesting because let’s go back to this kind of food chain or workflow. So, you start with a… let’s assume you’re building a trend model and you try to fit it into those steps. So, you have a macro agent that says, okay, what is economic regime? What does the trend followers say? Well, I don’t care about the regime, but if the price goes up, I’m going to buy.

And obviously, the fact that I observe a rise in my price is nothing more than an indicator of the regime. Equities you’d be long more often than not when growth is there. And by the way, that’s what economic trend does at the end of the day, you’re looking into a regime and then you infer that your regime estimate is correct. So, you buy the asset that was performing the regime.

So, pocket number one is our signal, then asset class agents. So, which is the expected return and the volatility of each one of those markets that we have, or what is the confidence around that? Well, I guess when we do trend following, we say the expected return is our past return, maybe scaled by some measure of vol or whatever. So, in a way the confidence we have is how much volatile it is, which shrinks or expands our forecast. And the magnitude is coming purely by how strong that trend is. And these are not compatible the same way as capital market assumptions are. So, that’s step number two.

So, the signal in, I guess, one and two are kind of similar, but I guess the macro agent would look into predictors of performance, whereas the asset class agent is more about the relative strength. The CRO, we do it, it’s the covariance estimation. However, you want to estimate your covariance with a variety of techniques to make it robust.

The fourth is portfolio construction. This is the place that typically we go for one or two, and that’s about it. So, I think here there is something that somebody could spend some time on and looking at. Now, how that then flows into an investment decision in the current framework is about all the scoring mechanism and which portfolio favors the rest of portfolios and how they change their kind of their assessment of their peers. It’d not clear to me how that comes into trend following. But it’s not that those steps are too abstract to how we would typically operate in a research environment. That’s why I found it quite interesting because I can see, as researchers, going through those steps, sometimes consciously, sometimes subconsciously, to eventually go from raw data, to a strategy, to a single line.

[00:57:57] Alan

Is it fair to say with the evolving emerging technology that if you’re trying to satisfy or build a portfolio for multiple objectives, that that’s maybe easier now? Mean, just thinking, in this example, obviously, they had an IPS and I think it was relatively simple objective of whatever it was Sharpe maximization or a certain return for certain vol. Obviously, if you’re a trend follower, you’re trying to generate absolute return, but you want to tilt the portfolio to first crisis alpha or performance in a downturn. You want to minimize drawdowns.

[00:58:35] Nick

Yes. All this can come in… at times conflicting.

[00:59:14] Alan

Multiple objectives, at times conflicting or at times, yeah, I mean, hard to define. You may have more objectives than are just defined by mean variance optimization. That’s the point, I guess. Does that become easier now in this kind of framework, do you think, or not?

[00:59:00] Nick

I mean, I think it becomes more accessible, let’s put it that way. You would still have to, I would imagine, to utilize the expertise of running those portfolios to present something that is both sensible as well as tractable. Because ultimately those decisions are made by human beings and investors, institutional investors, that hold the responsibility of running those portfolios. But I think it will make it more accessible, just to be clear.

The first paragraph of the paper is actually quite truthful. It says, and I’m kind of reading through, as you were kind of speaking, it strikes me that basically what you’re saying is kind of in this paragraph. It says, “the most binding constraint in asset management is not data availability or model sophistication.” As in, we can get the data and we can be sophisticated enough, but it is the finite bandwidth of human decision makers.

And then they go about talking about a CIO that perhaps can oversee 10 to 15 departments that are organized in lines of asset classes, and then a fundamental analyst can probably cover 10 to 20 stocks. And a committee that is kind of coming together monthly or quarterly, sometimes for a few hours, to just say, okay, let’s go and see you in the next committee.

It reminded me actually of a client that said, around the inflation regimes in 2022, “why trend following is useful: Above all empirical data points it is going to take opportunistic shorts or longs, depending on where you are, without expecting the next investment board to say, okay, let’s change the allocation.” So it was purely an attempt to shrink the time that it takes between a decision and deployment, which is a limitation in the way that institutions operate. Just because time is finite and it kind of ticks deterministically, but it takes time in this way. So, one of their arguments was we like trend following because it can be opportunistic and deploy the exposures faster than our ability to bring together the investment board and do the analysis. Because by the time we do that, it’s already gone.

[01:01:23] Alan

Yeah, interesting. I mean, one final one. I mean, obviously we have, in our world, programs built by machine learning. I mean, this is framed slightly differently with distinct agents operating in very well-defined capacities. I mean, how soon do you think we will see products in the market that are clearly labeled as run by AI agents?

[01:01:48] Nick

Hard to foresee. I think the asset management industry can utilize, I would imagine, those technologies. I mean, we all can utilize the technologies, but on our side, obviously designing index products.

[01:02:00] Alan

Yes.

[01:02:00] Nick

The bar is higher for the obvious reasons. I think the reason why this is different to just utilizing historically coding and backtesting platforms is that the agent is ultimately speaking with its peers in natural language.

[01:02:22] Alan

Yes, exactly.

[01:02:22] Nick

So, we can then go back and vet, and read, and understand, and check the debate. And this is all becoming much more auditable and tractable. But that in itself is stochastic.

[01:02:36] Alan

Yeah.

[01:02:36] Nick

It’s not x plus y equals to 2, if x equals 2, what is my y? And I think that’s the beautiful challenge in this regard. But the benefit of the agentic that it speaks the language rather than just running the math.

[01:02:54] Alan

Okay, interesting. Yeah, definitely a wake-up call as to how far things have come along in terms of putting not just the use of AI, but multiple agents together. And as you say, when they’re speaking to each other and some of the additional benefits that come with that in terms of an audit trail, etc. So, I think, yeah, it’s definitely… It does get you thinking as to what’s coming next.

Well, let’s leave it there for today. Thanks very much, Nick, for bringing those research topics and always fascinating to get your thoughts. And Niels is away next week again, so I’m back again next week speaking with Cem. So if you have questions, please send them in. But Nick, thanks for coming and to everybody out there, thanks for dialing in and we’ll be back soon with more content.

[01:03:55] Ending

Thanks for listening to Top Traders Unplugged.

If you feel you learned something of value from today’s episode, the best way to stay updated is to go on over to your favorite podcast platform and follow the show so that you’ll be sure to get all the new episodes as they’re released. We have some amazing guests lined up for you and to ensure our show continues to grow, please leave us an honest rating and review.

It only takes a minute and it’s the best way to show us you love the podcast. We’ll see you next time on Top Traders Unplugged.

This podcast expresses the views of its hosts and the guests appearing on the podcast as of the date of its recording, and such views are subject to change without notice. Top Traders Unplugged do not have any duty or obligation to update the information contained herein.

Furthermore, Top Traders Unplugged make no representation to its accuracy and it shall not be assumed that past investment performance is an indication of future results. Moreover, wherever there is a potential for profit, there is also the possibility of loss.

This content is made available for educational purposes only and should not be used for any other purpose.

The information contained in this podcast does not constitute and should not be construed as investment advice or an offer to sell or a solicitation to buy any securities or related financial instruments in any jurisdiction.

Certain information contained herein concerning economic trends and performance is based on or derived from information provided by independent third-party sources. Top Traders Unplugged may believe that the sources from which such information are obtained are reliable.

However, Top Traders Unplugged cannot guarantee the accuracy of such information and has not independently verified the accuracy or completeness of such information or the assumptions on which such information is based. This podcast, including the information contained herein, may not be reproduced, copied, republished or posted in whole or in part in any form, without the prior written consent of Top Traders Unplugged.