Sebastien Page| Pt. 1– Portfolio Construction Masterclass: Unveiling Hidden Risks

Alan, Jean-Paul Page is my father. And in my asset allocation book, Beyond Diversification, I have a mystery author that I quote at the beginning of every chapter, and it just says JPP. Only at the end of the book do I reveal that JPP is my father. He was a finance professor for 40 years. So you can imagine, Alan, I grew up around finance. He never pushed me to say, "You have to study finance, this is your destiny." Actually, he was quite hands-off. But growing up, I would witness discussions at the dinner table between him and other academics, his colleagues, and they would have a glass of wine and discuss the Capital Asset Pricing Model and debate. There was so much passion around the table, and I was a 13-, 14-year-old hearing this, not knowing what it was about, but it's certainly, Alan, part of my DNA. Ultimately, it was somewhat predictable that I would grow up and study finance, and here we are 25 years into a money management career.

Not really, because I didn't know what it was about. But it really matters because what I learned was the passion they all had for it. They could be anywhere else on a Friday or Saturday night late in the night after dinner eating dessert, but they decided to sit together and debate these ideas. So even though I didn't really know what was going on and I wasn't participating, I grew up around it. My father had an unbelievable work ethic. That's the number one thing I learned from him. He wrote textbooks, and they were very intuitive textbooks focused not on publishing academic papers, but on teaching students the basics and then the advanced topics in finance, economics, and money management. Alan, if you were to stack the textbooks he published in his academic career, it would go higher than your dinner table. That's the work ethic he had. Not only that, but he ran a working farm basically by himself. Farming is incredibly hard; it's a full-time job. He was a full-time tenured professor on top of that and kept publishing academic textbooks. There was no doubt at some point, when I started studying the topic and participating in those dinner discussions with those academics, that I would end up in that field. But again, just for the record, he never actually said, "Sébastien, you should study finance." He was just teaching by example, working hard by example, and having passion for what he did, which is the best way to teach by example.

I reflected on what I learned over my entire career and realized that the best way to organize the content was in line with what we do in the practice of asset allocation. It begins with forecasts, and some people forget that. We're enamored with the historical averages, volatilities, and correlations, but it begins ultimately with a forecast of returns, then a forecast of risk. Then you put it all together in portfolio construction. The structure reflects how we do asset allocation in practice. For each of these, there are countless debates to be had because while asset allocation and portfolio construction has a scientific component, Alan, a lot of it is judgment. A lot of it is integrating quantitative methods with judgment. That's one of the contributions I want to make with the book, and sometimes integrating bottom-up insights with top-down macro insights. That's also another contribution I'm trying to make. So there's a lot in there, but the organization is all about how we do this in practice.

That's what we do in investment management: we make decisions under uncertainty. And uncertainty is not necessarily the same as risk. Risk you can theoretically model; uncertainty is much more difficult to model. There are plenty of biases when we as human beings make decisions under uncertainty. For example, how do you frame the information? Every day, Alan, money managers get presented with information that's framed in one way or the other to be bullish or to be bearish. There's a famous study that goes back to the early '80s on decision-making under uncertainty, where Tversky, who ended up being co-author with Daniel Kahneman who ended up winning a Nobel Prize, and co-authors asked a group of about 400 people to make decisions under uncertainty, and they found that the frame greatly influenced the responses people would give. It was, to summarize, a choice between how do you solve a deadly pandemic. Assume there's a deadly disease and you're in a room with 300 people, and you give people a choice: do you want to save 100 people or take a one-third chance that you're going to save everybody? Now, if that doesn't go through, though, you won't save anybody. So it's basically a choice between a certain thing—100 people saved—versus a gamble. And there's no right or wrong answer; it's are you risk averse or not? And over 70% of the people chose the certain thing. Then they gave people another choice, two other cures. And they said, "Okay, the first one is going to kill 200 people for sure, or you can take a gamble that has a one-third chance that you won't kill anybody and a two-thirds chance that you'll kill everybody." Now, for our audience—I know your podcast, Alan—you've probably all realized already that the first two choices and the second two choices are exactly the same. It's just a different frame. One is expressed in terms of killing people, one is expressed in terms of saving people. When you offered people the chance to save 100 people for sure out of 300, 70% will say, "Let's do it. Let's save 100 people for sure." When you offered them "you kill 200 people for sure," then it dropped to below 50% of the respondents.

So decision-making under uncertainty involves framing the information. And every day in financial markets, whether we look at quantitative models outputs or just turn on CNBC, we're looking at information that's been framed around a certain narrative in a certain way—sometimes by cherry-picking data, sometimes simply by showing data in a different light. I can say, Alan, the unemployment rate is up by 90 basis points, which it is right now, and make the claim that there's no time in history when that happened that it didn't continue to go up and end in a recession. So that's a strong recession signal. I can also say, however, that the unemployment rate right now is 4.3%, and that the long-run average is 5.7%, so we're actually running at full employment. Now, if I look at the level to frame the employment situation, I wouldn't expect a recession anytime soon. If I look at the change, that's a clear recession signal. Again, framing. So, Alan, it's super important, sometimes underrated. Most of us don't think, "Oh, what is the frame here that I'm looking at?" And that's super important.

Absolutely. This might be one of the greatest mistakes that we make as an industry: to rely on averages. Averages can hide extremes; they can hide different regimes. Now mind you, historical data is all we have. I remember early in my career, I was presenting to clients models for expected return and ultimately asset allocation recommendations. Sometimes the clients would say, "Well, okay, but your framework uses past data, so I'm not sure it's valid." At one point, I was jet-lagged and a little short with a client, and I said, "Look, I don't have future data. All I have available is past data. I cannot find future data on Bloomberg." But here enters conditionality. What are the conditions now that are relevant for investing—the valuations, the macro, the fundamentals—and can we find periods in history, either in terms of returns or risks, that resemble those conditions to make our forecast more relevant to current conditions?

The best simple example you can think of is expected bond returns. You can say, "Here's the average 100-year bond return; I'm going to use that past data, it's all I have." But if you know the current yield, it's actually a remarkable predictor of future bond returns. And we just went through a period where yields were essentially zero for the front end of the curve and zero for most of the curve after inflation. Therefore, current conditions—conditional expected returns—should account for current levels of yields. This sounds overly simplified; it is. I'm just trying to make an example. But every time in finance when we do forecasts, we have to think about current conditions and how we map those current conditions to make the historical data that we have more relevant. Recent work published after my book by my old mentor Mark Kritzman and his co-author David Turkington is focusing on what they call relevance. And this is super important for our industry; I would encourage your listeners to look into it. But what Kritzman, Turkington, and Kinlaw do is they look at current conditions—macro, valuation, fundamentals—and then they look back at the data and look for similar periods in history. They then reweight their prediction models, say a linear regression, to emphasize those data points because presumably they're more predictive because they resemble current conditions better. They wrote an entire book on this. It's much more complicated than I make it out to be, but these are advances in quantitative methods for return and risk forecasting that are continuing to evolve. They're as important as anything else right now. This is the cutting edge that we're looking at.

Yes. Up until pre-pandemic, we've just been through a 30-year bull market in bonds with constantly declining interest rates. I don't think that's the prevailing environment anymore. This has implications for how we think of strategies like risk parity, where you use leverage to increase the contribution to your risk budget from bond returns. It changes asset allocation decisions ultimately. I also think we're in a regime with greater inflation volatility, and the ability to capture a regime shift is actually very difficult, but it's one of the most important things a financial forecaster can do: understand what are the regime probabilities. You don't need to predict exactly that we're shifting from regime A to regime B, from declining rates to rising rates, but you can say generally the probability that we're in a rising rate environment is much higher than the embedded probability in my 30-year historical data.

Valuations are important, although from a cross-asset perspective, they haven't worked as well as they have in the past. You also need to account for macro factors, fundamentals, earnings growth, margins across asset classes, and sentiment, which I would also include positioning. You put these variables together and then you use your judgment, quantitative models, dashboards, frameworks, capital markets assumptions. It's a mosaic. On the risk side, I think the clustering is very, very important. There is predictability in risk much more than in return. I talk about this in Beyond Diversification. If you look at daily volatility from one month to the other, you actually get a correlation of about 60%. If you look at average returns from one month to the other, it's essentially uncorrelated. So risk is more predictable in that framework than returns.

You can say, "Look, I recognize that markets behave differently when we're at extremes, whether extreme positive or negative returns." If you use monthly data, a very simple thing you can do is cut it up into subsamples and say, "I'm going to look at correlations, exposure to loss for periods when the monthly return was minus 10% or below, for periods when the monthly return was plus 10% or below." And that's a good way to look at conditional expectations or conditional realized volatilities and returns and tail risks. However, that's not perfect, Alan, because you don't really know how to classify the regime you're in. I can have three months that are very extreme in a row or three days that are very extreme in a row, but then get a quiet day or a quiet month in my data sample. Let's say this is today. Well, what is it? Am I now all of a sudden in a quiet regime? I just had three really volatile days. And the Markov models and clustering models and regime models try to say, based on the path of volatility and the observations, even though you might be looking at a quiet day today, if you've had lots of turbulence leading up to it, it's much more likely that you're still in that volatile regime, that risk-off regime, than otherwise.

I like to use this analogy. I use a Garmin watch to keep track of my heart rate, my sleep, and all that stuff. A data person is going to always be a data person in all aspects of their lives. Let's say you don't know me or see me, you don't know what's going on, but you get my heart rate data. The heart rate data actually clusters, and you have regimes. You have the asleep regime where I might get up, my heart rate might go up, but just temporarily. And then you have the awake regime, and you probably have a third regime where it's exercise, where the heart rate looks even more different. But there are periods within those regimes that look like I'm awake, but I'm really sleeping. And there are periods during the day, if I'm just sitting watching Netflix, it kind of looks like I'm sleeping. But if you have a statistical model that recognizes the clustering and the patterns in the heart rate, you can tell without seeing me, without knowing anything else, most likely whether I'm awake or asleep or exercising. Same thing works with volatility in markets.

And Alan, if I can jump in, because you just mentioned forecasting as well: you don't have to say, "I'm in a high risk-on/risk-off, I'm going from this regime to that regime." If you have modeled, say, a risk-on and a risk-off regime, you can just reassign the probabilities. So a risk-off regime might represent 5% or 10% of your historical data sample, but if something just happened—like Lehman just defaulted—and you have a qualitative probability that if we just had April liberation day, the market's going to be volatile for the next few days, so you might say, "Okay, I'm going to reweight my risk forecast to put an 80% probability on the 10% most volatile periods." Right? So it's more about a sliding rule of subjective probabilities that you use to reweight your historical regimes.

Look, the holy grail is to try to forecast fat tails, but it really is kind of impossible because what makes a market event extreme? It's the fact that no one saw it coming. So by definition, fat tails are going to be very difficult to predict. I give it a shot in the book when I show that volatility is predictable. I have different trailing windows to build the forecast and then different forecast horizons. Then I do the same with skewness: is the distribution symmetrical, or kurtosis: how fat are the tails? And unfortunately, while I find that the volatility month-to-month is 70% autocorrelated, I get basically zero for kurtosis and skewness, which is kind of expected. However, this is key: you know that certain asset classes have fat-tail properties fundamentally. Credit, for example, in its fundamental form involves selling a put option. If you're a credit bondholder, according to the Merton model, you're selling a put option. So an asset class that loads up on credit or other types of risk premia, where you're essentially selling a liquidity option or another form of option, while you can't predict when you're going to hit the fat-tail event, you know that that asset is prone to fatter tails. You can model that, and what you want is to build the plane to be solid enough to withstand the turbulence, not knowing when it's going to occur.

And that is key, and it's well known. I'm not saying anything new. Academic finance has talked about this; even Markowitz in 1952, mean-variance optimization, even back then he recognized that if you read the original paper. So it's nothing new, but there's effort to be made in practice to better handle the risk-return trade-off with a focus on: am I getting paid not just for the volatility, not just the Sharpe ratio, which uses volatility as its measure of risk, but for the actual exposure to loss? And am I using scenario analysis not just after the fact to convince myself that I've constructed the right portfolio, but before the fact to make sure that I'm going to get paid for the exposure to loss in a way that's sensitive to realistic scenarios? So I had to pontificate a bit, but it's so important. It's part of risk management in practice. We all know this in our industry. I think Taleb is a great writer, but he kind of makes it sound like it's a discovery and practitioners are completely naive and don't know anything about fat tails. No, not at all. But the tools to handle those and to maximize your risk-return trade-off continue to evolve, and practice continues to evolve in a way that is becoming much more attuned to how markets really behave.

That's a great question. I started working on that before the Great Financial Crisis. I started with US versus non-US stocks. I looked at the correlation using monthly data going back to the '70s when US stocks were down by 15% to 20% or more in a month, and I saw a correlation for that subsample that was in the range of 90-plus percent. That's not that surprising. But importantly, as you hinted in your question, the correlation absolutely cratered when US stocks were rallying. So diversification, not only did it leave you when you needed it, but it also showed up uninvited when you didn't want it, right? So it was the exact opposite. Alan, I was working on this, and I got a call around 2007 from Charlie Henneman, who ran the conferences at the CFA Institute. I was a young sort of quantitative analyst; I didn't give a lot of presentations at the time. And Charlie said, "Sébastien, what are you working on? We might have a speaking spot for you." And I said, to your point Alan, "The title of my talk would be 'Undesirable Asymmetries and Conditional Correlations.'" There's a silence on the other end, right? Someone's organizing a conference; they don't want a nerdy talk. And after just a pause, Charlie Henneman said, "Okay, I have a better title for you," because he actually understood what I was talking about. He suggested the title "The Myth of Diversification," which is a wonderful title.

We ended up writing a paper because we found those asymmetries across almost all asset class pairs. And the more research we did on this, controlling for what should happen under a normal distribution using different ways of slicing the data, the more we found that we kind of know this asymmetry is undesirable, but we thought the industry didn't realize how widespread this was across asset classes and how consequential it could be for risk management and portfolio construction. Alan, the fascinating thing is, A, the title of the paper became "The Myth of Diversification," same as the title of the talk; B, it came out right after the fall of '08, and it was just good timing because we actually used data from before, and then everything that we described in the paper happened in the fall of '08. So we had a great title and we had remarkable timing on the paper, so it won a Best Paper of the Year award. I'm convinced most people didn't read it but voted for it because it had a great title and it was great timing. Now, there's an asterisk to this story: we found out years later, actually recently, that there's a better way to slice the data to look at how correlations change in function of what the returns are—conditional correlations—and we kind of corrected the original paper, and that paper correcting our original paper won the award as well. So no matter what you do, if you talk about asymmetrical correlations, you have a shot at an award.

That's a tough question. I don't think the amount of leverage in financial markets is nearly as high as it was back then. We're worried now about high valuations on the stock market, kind of comparable to the dot-com bubble. I think we will make those mistakes; the history of humanity and financial markets is that there are bubbles. At the time as we're recording this, I don't think we're nowhere near the level of speculation that we had in the housing market in '08 and the irrationality of it, and the fact that financial institutions packaged those loans with tranches and derivatives, and it became a house of cards, and all the banks held those assets on their balance sheet. In hindsight, what a massive amount of systemic risk. I don't think we're in that situation now. We might get there, but right now, US corporations and public companies have actually solid balance sheets. The financial system is resilient. And of course there's speculation—there's meme stocks, there's a lot of liquidity—but I don't think the conditions are that right now as we speak there's impending doom and a massive global financial crisis like we had back then. Again, we'll have some of those crises over time; I just don't think we're here right now.

One client told me once, "If you can't diversify, hedge." Of course, hedging is costly. But in some of our portfolios for people that are near retirement, we have a slice of our equities that is hedged, and we manage the hedge. We manage the downside explicitly with different methodologies to reduce the cost. Overall, we recognize there's a return drag, but for that part of the portfolio, we know we're going to do well if the market crashes, or we expect we're going to do well if the market crashes and bonds don't show up as the diversifier like in 2022, right? If it's an interest rate or an inflation shock, even US Treasuries might not be your diversifier. And it's extremely difficult to find assets that will act as a hedge or diversifier in a market sell-off, because that's when investors panic, that's when you have liquidity issues. We talked earlier about my father, a finance professor. He loved to use analogies, and he had an analogy to describe a liquidity crisis to students that I think is relevant when we think about why things all tend to fall together and what happens when things go bad. He would tell the students, "Imagine the building's on fire. Of course, everybody's rushing for the door. The...

The difference, he would say, in financial markets is that in order to get out of the building, you need to convince someone to take your place. If you're holding an illiquid bond in a market selloff, the only way you get out is by finding a buyer, and that creates gaps in prices. It creates assets that go down together just because they have similar liquidity characteristics. It creates panics, and that's why fear is more contagious, or it's related to the observation that fear is more contagious in financial markets than optimism. What do you do about it in portfolio construction? Well, I talk about in Beyond Diversification how you can construct portfolios using optimization models that are more sensitive or more calibrated to the tail risk of different asset classes—you know, instead of just running a Sharpe ratio maximization and loading up on credit and hedge funds or liquid assets that all have negative skew, they're all sort of selling options. If you calibrate your model using full-scale optimization, then you can control the tradeoffs with much more directionality and much more optimality.

Ultimately Alan and I started working on this early in my career. It was a really interesting moment. I got a call from my mentor Mark Kritzman, and at the time I was doing work for him. We would schedule our meetings. He was pretty busy. I was a quant running numbers and building models. When we met, it was scheduled, and I had just a limited time to explain my findings. Then he would help me interpret and think about next steps. This was a random call. So I'm sitting at my desk, phone rings, I get a random call and Mark says, "Seb, can you just walk over?" His office was next door. I walk over. Mark is sitting there in his office with a sheet of paper, and he's frowning a little bit. He hands the sheet over to me and he goes, "Do you understand what this is?" It was a handwritten fax, Alan. And the fax was from Paul Samuelson, Nobel Prize economist—he's passed away since then—one of the most famous financial economists of all time. And Paul was writing a handwritten fax to Mark to complain because Mark had just published something about mean-variance optimization, explaining how useful it can be to deal with different problems like currency hedging, understanding risk tolerance, and so on. So Paul Samuelson was saying, "You're forgetting fat tails, and here's how in portfolio construction you address fat tails: you optimize over the entire probability distribution. You don't need to model it with parameters. It's called full-scale optimization."

This, Alan, kicked off a project for me to start running full-scale optimizations to build portfolios of hedge funds, comparing those with mean-variance optimization, and it became a framework that we eventually started using in practice and writing about and publishing about. But it's a story to say there are ways to address fat tails in portfolio construction that have been published over the last couple decades, and they work quite well if you know how to use them. Now, at the time, what did Mark do? Because at the end of the fax, it was written on the margin by Paul Samuelson to Mark, "And Mark, therefore you and Markowitz are both wrong." And when Mark tells this story, he likes to say, "Well, then I called Markowitz and I asked him, 'Harry, what did you do wrong?'" But it was a debate of the titans, right? A debate of the academic titans between Markowitz and Samuelson about fat tails. And it's an important debate. It's part of the debate I cover in Beyond Diversification, part of the debate we're having today.