Chris Carrano – Designing Practical Factor Models (S7E20)

Hey everyone, Corey here. Thanks for tuning into another episode of Flirting with Models. If you're enjoying the show, I'd greatly appreciate it if you'd take a moment to rate, review, and most importantly, share with a friend. A word of mouth is how this podcast grows. And if you'd like to learn more about newfound's platform of Returnstacked Mutual Funds, ETFs and model portfolios, head over to Returnstacks.com. Now on with the show. 3, 2, 1, let's jam. Hello and welcome everyone. I'm Corey Hofstein and this is Flirting with Models, the podcast that pulls back the curtain to discover the human factor behind the quantitative strategy. Corey Hofstein is the co-founder and chief investment officer of newfound research. Due to industry regulations, he will not discuss any of newfound research's funds on this podcast. The audience expressed by podcast participants are solely their own opinion and do not reflect the opinion of newfound research. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of newfound research may maintain positions and securities discussed in this podcast. For more information, visit thinknewfound.com. In this episode, I speak with Chris Carano, Vice President of Strategic Research at Ven by 2 Sigma. Scientists have had the rare vantage point in the world of factors, spanning smart beta, long short hedge funds and risk modeling. And that experience has shaped the thoughtful view of what factors really are and how they can be practically used. We dive into the philosophy and design behind Ven. Why it uses just 18 orthognalized factors, how it blends lasso and OLS to reduce overfitting, and why it prioritizes interpretability over complexity. We also tackle messy real world challenges, how to analyze private markets with sparse data, how to trust synthetic return streams, and where to draw the line when using monthly snapshots that embed structural portfolio shifts. Finally, we explore what it means to make factor results actionable by the through stress-testing, residual interpretation, or portfolio diagnostics. Please enjoy my conversation with Chris Carano. This welcome to the podcast. Thank you for joining me. Excited to dive deep in all things the world of factors today with you. Excited to be here. Thank you for having me. It's an honor. So let's dive in. I usually start with each guest a little bit about their background, but I would love to tilt this question. Obviously, I want you to share your background and how you got to where you are today. But a lot of your exposure through your career from what we discussed has been different vantage points on factors themselves, starting somewhat in the smart data side, long short hedge funds, and now working in risk models. But everything including that factor-based lens, how is your view of factors evolved throughout those roles? So that's a great question. I am going to start it by reading a disclosure very quickly and just say that before we get started, I'd like to make it clear that what I share on this podcast is not necessarily an endorsement by two sigma of the views I express is not intended to be relied upon as investment advice. Please refer to Ben's website for important disclaimer and disclosure information. Okay. So now that we've got that out of the way, great question on my experience and vantage point on factors, the way I'll start it is thinking of my career trajectory. So I started at an ETF shop that was smart beta, which I don't even know if that term is used as much as it was back in the day. I've moved on from that part of my career a little bit, but at the time, smart beta was everywhere. So this shop did dividends waiting, earnings waiting, long only equity exposures for the most part. And that's where I really got exposure to factors because dividends and earnings lend themselves to quality and value. So that's where that started. Now we did some multifactor there, bottom up, for example, and other things of that nature. So for me, that was getting factors as an enhancement to an investment product, the vantage point there. Okay, this is a little bit off the topic, but something I really valued about that time was I was really working with the data at that point in my career. So I got to know what it really felt like to build a PE for 5,000 stocks. You have some of the data for some of them. You don't have the data for other ones. I spent like an hour debugging code just to find out that my dates were misaligned. I think that type of experience really adds perspective to what the actual inputs are into these kind of factor products that are out there. So that's just something I really appreciated about that experience. Then I moved a little bit away from the data. I moved into being a product strategist for a factor suite. So this included factor rotation products, long short factor products. And this is where I got into what I like to call common fundamental factor discussions, which is a mouthful. But it's basically the exact conversations that are probably happening in a financial advisor's office right now, which is why is leverage better or worse in a quality factor? What's intangible versus tangible for value? What are the different reasons why equity style risk premiums exist? So that's where I really got a lot of that exposure. And some of the products I was covering there actually had risk models as the underlying engine. So that's where I started to get exposure to this idea of factors as a framework. And the idea that not all factors are returned seeking, the foreign currency factor is just trying to explain risk. And that really changed my whole view on realizing that factors as a return enhancement for an investment fund, it's almost like a second derivative in my opinion with the real use cases, which is for holistic understanding of risk for multi asset portfolio. So that's where I got that exposure risk models and things like that. So this was a natural segue to where I am now, which is Venn. Venn is a risk analysis platform. So just to be clear, this is portfolio analysis performance reporting. It's really the whole gamut of stuff you would expect. But there's a risk model that's really the engine gives factor analysis and all these things for institutional multi asset class portfolios. And I think what really drove me to Venn is the risk models I had been engaging with before were extremely granular, hundreds of factors and really a lot to take in. And then really takes an opposite approach, which is it tries to be as simple as possible. It tries to approach factors through accessibility and making it so it doesn't have to be like an analysis paralysis type exercise. So I think that's where I'm at today. Still thinking about factors as a framework, but trying to make it as simple and accessible as possible. I know a lot of Venn's client base is institutional allocators. Maybe we can take a big step back here and start with the fundamental question of why should institutional allocators even adopt a factor based lens in the first place? What kind of questions does it help them answer that a traditional asset class view, which has been used for decades, can't answer for them? Let's start with what the reality is today, which is most institutional allocators are using asset classes. So what I mean by that is the most simple example is like a 6040 or something like that. And when they go to make an asset allocation decision, they're saying, I'm going to lower my duration and they're going to look at their bond managers and they're going to start playing around only in their fixed income sleeve. This is the asset class framework of thinking that is typically being used today. So the problem I think while asset classes can be helpful, just to be clear, this is not an attack on asset classes. They can be very misleading because at the end of a day, an asset class is just a label. It's not really a description of the underlying risk. So something I feel very confident in saying is that if you think about some different asset classes like bonds, real estate, fixed income, all of those have exposure to interest rates. So interest rates is the fundamental factor that is shared between them. So even though you may be diversified across asset classes, you may actually have risk exposure to a much smaller set of fundamental factors like interest rates. So the whole idea is let's think things like, what do we do about duration? It's thinking about that only now in a risk world rather than an asset class world. And when you're using factors, if you're analyzing a manager, a sleeve, a total portfolio, you're doing it using the same set of factors in the same language of risk. So you're applying that to all your different asset class managers. And I think that is what really makes true total portfolio analysis possible in a way that a traditional asset class approach just can't replicate because you're continuously thinking in that siloed way. So I'll just say quickly in terms of adoption, I won't lie. I've been in factors for more than a decade. Over my whole career, there's always been this feeling that factors are going to be mainstream tomorrow. It's like it's right around the corner. And it's never really materialized. But at the risk of being ironic, I do think it is right around the corner. And part of that is because there's an organization, Kaya, they just released a paper called the Total Portfolio Approach. You probably know Kaya, a huge organization, touches many institutional investors. And that talks about factor analysis as the engine of a total portfolio approach. It's the thing that helps you think about total portfolio outcomes. So I do think people are going to realize more and more the role that factors can play. And I think that paper is a response into how portfolios are getting more complex. More hedge funds, more liquid alts, more private assets. This is what factor analysis is really trying to help you understand those hidden correlations and volatilities that it can be hard to really digest otherwise. And I'll just quickly say that's where it then fits in. So it's like I'm thinking about factors and then then fits in. to help bridge that gap to actually implement. A significant part of the factor analysis equation is ultimately what factors you're using to do that analysis. And Venn takes, at least in my opinion, a notably parsimonious approach. There's 18 factors, the hierarchy of factors, the factors are all orthogonalized to each other. What's your theoretical argument for a less is more approach, particularly in the context of multi-asset portfolio risk modeling. Again, at least in my experience, the number of factors that Venn uses is significantly lower than unnamed competitors. I don't think anyone's going to bat an eye twice if you say 18 factors is a small number of factors. So I think you're okay. So definitely a parsimonious approach for sure. So just to recap, parsimonious about really being selective and disciplined in what you model. So I think if you cast too wide in that and you try to fit every risk signal under the sun, there's clearly problems with that in terms of overfitting on historical data, so it won't be predictive out of sample, less interpretable. You get a really sensitive and fragile model. So we're trying to take that in the opposite direction, which is a parsimonious, simple, understand, digestible lens. The type of factors we use are the greatest hits of market risk. So if you think about macro drivers like equity, interest rates, credit, commodities, these are things that are going to jump out and slap you in the face for any multi asset institutional portfolio. And then we also have style factors. So these are the risk premium drivers, so momentum, value, inequities, trend following as a macro style factor, for example. And I think the important thing here is that the factors are selected and constructed to explain as much risk as possible. And if you're taking a parsimonious approach, obviously that's really important. So just to put that in perspective, our value factor is not some super unique high alpha our best approach at a value factor, because if we did that, then the broad institutional landscape of asset owners are not going to correlate to that value factor, so we won't explain any of their risks. So that's not the point. Now of course, we want it to be representative of the academic return premium and all of these other things, but it's really about selecting and building them to explain as much risk as possible. Now, I'll say that we don't stop there in terms of parsimonious. We also use a two step regression process, so the OLS tries to fit everything you throw at it. We first use Alasso, which is a little bit of a factor selector. So when people actually conduct this analysis on then, we're often not even using all 18 of our factors. We're often using less than that. So again, the goal, be resilient out of sample, be transparent, give analysis that's really actionable. And I do think a little side benefit is that the investment dialogue is just so much easier. So whether it's an investment team talking to each other, trying to do manager due diligence, and you're talking about 18 factors in these big fundamental risks. If you're talking to an investment committee or an end client, you don't have to be like, don't worry about the model. You can actually talk about the model because it's digestible. So that's the idea. In your answer there was the acknowledgement that to some degree, the choice of factors is inherently subjective. There's no clean platonic set of factors that every single investor will necessarily agree on. The common critique of factor models is that they risk overfitting, especially when you're talking about shorter term time series. You mentioned that then applies this LASO-based approach to try to combat this. Actually, it's a two-step approach, is my understanding LASO, and then no LAS. Can you walk us through this two-step process that you apply that tries to improve the robustness of fit without the problem when you start to do too much? Is it makes the system opaque and potentially not interpretable? Again, how do you fit that balance of robustness versus interpretability, which is the end goal here? Let's take the LASO out for a second. If we did an OLS with all 18 of our factors, that's going to be the best fit possible. But it's also going to be the highest risk of overfitting and not being predictive out of sample. How do we try to combat the fact that an OLS just fits everything you throw at it? That's where you introduce the LASO, which kind of acts as a factor selector first. The way I like to think about it, I know we're on a podcast here, but if you can picture it in your mind, you have the regression, you have all 18 factors as the independent variables. I like to think of the LASO as like, this is dial, like you'd see on a washer machine or something like that. I call it the factor shrinking dial. If you turn that one notch, what's going to happen is those 18 factor betas are going to shrink towards zero, and the less influential or the smaller betas are going to go to zero. Maybe we turn that factor dial once, and it goes from 18 to 16 factors, with two of them being at a zero betas. We're basically dropping those. We turn the factor dial again, and maybe four go to zero. If I'm doing my math right, I think we're at 12 now. We basically can continue to do that process until you're eliminating these variables from the equation. This dial is not a perfect answer, because if we don't shrink it at all, it's just an all-est regression with all 18. We shrink it all the way, and then you got nothing in it. How do you decide where's the right place where you can balance the goodness of fit with a number of factors? This is where we use a score called corrected AIC, and it does exactly that. It calculates goodness of fit number of factors. Here the best score is, which is actually the lowest score, is where we set that dial in terms of the number of factors that we allow to pass through to the OLS. We use a corrected AIC, which is basically an adjusted for a small sample size. Let's say the portfolio we're analyzing is only 36 data points. The score is going to prefer an even simpler model. It's really penalizing the less data the less complex the model is. That's acknowledging the higher risk of overfitting with less data. If you don't have a lot of data, then it might only spit out one, two, three factors, because that's what it's confident in terms of not introducing itself to that overfitting risk. Once you have that factor selection, you have the right dial setting, you've turned it, you've got maybe 10 factors that make it through. That's what you put into the OLS, and that's the workhorse. That's where you estimate the data as you get the T stats, and you're fitting everything you throw at it. Just to be clear, it's not as if this is some super magic bullet. It's returns-based factor analysis. There's still times I might analyze a long-only equity active manager, and there'll be some fixed-income carry data in there. You made it through the LASO into the OLS, and I still have to have a good head on my shoulders. That's probably a spurious correlation. I don't know that fixed-income carry is relevant here, so it's definitely not a magic bullet. But I will say something that's interesting is because we're taking this approach with only 18 factors, it's really easy to know what got filtered out by the LASO. If you only have 18 factors and we give 12, you're like, "Okay, well, value is clearly not here." It's really transparent in that way, where you know what's being used in the OLS to basically model your portfolio returns. How many full portion of the factors that you have in your library are what I call style factors, the value, momentum, quality, defensive, even trend? These are places where, in my experience, implementation details can significantly influence factor behavior. If you define value one way versus another, it can significantly change the loading that a factor regression can result in. What's even more difficult here is that one allocator's definition of value or momentum or carry might actually differ from Venn's library definition. How do you think about balancing this precision in factor construction with the need for users to align the outputs of Venn's factor model with their own mental models and definitions? Fair question. We have equity styles in Venn, value, quality, momentum, low size, low risk, and crowding. Definitely all of those have different metrics and the metrics matter. So 100% agree there. I'll just say that we don't really get too cute in terms of what we use. So you're not going to see a quality fundamental. We're like, "I've never heard of that. It's going to be R-O-E, things like that." That doesn't mean it doesn't matter and everything you said is true, but what I would maybe do is shift the question a little bit to something that I think is sometimes underrepresented, which is not necessarily the metrics, but also how important the portfolio construction is of those metrics. Whether you're creating an investment fund or a factor, how you ultimately represent that return stream matter. For example, all of our equity styles have the same portfolio construction, different metrics, of course, same portfolio construction. One of the most influential parts of the portfolio construction I think is being market neutral or beta neutral. They're long short, but also explicitly aimed to have beta of zero to the market. Where I think this can break down in terms of the mental, math, or expectations and intuition of investors is especially with the low risk factor. Because if you think about it, there's two constructions you can kind of approach the low risk factor with. A dollar neutral, which is long low risk stocks, short high risk stocks, but an equal dollar value between those long and short exposures. Then the same market neutral, but equal beta value between those two. If you think about the dollar neutral, it's short the equity risk premium because the long basket has a lower beta, high basket, probably higher beta. So you're naturally short the equity risk premium. Let's imagine the COVID crash. I say let's imagine, but we actually did research on this between the difference between the dollar and a market neutral during the COVID crash. The dollar neutral was positive during the COVID crash. Obviously, short the equity risk premium benefit from that tailwind. The beta neutral version of our low risk factor was actually negative during the COVID crash. And the comment here is it doesn't have the tailwinds from being short the equity risk premium. So now, we're going to look at the next one. Now it's relying more on the spread between high risk and low risk stocks, essentially adjusted for risk, ironically. That can be hard to wrap your head around, but one of the benefits of this is orthogonal to the equity factor when you run a regression. But let's think about that practically. You're an investor, you have something simple, like a low risk ETF, and it outperforms the S&P 500 during the COVID crash. With a market neutral version in our model, it would essentially just be telling you that you should have just had less equity beta. If you had perfect foresight, you should have just had less equity beta, and actually introducing low risk was not a prudent thing to do over that period. If you think about it, that's actually a much simpler takeaway, because the equity factor is more liquid, it's more actionable, and it was just that that was driving and explaining out performance of that low risk ETF for the S&P 500. So incredibly different model outputs, whether you used market neutral or dollar neutral for low risk, you would get a completely different answer in terms of what was driving out performance over that period. Now, something else, this is just a little different that I like to talk about, it's the betting against beta idea, but a dollar neutral low risk factor is also fighting the equity risk premium constantly. In my opinion, you're basically hurting your risk premium there for no reason. The market neutral version, I think personally, that would be my preferred implementation, not only for orthogonality reasons with the risk model, but also just in terms of seeking a risk premium over a long period of time. So definitely it could be argued not the most intuitive at times, but that's what a risk model should do is really get at the true drivers of return. We've used the word orthogonal probably at least a dozen times now, but it is a really critical design choice that you made. And I think it's important both from the mathematical side of making sure you have a good fit to these factors that they are truly independent from each other, making the fit easier. So you've also written it with then that the orthogonalization allows factors to provide clearer insights, which I think is what you just spoke about this, but sometimes those clear insights can be at odds with an allocator's intuition of observed market returns. For example, emerging markets may go up while the emerging market factor is flat or even negative due to residualized macro risk. How do you reconcile these trade-offs of goodness of fit, explanatory power, and again, the end users need for interpretability to create actionability with these factors? Great question. Just taking a step back. We just talked about being orthogonal with our equity factor through long short portfolio construction for our equity styles. We're not able or we don't do that with all of our factors. So for other factors, we use something called residualization. You mentioned emerging markets, so I'll talk about in the context of EM. Our EM factor, before any residualization or thogunization, nothing fancy, is just three indexes. It's market cap weighted, EM equities, market cap weighted, debt, and then EM currencies, equal weighted. So it's really not a fancy game-changing interpretation of what EM could be. Now that basket of EM indexes, we think, has some exposure to macro factors. There's probably no one on this podcast that would argue that that basket I just described has some global equity exposure, because 33% of it is EM equities. So what we do is we try to measure how much exposure to these macro factors this EM basket has. We do that using a rolling three-year, exponentially weighted regression. We found that this is a long enough period where the relationship between that EM basket and those macro factors is not overly sensitive and noisy, but it's short enough where when markets change, it reflects that. So conceptually, what we do is we go long that EM basket of indexes, and then we short the macro factors with the magnitude of those shorts based on that estimated rolling regression. So that's conceptually how we basically try to get to that pure EM exposure. Now in terms of trading that off with an allocator's market intuition, our EM factor is negative over long periods of time. So the residualized EM factor, the independent one, is about 150 bips per year, it's negative. So imagine you're an institutional allocator, and you load your portfolio and vendor, you build it, and you've got an EM sleeve, and you've got five EM managers, and you're like, oh, wait a second, the EM factor and Venn's not even rewarded over time. So that definitely can challenge intuition, because they expect EM growth potential high risk, but high reward, when our factor would suggest that it's actually just the macro factors doing the heavy lifting that are embedded in EM. So maybe there's an EM rally, not all the time, but most of the time you probably would be better off just having a higher equity beta, not necessarily being in EM specifically, because that may be lagging behind the EM, though global rally. That can be challenging sometimes, but at the end of the day, similar to before, like that's what a risk model does, it tries to get at the independent actionable truths, and this is our framework. So if someone really disagrees with that, they might go find a risk model where EM is rewarded, and they just have to kind of accept that. And not to end the question with another bomb drop, but actually small cap factors not rewarded in our lens either. The market neutral implementation of a small cap is negative over the full period. So that's another one where people sometimes have to wrap their head around it. One of the maybe benefits of sitting in your seat is you get to work with a lot of institutions, and you get to see a lot of institutions put their portfolios into then and see what comes out on the other side. And one of the things we talked about in our pre-call was that fairly consistent theme is that most institutional portfolios are maybe far less diversified than they appear when they go in. There's a lot more factor concentration that comes out. I hope you could maybe expand upon that and talk about some of the common surprises that emerge when Ben is applied to institutional total portfolios. So this is going to hit on some of the things we've kind of alluded to throughout. I think UNI both have done this. So when you think about institutional portfolio, they come in and they're ultra diversified from an asset class perspective. You've got real estate, liquid, ETF, some mutual funds, you've got hedge funds, some private assets, and then they'll run it through Ben and I'm not going to say it's common, but it's not uncommon for that to spit back out an equity and interest rate factor as the main drivers of risk. Sometimes upwards of 80 to 90%. Now I'll just say quickly that could be the lasso. Maybe they're coming to the table with 36 data points in the lasso is spitting out a simple model. But in my personal experience, that's not always the case and it's often not the case. So we had mentioned this residualization idea of taking out macro factors from other factors and consolidating that into the macro factors. So we actually have a tier system that does that. So imagine it like a pyramid. And the highest tier macro factors in the two sigma factor lens are equity and interest rates. So just think about this as currency hedge, equities and bonds. That is the world with which we're trying to consolidate as much risk upwards as possible. Then tier two is credit and commodities, tier three foreign currency emerging markets and local inflation. Again, residualization is the tool for the most part, equity styles we describe market neutral. But the residualization is the tool that passes that risk upwards to the highest tier macro factors. So that's a decision choice to put one factor over another in the tier. And the reason why we really make that decision is because we think they're more liquid and actionable is of course, if you want to implement portfolio changes, that would be the best place to start. And going back to the original question, the surprises you have this mix of asset classes and all of this stuff. And either it all has very similar underlying risks or your portfolio mixes in such a way that it can just be described by these main high tier factors. But I definitely think that can be eye opening for a lot of people. The benefit is the payoff of that is you get a more actionable understanding of what's really driving the portfolio outcomes. You introduce a tool to the process that I think it can define diversification in a different and meaningful way. And you have to accept or not accept that perspective as being valuable to your asset allocation process. And I'll just say what is usually the next step if that happens? Well, now you start to play around with managers. You start to see how can I introduce more betas to other factors that Ben has. And then when you are diversified across Ben factors, there's this feeling, you know, understanding this risk framework. Now I really feel diversified in this risk framework. So there's a clear payoff to it in terms of I have this surprise. Now what do I do about it? One of the biggest challenges when working with institutional portfolios is the incomplete transparency you may have, particularly with the alternative managers that they allocate to. And here I'm thinking about both privates were maybe the return data has issues in itself, but also hedge funds or emerging managers with limited data. Those that don't necessarily show up as much when you're talking about your average wealth client who's allocating to highly liquid mutual funds and ETFs. How does Ben approach this problem for institutional allocators? My first thought is to actually go back to this idea of the total portfolio approach. So the factor engine is this necessary tool to do the total portfolio approach. Now let's put aside for a second, you have to select which factors to use for your model. You have to manage and procure the data. You have to make sure everything's independent with each other all that aside. If you're using a holdings based approach, you also need the holdings for your portfolio. If you have a hedge funds or private assets where you're probably not going to get those holdings, even if you did build a factor lens and do all this great stuff in order to to conduct that holistic total portfolio analysis, you're basically ending before you began because you just don't have the data to do it. So that's one of the fundamental reasons why then took a returns-based approach. The only real holdings that are happening are the holdings used to build the factors, but then's doing that. And then we spit out the factor time series of return. So you as a user, you're either getting from then the return of the S&P 500, or you're uploading your own returns, and then it's purely a statistical or regression-based approach from there. For a hedge fund, with no holdings, as long as you have enough return data, and you believe the quality of that return data, you can do statistical analysis on it. You can see maybe they're turning over their portfolio of ton, but they always have momentum exposure. Maybe I could get that within ETF, or maybe they have tons of residual that can't be replicated and it's all been positive. So that's a great thing. Private assets get tricky because you mentioned they have all types of problems when it comes to time-weighted returns for private assets. It's not just about having the data. Often that's just the first step. If you think about a private asset index, for example, like a private equity index, it's going to have some major issues with it. So it's probably going to be smooth. I say probably, almost certainly it's going to be smooth, which is going to lead to lagging public markets. So public markets rally, this index may rally three quarters later or something like that, and artificially low volatility. And these two components you may have heard, Cliff Asnist referred to as volatility laundering, which is basically the idea that unlike public markets, which are going through price discovery, these private assets are being marked evaluation. How they're being marked evaluation is often a black box and creates the opportunity to market in such a way that the smooth volatility profile emerges. So that's a major issue. Now you can imagine there go your correlations with the liquid factors that you're trying to regress against. Another issue is they're usually only quarterly. So if you have total portfolio analysis that you're trying to do on a daily basis, well now you have to downsample to basically quarterly because the private asset you're trying to include is part of that. And now you're only capturing the risk dynamics of the slowest moving parts of your portfolio. And then private assets typically are super out of date. There are as are like three quarters ago or two quarters ago. So the idea is how can we use return enhancing features to try and get us closer to what marked to market version of private asset might look like. And you can imagine as a return based platform, then puts a lot of time and effort into return adjusting features to help you get better time series. So in the case of private assets, there's three things that we use that are similar in spirit, dismoving, interpolation, and extrapolation. And all three of them rely on a good public proxy in terms of using that as a reference point. So if I try to use these features on a private credit index with Apple as my public proxy, it's going to be garbage and garbage out. Dismoving is what tries to mark the private asset to the market. That typically raises its volatility. It's definitely the most complicated in terms of what's going on in the back end. For that, we use a process from Gitmansky Lo in Macarov from a paper in 2003, basically reverse engineers, the smoothing process. Interpolation were changing the private asset from quarterly to daily, again, using public proxy as a reference point. And extrapolation estimates where it would be up to date. Extrapolations regression-based, to smoothings regression-based, interpolations more like straightforward arithmetic. But at the end of the day, the goal is to get a hypothetical, what if my private assets were marked to the market, what might they look like? And you can imagine not only does this enhanced factor analysis and make it more interpretable, you tend to get lower residual, because now you're correlating more with the factors and things like that. You're explaining more of the risk. But also just from a workflow perspective. I need to go to my investment committee and give them an idea of what my portfolio looks like today. I can't be doing that as of six months ago. I have 10% in private assets. 90% of my portfolio is liquid daily. I want to use my total portfolio in a daily lens. So it's really about just getting somewhere that is probably academically more correct. But at the same time, unlocking total portfolio analysis in a meaningful way. So you have these three techniques, de-smoothing interpolation and extrapolation, to ultimately transform what are private asset return streams into something closer to daily, marked to market time series that you can get a more robust analysis on. But ultimately, these are synthetic return streams. They're estimates. And maybe you can talk a little bit about the limitations. I would presume that an institution that has 20 private equity managers, you can probably just talk about a generic private equity index there versus if they have a single private equity manager, there's a lot more idiosyncrasic risk. So talk about maybe a little bit of the friction and the difficulties there. And what that ultimately translates to in my question is how confident should allocators be in the factor conclusions that are drawn from these synthetic return streams? Yeah, so just quickly to that first point you made, definitely we're working with indexes here, because you do see a certain amount of capital inflows and outflows in order to transform those IRRs to time-weighted returns. We're working with representative index exposures here. How do you trust those synthetic return streams? I'll get into that. But I just want to make sure we know where we're starting from. How do you trust the smooth returns? You're starting from a place where you have the same level of mistrust. So if you look at something like a broad private real estate index, if you go back to 2004, that's got a vol of 10%. If you look at public real estate over the same period, it's 23%. Now you can say what you want about the opportunities of private assets. We certainly think there's tons of opportunity in private assets. But I do think it would be ill-advised to think that that 10% vol, 13% points less than public markets. So that's the starting point. So then the question is, let's use the smoothing as an example, because I think that's probably the most unique of the features. How do you verify that this is getting you closer to a better answer? I always think about this in Giggle, because there's no marked to market private asset return stream. So you're building this model to dismoose something, to make it closer to a marked to market return. That doesn't exist. So how do you get closer to something that doesn't exist, which is always this funny idea to me? This is where, in order to do something like model validation, you can use Monte Carlo simulations. For example, we would build our own 1,000 quarterly marked to market return streams. Then we would smooth them. And then we would use our dismooting process to see if it's actually getting closer to those true underlying returns. And basically, we do that by measuring if on average, the D smooth returns are closer to the true returns, then what the average difference in returns are for the smoothed version with the true one. Basically, are we getting on average closer than if we did nothing? That's the idea. So it's dismooting getting you closer. Of course, there's certain conditions that have to be met-- like an example is correlation to the public proxy. There's a certain correlation that needs to be met there. So we have these certain conditions that we recommend. And in those conditions, 90% of the time our desmooting model gets you closer to what the true underlying economic returns are-- again, asterisk. That doesn't exist, but our Monte Carlo simulations. 90% of the time that happens. Now, if the conditions aren't met-- for example, public proxies not correlated enough, that's an example where we won't recommend to smoothing in the app. I think the takeaway is, for allocators, have confidence that these are empirically validated ways that are trying to improve your understanding of risk in what is a undoubtedly opaque, no-win scenario. But stay humble in terms of these models or simplifications. You're trying to get a better idea of what market-to-market risk might look like. I personally think you're certainly getting closer to a true underlying feeling of what the risk is, especially when volatility is just so much lower than public markets. It's an easy place to start from in terms of trying to improve on that. But again, any type of private asset modeling, there's got to be a certain amount of humility involved. And just making sure you're understanding it as a form of context. And I think specifically, in terms of total portfolio analysis, a way to break down the barriers that private assets introduce, like quarterly, not being up to date, and just allowing you to do public and private analysis side-by-side. The return space analysis that we're talking about here is really the only option when a full holdings look through isn't available. But it does come with some of its own problems. So as an example, I'm imagining an institutional allocator that has meaningful asset allocation shifts over time, or has meaningful manager turnover. What that means is that old returns aren't necessarily representative of the portfolio as it currently stands. How do you think about the interpretability of factor results in that context? And where do you, again, draw the line as to how much manager turnover is OK? How much allocation shift is OK that the general results can be trusted versus, hey, this is going to be garbage and garbage out? Totally fair question. Obviously, return space analysis allows a lot of getting over data hurdles, but there's cons to that too. Let's talk about an example where maybe you have an endowment. And let's keep it simple. Let's just say they have a 60/40 allocation. So they've had that allocation for 10 years. And then one day, someone from the investment carry a wakes up and they say, you know what, we're going to make a Navy 20. We're going for it. [LAUGHS] So that day, your beta is going to look exactly like a 60/40, because that was estimated using those 10 years of returns. And as every day is more of that 80/20 portfolio, it's going to start to tilt more and more. But that's going to take forever if you have 10 years of history. Return space factor analysis does not pick up on these asset allocation changes immediately. So what are the ways around that? You could shorten your window. the less you make that window, the more. percentage of contribution you could call it the 80/20 allocation is going to have on the betas, then you can do rolling factor analysis. So it's not just about literally shortening the range, you could look at trends in rolling three year one year. So if you look at rolling one year, portfolio over 10 years, the most recent one year period is going to pick up that 80/20, the equity betas going to drift higher faster. But even then, that's not a magic bullet. It's not picking it up immediately. So the most realistic way, if you're trying to say my portfolio today, how do I use this factor analysis to model and stress tests what it would look like in the future. If the past returns aren't representative, you use a pro forma portfolio essentially. You would just put in an 80/20. You would build that in Venn. You would basically throw out the actual portfolio returns that you've had and you would just model using an 80/20 over the last three years or five years. Now you're still implicitly assuming that those past returns will be informative for future modeling, but now you're using a pro forma 80/20 which should represent your risk that you just introduced to the portfolio. So that's the basic idea. With that being said, I think it's totally fair to say that's not a solution for everything. Returns-based factor analysis is definitely more powerful for strategic asset allocation that aligns with most institutional portfolios. Again, the parsimonious approach, the factor lens, returns-based analysis, this is all gear towards the strategic asset allocator. It's great for analyzing hedge funds, but not necessarily the hedge fund manager who's trying to trade stocks and bonds and things like that. So where do I not trust it? It's really when you have complete regime shifts. It's when you don't trust in the past return patterns. You can't build a pro forma portfolio. The relationships between the factors themselves may not be as expected in the future. Those types of regime shifts are where you have to look at yourself and say, like, I understand the limitations of the tool. This is an example where a higher frequency holdings base, something like that may be useful as a supplemental understanding. And in general, returns-based analysis is one piece of a puzzle. It's something, holdings, returns, they can all work together to paint a more full picture. So that's some of the examples where it can be challenging and it's just about understanding the limitations of the tool. I want to talk about the interpretability and actionability of these factor-based models. Along the same lines of an earlier question, orthogonalization can lead to some account or intuitive, but intentional outputs. For example, you can see interest rate and equity exposures showing up when you compare FX to equity versus FX un-hedged equity portfolios. How do you think about educating users on interpreting those outcomes, particularly where the goal is not just information, but action? I love that example. Corey, we're going to go on a journey here. We're going to try to get real detail oriented in this and go through this example. What I like about hedge versus on-hedged equities, again, the assumption here is that the underlying equities are the same, the only difference is the currency hedge. One of the unique opportunities where then clearly shows the output of residualization in the tier system everything we talked about. The intuitive at face value, what you would expect when you look at the relative return stream between a hedge and un-hedged, is that then foreign currency extractor would explain everything. I think that's the first layer of thought intuition. So in a hypothetical world, let's say you had lower dollar, lower rates happening at the same time. Stocks hold held in non-USD currencies would benefit from that un-hedged without perform. In this world where rates and the dollar are falling, that correlated risk essentially is going to explain the outperformance of hedge versus un-hedged. But where do you put that correlated risk? Do you assign it to foreign currency? Do you sign it to interest rates? This is a Venn decision. You could do whatever you want. But that correlated risk for all the reasons we mentioned, we think interest rates are more liquid, more actionable, higher tier. So we choose to take that correlated risk and put it into interest rates. When you're looking at the relative return stream between hedge and un-hedged, not only do you see foreign currency, but you see that interest rate beta 2, which is pulling out that correlated risk. Now it doesn't mean that there's no foreign currency risk. It's actually the opposite. Foreign currency is still the largest relative beta between the two return streams. It's just that represents the unique uncorrelated part with interest rates. Just to be clear, interest rates is not the only one equity. There's other macrofactors that it's residualized with. So now you've identified this correlated risk. You've chosen to strip out that piece and put it in a higher tier macro factor that is more liquid and actionable because we're talking about, remember, interest rates factors just bonds. And on that same note, because it's just currency hedge bonds, imagine you're an allocator, you probably are familiar with that exposure throughout your career. You know which managers have some exposure there. You have a good idea in your portfolio. What are the levers I might pull to affect change in that interest rate beta? That's part of why those macrofactors are built to just be broad actionable indexes. So now you're looking at the difference between the two and you want to hedge some of that risk, you're not forcing yourself into only using currency forwards or something like that to affect the foreign currency beta. Now you're able to look at interest rates since the advancements like bonds or things of that nature. That was a lot. Think about running analysis on then you're not going to do that mental exercise every single time. You're not going to scrutinize why this little beta is popping out somewhere because we had this perfect example. We know it's residualization, but when you have a multi-asset portfolio, you're not going to necessarily get that clean feeling of this one factor and then how it's being other factors are being stripped out of it. So I do think that first and foremost, it's important when you are using a risk model to understand the risk model you're working with. For then in particular, it's really the tier system. The residualization process is this something that I believe in because at the end of a day, a factor model is surfacing these little volatility and correlation nuances that you can't just read that in a headline somewhere. So there's some level of expecting the unexpected with a factor model. So I do think it's about accepting the risk framework and then I don't want to say ignore the analysis. If you have a red flag going up and you should absolutely do that, it's good to be skeptical. But I do think you need to start by accepting the process and you can't question every little unintuitive thing when you're working with a factor model, at least in my opinion. We want to talk about stress testing and scenario analysis. Most allocators think about these in terms of asset classes and historical events. That's often how I see this type of analysis performed. What does it look like when you move into the factor space instead? And what sort of insights can that unlock? So we have something called sensitivity analysis. Imagine a user goes in, they pick a market index like the S&P 500 and they say if the S&P was down 5% over the next 30 days, what might multi asset portfolio do? What we do is we use regressions and we translate that index shock, the S&P down 5% into a factor shock. You can imagine in this simplified example that probably looks a lot like the equity risk premium in Venn's factor lens going down around 5% something like that. So now once you've translated that index shock into a factor shock, if you know your portfolio is exposures over, we like to use the last three years as representative, you can translate that factor shock into a portfolio shock. That's the idea. And just going back to our other example, if I was going to do this and I just switched to an 80/20, I'd be shocking a pro forma 80/20 portfolio. Now everything I described index shock to factor shock factor shock to portfolio returns. That's linear and it's based on historical factor relationships. So it works best for mild to moderate shocks. Factors, as we know, are not always well behaved, especially in a crisis for example, you can actually see the covariance matrix, the correlations between the factors change. So this is where we use something called a Gaussian mixture model. And this is essentially for extreme user shocks. So now we're talking the S&P down 20%, not 5%. So what a Gaussian mixture model does is it basically takes those all the returns of the two sigma factor lens and it clusters them into overlapping normal distributions. Now it's a machine learning technique. We don't know what those clusters are. They're not named for us. But when you look at the factor returns and the profile of all the, it's clear like they're kind of economic regimes. For example, there's a crisis economic regime or a cluster and the equity factors down, I don't know the number off the top of my head. Maybe it's 50% 40%, whatever the case is, it's clear that that's representing a world with crisis happening. So the four clusters and regimes that we use are crisis steady-state inflation and walking on ice. Basically what happens is when a user puts in a big shock, it will try to map what the probability is. A little bit out of 100% that it's a crisis. It's a steady state. It's an inflation regime that we're working in. And then there's basically a weighted average factor covariance matrix that goes in to translating that factor shock into a portfolio impact. So that's the way we try to model the abnormal factor relationships that can happen in extreme environments. That's when you get a breakdown of diversification, the systematic relationships that you know. And then ultimately what that means for the portfolio or asset that you're analyzing. That's just one example. This sensitivity analysis like you can also, I won't go in depth, but quickly I know my factor exposures today. Maybe I only have five years of history, but I want to see what the portfolio would do in 2008. Well, you have factor returns back to 2008. So you can go see, you know, given my exposures today, what might my portfolio have done during the global financial crisis, stuff like that. But again, all of it is on the assumption that those past return patterns are indicative for modeling. One of the other outputs that then provides. or produces this estimate of factor contribution to risk return and diversification. And I'm curious how you help users translate these diagnostics into action in their portfolio, especially when portfolios as traditionally constructed don't usually allow for direct rebalancing along factor lines. Moving one manager or asset allocation choice will impact multiple factors simultaneously. This is a great question and I think it gets like a higher order theme here. Where does ven sit in the asset allocation process in total? I mentioned about factor adoption earlier and we're closed, but we're not there yet. Institutional clients are not typically factor rebalancing. That's not in my experience where we're at today. As a class rebalancing and that form of thinking is still where people are starting from. So for me, all of the factor output you mentioned contribution to risk return diversification. This is really just empowering investors to be more transparent about their understanding of their portfolio. To surface hidden insights and hidden risks that they otherwise wouldn't be able to access because those volatility and correlations are not always jumping out at you. And that's what the factor model brings to the table. And really just to take the entire asset allocation process and be a lot more intentional and sharper with the decision. I don't think anyone's out there doing an entire investment plan using returns based factor analysis and rebalancing in that way. But there may be asking themselves does my high yield bond manager really belong in my fixed income sleeve? Does that make sense or are they actually more equity like? Is my hedge fund manager really earning their fees or are they just exposure to systematic risk? Am I actually diversified with all of my asset classes? I think just adding this layer of insight promotes better discussion, better portfolio building, better decision making. I think that in and of itself is cutting through complexity which promotes action ability. And then again, the way we approach the actual analysis itself, you know, parsimonious approach being built really with asset owners in mind. It's really just making that process easy for them. I don't know if I answered your question directly, but it's really just finding where returns based factor analysis fits as a way to just make any investment process more robust, more transparent, not necessarily just blindly trying to rebalance by factors and accepting all of the output. Chris, you guys have written quite a bit about Vens approach. You've been very transparent and public on his podcast, but again, in the literature that the team has written if people want to learn more and dive deeper into the actual process, where can they find information? So we have a website, event by two sigma's the name and there we have a blog with insights and resources. We have a help center where we really get into how calculations done or things like that, but I would definitely start with the blog as a place where people can come and learn more. And then there's tons of links all over that where someone can find an email or click on a form that would put them in contact with us. If they really want to see how their portfolio might look through a factor lens or Vens factor lens or things like that. So I would start with a website. I will just say as an independent source, you guys provide enough transparency that I was personally able to go through your methodology replicated in house, come up with the same results that you guys were actually producing in Vens, which was really nice because it gave me a lot of deep understanding of how your process works and then to see a translate one for one was really phenomenal. So I love that transparency. I went to the episode I want to ask you the same question I ask every guest this season, the season is dragging on far longer than I expected. So we'll continue with the same question though. What are you obsessed with today? This could be an idea. It could be a book. It could be a show. It could be something you're doing physically. Just what is something outside of work that you're currently obsessed with? Oh boy, I mean, I know the exact answer that I want to say, I'm just going to go ahead and say it. So I collect Pokemon cards. Okay, Corey, let's just get that out of the way right now. What's really interesting about Pokemon cards is they have this entire market to them sealed versus unsealed. Each Pokemon card has its own time series trend and I've been obsessed with thinking about like can you build a Pokemon card index that you could use as a benchmark for that market because I got news for you. If you took all of the different sealed products and bought one of them, I'm pretty sure I haven't built the index yet, but I'm pretty sure would outperform the S&P 500 by a lot. So I don't know, I think about that a lot. Like what's the methodology I would use which cards would be representative what I use seal collections what I do equal weight what I do market cap weighted where like charizard the charizard card would be 70% of the index or something like that. I don't know if that's a good answer or bad answer, but I think about that all time. As someone who grew up with Gen 1 Pokemon on the original Game Boy, I am all in with you, but I'm a little disappointed in the answer and that it's not you weren't building a factor model for the cards. Yeah, well, this is an example. So because this return based analysis, if you built a time weighted return series of Pokemon cards, you could upload that time weighted return into then and see if it correlates with economic growth or the equity factor. Because one would think when there's good economic growth, you would be buying more Pokemon cards and when economic growth is not doing well, you would be buying less Pokemon cards. So there could be that the Pokemon exposure is highly explained by the equity factor. We'll see when I build the index if that's the case. We'll have to do a follow up of Chris. This has been fantastic. Thank you for your time. I really appreciate you joining me. All right. Thank you, Cory. It was an honor and thanks for having me. [BLANK_AUDIO]