Bridging the Gap between Strategic Allocation and Investment Risk

Hi, friends. Welcome back to another episode of Peer Connections, the podcast series brought to you by the Global Peer Financing Association, also known as GPFA. These podcasts offer our GPFA members and global beneficial owner friends a forum for information sharing and discussion on topics most important to them. And we hope you, our listeners, appreciate the insights, best practices, and transparency offered from our members and industry friends about securities, finance, or related investment areas. Now let's get into the episode. Thanks for joining us for another Global Peer Financing Association Peer Connections podcast. My name is Chris Benesch. I'm a Portfolio Manager at the State of Wisconsin Investment Board. Today, I'm joined by a couple of colleagues, Jackie Lee, a Senior Managing Director in the Total Portfolio Group at Healthcare of Ontario Pension Plan, as well as Redwan El-Khami, an Associate Professor of Finance in the Rotham School of Management at the University of Toronto. Welcome, guys. Thank you for having us on. So today we're talking about a paper you published called Bridging the Gap Between Strategic Allocation and Investment Risk. I think this topic is really interesting to a lot of our members. It sort of crosses boundaries between asset allocation, risk management, modeling. Maybe to start with, why don't you guys share a little bit about where the idea for this paper came from? Yeah, sure. First of all, thank you for having us on. As to where the idea comes from, it really started many years ago when I was working at the Ontario Teachers Pension Plan in the risk division. And there I do a lot of modeling work. And one of the key characteristics of a lot of risk systems is they involve many risk factors, sometimes a couple of thousands, combined with tens of thousands of risk factors. And because of that, the time series data that you use in the risk system is typically short. And what I mean by short is it could be 10 to 15 years, starting from maybe the mid-2000s is where the data usually start. But as I progressed in my career and I moved into more of a portfolio construction role, I have been asked to build asset allocation models. And those models are quite a bit different. They involve much fewer factors, typically only a few hundreds, and they use much longer time series. In fact, as long as possible. So what that means is the time series we use, the minimum would date back to the 1970s, if not earlier. So when we try to come up with asset allocation decisions, we use the asset allocation models and it will come up with a result. So it could be buy more equities, buy less equities or more bonds and so forth. And it would produce some result in terms of the expected improvement in sharp ratio, diversification benefit and whatnot. And then oftentimes we're going to rerun those portfolios in the risk system. But as I mentioned, the risk system uses a much shorter time series and history. And sometimes this mismatch in time horizon would lead to different conclusions. So one quick example, and we're going to elaborate more on this later, is that a decision that the asset allocation model would suggest that it is risk reducing, maybe risk adding from a risk management perspective. So this wouldn't be a problem if you're not running tight in the price risk limit, but when you're close to it, then you're going to have a difficult conversation between the portfolio construction group and risk management group, because they're going to have different views. Portfolio construction guys will say this is risk reducing, but the risk management guys may say this is risk adding and vice versa. And we know where the problem is. The problem is the mismatch in time horizon. So at this time, we have been thinking about how we can address this. And we have an approach that I think could work and involve some mathematics for sure. I wasn't sure if it's a good idea. And Wetwan and I have known each other for a little bit with a consulting project. And so I went to Rotman to visit him and we were talking about some other topic. I couldn't remember what it was. But when the lunch ended and we were walking on the hallway and I told Wetwan, hey, Wetwan, I actually have this idea. Probably not a good idea. I just want to see what you think. So I'm going to tell him about this idea, basically using some statistical techniques to manipulate the result of a risk system to make it look like it's closer to the asset allocation models. And he thought it was a good idea. So we didn't know if this would attract any academic attention at all, but we went ahead and decided to write a paper anyway. I think I remember what was the topic. The first topic was we were actually talking just generally about leverage. At that time, that's actually the interest. And we spent a lot of time on it. But then this idea comes up, as you said, on the whole way. And we said, Jackie was mentioning, there is an idea I have. And what do you think about it? And basically, since I've also been working as a consultant to that point for other Canadian pension funds, especially, I realized, well, wait a second, this is a serious practical problem. I've encountered this also in other firms, basically, as Jackie was mentioning, there are risk systems, there are asset allocation models. For the audience, really, this is what I wanted them to keep. The idea is really bridging a gap. And that's where the title really comes from. Because usually people work in silos. The risk managers try to be as accurate about measuring risk. And usually they have limitation of the number of data. The asset allocators is about the vision of long-term investment. They use different ways. And there was a serious practical problem. I remember really well talking to Jackie because that was his, he initiated the topic. I said, well, someone in academia must have done this. It's too practical, but maybe somebody have looked at it. And we went back, we did some research, literature research, as all the academic papers does. And this is really how the idea comes from. I looked at it. It was no paper that addressed it. We chatted after and we say, this idea seems worthwhile pursuing. Why not doing it? Because there is a need, there's a problem to solve. And there is a technique that I thought is very smart and cute. And the first impression was cute idea. I loved it. And then we continue. This is really what I have to say on this one. I like it. I like that you identified the issue, you collaborated on what it means, and you really saw the, as you said, you saw the gap that was out there and decided to write the paper to fill in the gap for market participants in general. I think that's admirable. Maybe focus for a minute on when it comes down to the real crux of the problem that you're trying to solve? When you guys went to write this paper, you know, what was it that you were hoping would be the output for the consumer? Before just getting to what the problem really is, usually when you write a paper, and I've been writing many papers before on different topics for the academic side, but for this one, actually, it was very, very, very easy decision. Because as you said, we identified that there was a practical problem and there was no other solution to our knowledge at that time that exists in the literature. And our decision was, are we going to add value by doing this paper? And it's not firm specific. Maybe the problem has been identified in Jackie's experience working at Teachers Ontario Fund. But I had the problem I was encountering in other pension funds. In fact, other three that are not even Teachers Ontario Fund. And we decided, really, this is the problem we need to solve. And the mismatch, and I think, Jackie, you can probably elaborate on this one. The question was, you're asking is, what's exactly the problem you're trying to solve by doing this? So the crux of the issue is we want to resolve the mismatch between the model assumptions, between the asset allocation model and the risk models. The modeling typically is very similar. So it really comes down to time horizon. At the highest level, an asset allocation system is really just a risk system if we ignore the expected return aside. So really the only difference is what history do you use to calibrate the model. And I'll give you one more concrete example. So at least prior to COVID, US dollar CAD has been known to provide a pretty decent offset to equity drawdown since the great financial crisis. However, that may not necessarily be the case in the 1980s or the 1970s. So when you make an asset allocation decision on how much, for example, foreign currency or specifically US dollar exposure do you like as a Canadian pension plan, you have to consider the correlation between USDCAD and equities. And if you use recent history, at least before the COVID crisis, the correlation was quite high and quite negative. But if you were to go back further in time, that's not necessarily the case. Without the Canadian dollars, the US dollars, there is even one which is even simpler to asset class. Take equity and bond. And I really wanted the audience to see the real problem in organization is because think of this, you compute correlation or whatever it is that the relationship between equities and bond in a risk system and the risk system as we identified and this is not specific to one risk system is mainly the majority of them take aladdin take sandguard and sky take many other risk system algorithmics the majority of them actually basically have because and it's it's a constraint that we live they have to do it for hundreds and thousands of assets or securities they are limited in the data they use and because of that limitation, where asset equities and bond, if you add bond to your portfolio, it doesn't become risk reducing, whether it becomes, for example, risk add. Like the question, if you look at 15 years of data between equities and bond, or you look at 50 years of data, because equity is a bond, gives you two different ideas how much bond you need to add in your portfolio. And now it becomes a serious problem of organization. I want to do the right thing as an asset manager, but the risk system limitation is forbidden me from doing it because it added to the risk. And adding to the risk, you may actually reach a limit. And that's one of the problem actually you run into. Even though everybody admits, we know that there are longer horizon equity and bond, but I cannot calibrate it in a risk system today because we cannot have the whole variance matrix for 50 years of data. This is again, the real gap we deal with. USDCAD that Jackie talked about is one of them. Bond is another one of them. I can even give you a recent one right now, which is everybody talk about inflation these days. Take a risk system that you know, pick up whatever one that you have, and you're going to realize if you take the 15 years of data that you have in your system or 20, whatever, the tail scenarios that you computed, your value at risk, name it, whatever metric, I don't want to even use acronyms, whatever risk metrics actually you have used, it's going to show you that the tail of the disaster is actually a disaster. I'm talking and just place yourself a little bit before COVID, it's going to show you that really we have a deflationary environment being a tail risk. But if you take the data out to the 70s, you may actually have thought, well, wait a second, there is also a possibility for having an inflationary being a tail risk. But it cannot be identified in a risk system data. Our approach actually bridged that gap. I'm still using the same product. If basically you can limit yourself on a few factors that you can have data on over a long time, you can actually utilize that basically to solve the problem that we run with in a risk system. These are just three examples that are practical as it gets. How much USDCAD you need to have in the portfolio, how much bond is risk reducing or not, and what are your tail scenarios? Is it inflationary or deflationary in the portfolios? We think we kind of have found a cute way to do it. And I think this is why this paper have attracted some attention. Another point is that we were very excited to tackle this problem because it was, And it is still a very difficult problem because you're talking about two systems, one being risk, one being forward construction. One uses a lot of risk factor. One uses very few risk factor. One uses very long time series. One uses relatively shorter time series. I mean, how do you make them talk to each other in a consistent way? And the approach that we found was we liked it because we're able to essentially get the risk model to talk in the portfolio construction language, i.e. broaden the lens that the risk system can see, not just with 15 years of insight, but with 40 years of insight. The crux of it is we want to inject 40 years of insight into a risk model that builds using the 15 or 20 years of data. So that was going to be my question. How does the model do that? How did you guys solve this problem? Let me try to take you step by step but in very high levels so our audience can understand. So first, you know, take that many commercial risk system use scenario to analyze risk, be it historical scenario, multicolor simulation, what have you. The system will generate tens of thousands of scenarios and then you're going to calculate your favorite metrics, could be far, standard deviation, expected shortfall, what have you. And in all those scenarios, typically they will generate using 15 or 20 years of data, as mentioned. And in essence, what our model does, it is we waste the scenario such that when you look at the scenarios all together, they look like, and I use the word look like in quotes, they look like they are coming from a longer history. So if you run 10,000 scenarios in a risk system, basically each of the 10,000 scenarios are equally likely by definition. You do 10,000 simulation, equal chances of picking each one. But maybe there are some scenarios that are more likely than others if you were to look at a longer time period. So to give you one quick example, suppose you are using a 15 to 20 year history. and when equity is down, inflation is probably down. Well, why is that? Well, because you have the great financial crisis and then you have a few more crises in between. And in those times, typically the break-even inflation actually drops when equity drops. And I'm speaking with a slight caveat here. This is pre-COVID because COVID did change the data set a little bit. But if we stay just before the COVID period, just look at the 15 or 20 years data prior to 2021, this phenomenon. And in this setup, you can also search for a scenario that is equity is falling, but inflation is rising, but it's very infrequent. It could maybe 0.5% chance or 0.1% chance or what have you. So what I'll post that is it looks at the long-term data and say, wait a minute, this scenario is actually more probable if you were to look at 40s of data than 20s of data. So what it does, it increases the weights. So instead of being equally weighed, what would happen is our approach would then reweight all 10,000 scenarios, and it would put higher weights on the scenario that are likely to occur in the 70s and 80s, more so than the past 10 or 20 years. In a way that once they are re-weighed, then the entire distribution would then look like, and I used to put in quote again, look like as if it is coming from a risk system that calibrated on the 40 years of history. But the beautiful thing is, and this is where I think we really like it, is we simply re-weighed the scenarios. we didn't buy a new risk system we didn't have a three-year project recalibrating everything we did not have a new project of recalibrating the time series and we build we use the existing scenarios and we weight them and the trick is then how do we come up with the weights obviously the details are on the paper that we're not elaborate here but essentially that's what it does one thing more than what jackie's saying i don't want the audience to think well well why not building a risk system with a 40 years of data or 50 years of data but that's actually not first you can't do it and even if you do it that's not really the purpose we want sometimes risk system to have a short window this is just give you the ability is almost think about it as an add-on as a trick as a cute way for your elaborate risk system that is developed as you want it over a short horizon but the history that matters for your holding let's say public equities and credit and whatever. But we allow it to embed whatever history we desire as an asset allocator. If I'm a sovereign wealth fund or a pension fund that really care about long-term horizon, and I need to know what the cycles look like for commodities, what the cycles look like for the correlation between bond and equities and what that matters for the portfolio, I can embed that to the risk system without changing the risk system. This approach, I think what I enjoy about it the most that can be applied, take your risk system from the shelf, as long as it has scenarios. And I think all of them now have scenarios. If they generate scenarios, Monte Carlos, whatever, Monte Carlos or historical hybrid, you can basically be able to embed our approach in that risk scenarios without changing anything. This is really the trick. It's be keeping what you have, using a history longer for few factors. We didn't want to talk about this a lot right now, But usually this is also relying on a factor framework, but you can use whatever factor you desire to use, whatever characteristics or factors you want, and you can embed them into the risk system. And in this case, you have decisions at the portfolio allocation that doesn't contradict the risk systems spitting out some risk that is actually not even true through the cycle. That's actually the inside of the paper, and we hope people actually can use it. That's really the purpose of even doing this, and we want people to use it. And to add to Web1's point, and the practical implementation, I mean, obviously it depends if it is a Fender model or in-house model, the difficulty varies, but it really boils down to this. When a risk system generator, a bunch of scenarios is stored in some internal database somewhere, and when a risk manager asks for a metric, it then calculates it. And when it calculates, there's an embedded formula, right? If you're calculating standard deviation, VAR, expected shortfall, well, the embedded formula just assumes equal weight. That's what it does. So even if you are using a Fender system or internal system, really all you have to do is instead of equal weighting the observation, just use the weights that you design based on this methodology to calculate expected shortfall or volatility. And there you have it. That's the result. So really it's just a weight factor that you apply to when you calculate the metric. So you don't really have to mess around with the engine at all. You just have to, when you calculate the end metric, that's where you apply the weights and then you reflect whatever characteristics you want to reflect. It's as simple as that. So it sounds like there's a number of uses for this methodology across statistical analysis. What do you guys think is next? There are multiple uses also in terms we didn't talk about. One thing that I kind of omitted, which is stress testing and sensitivity analysis, which we didn't talk about also. and basically a very simple question of liquidity risk, which is actually important actually for this audience. You can do, in my opinion, using our technique now, what you cannot do right now in many softwares that exist, unless you change it or you do it outside the models, which is anytime you see a risk system that does what we call, I'm going to do some scenario analysis or sensitivity analysis, and I'm going to fix everything else constant. If you remember this discussion for anybody working in risk or portfolio, you will know but no else stay constant things are changing even if you basically want to change i don't know your correlation between equities and bond and you want to increase it so much but other things also going to change at the same time and if other things going to change at the same time what is the output of today's model assuming everything else stay constant maybe more of a myopic way of viewing the world our approach actually relax that assumption and can allow you to embed, in my view, a holistic kind of approach of stress testing and sensitivity analysis. Now, your question was about what is next. What is next? Actually, for the paper, we just hope other people use it. We're not selling it to anybody. We're not using it. We've kind of forgot about it for a while. But what happened to us, to be honest, is this collaboration between academics and practitioners or industry people like Jackie has been at least very rewarding to me personally. and it had generated multiple other ideas it's not the purpose of talking about them today in the podcast but you can just say we expanded from that work this is like a five years relationship work and i realized that in academia when i teach basically theory sometimes i'm not sure really how to use of it and at the same time when jackie comes up and say well there is a problem for example mean variance with the portfolio there is a problem we basically collaborated on multiple other projects probably we have another four papers to talk about now we have five papers together this is just one of them and those paper already got some of them I think three or four already got published and I think it has been a rewarding collaboration for me and I hope it's also for Jackie and hopefully these paper can be useful to people like we just wanted to see our work making an impact and adding value in other institutions in my view there's no other next step except adding value to the industry yeah I might just give you a bit of a teaser so after we walk on this paper as I'm working at hoop now and obviously in portfolio construction role you factor investing was a hot topic, at least for the last several years in the portfolio allocation world. So we have been looking into that. We did some work on what does it mean to be factor investing in a multi-asset world? And what are the techniques to do it? Not to go too much into it, but one challenge has been, people keep saying, if you are building an asset portfolio, you have many assets, asset class, maybe 10 or 12 asset class, but factors, people only typically talk about three or five factors so how can you build a three to five factor portfolio with 12 assets you have more assets than the number of factors so what is the right choice that's kind of one example of what we're looking to we'll be happy to talk about it some other time just one last thing if i i think the title of this paper we talk about today is bridging the gap between x and y actually what is next for us what really happened it was bridging the gap between what we think in academia and what practical really people think we are trying to bridge the gap because what we do in academia is fantastic i speak to it is amazing but sometimes we are not really thinking about practical questions to solve the way that the practitioners are thinking about them they are definitely useful you can actually use them but they need some kind of bridge in the gap between the academic world and the practitioner world and then this is i think one of the papers we're going and working on working on is basically bridging this gap trying to understand what has been developed in the academic world, trying to address some practical questions and do it kind of rigorously. This is the one. We're trying to do it in a certain rigor that it's not just a thought process, rather than it has to be a little bit robust. This is really what our collaboration is trying to achieve. Well, I think it's obvious. This is a really great example of collaboration between you guys with a really, I think, valid outcome for a lot of market participants. So thank you both for sharing all this information today. For our members who are interested, where can they find the paper or find out more? This paper is published in the journal Portfolio Management. I guess you can find them on there. You can find it in the journal. There is appendices in the paper. We refer to some appendices. They don't make it sometimes to the journals. Those guys want the appendices because we can't share if it's published already, but we can share the appendices. Great. Well, thank you. And again, that paper was bridging the gap between strategic allocation and investment risk. Jackie Lee, thank you very much for joining us. Redwan Elkami. Thank you for joining us. Really appreciate your guys' time today. Thank you for inviting us. Thank you. Thanks for listening to another episode of Peer Connections by GPFA. We hope you found the information shared in this podcast interesting and beneficial. And as always, please feel free to reach out to GPFA with ideas or interests for future episodes. And if you liked what you heard today, don't forget to subscribe wherever you get your podcasts. Now for the disclaimer, the opinions expressed in this podcast are those of the presenters and do not necessarily reflect the views, or opinions of their respective employer organizations. This material is for your private information and does not constitute legal, tax, or investment advice. There's no representation or warranty as to the current accuracy of nor liability for decisions based on this information.