My guest in this episode is Sandrine Ungari, Head of Cross-Asset Quantitative Research at SocGen.
Sandrine cut her teeth in the industry as a fixed-income pricing quant, but made her way over to sell-side, investment quant research in 2006. Her early research focused on credit and macro, but since 2012 has been heavily focused on equity and alternative risk premia.
Our conversation begins with equity factors and Sandrine provides insight both into how factor construction has evolved over the last decade as well as her thoughts into where the field is headed. We broaden our discussion to include alternative risk premia, and Sandrine provides a useful mental map for categorizing this broad range of strategies. We discuss the risks of crowding, latent beta risk in levered factors, and the influence of macro economic factors.
More recently, Sandrine has focused her research in the application of machine learning in strategy construction. We discuss one particular example – the application of a recurrent neural network in trend following – and Sandrine shares her views as to how machine learning might affect factor investing going forward.
Sandrine also shares some interesting ideas about where future risk premia might emerge from – but you’ll have to tune in to hear!
Please enjoy my conversation with Sandrine Ungari.
Corey Hoffstein 00:00
321 Hello and welcome everyone. I’m Corey Hoffstein. And this is flirting with models the podcast that pulls back the curtain to discover the human factor behind the quantitative strategy.
Corey Hoffstein Is the co founder and chief investment officer of new found research due to industry regulations he will not discuss any of new found researches funds on this podcast all opinions 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 in securities discussed in this podcast for more information is it think newfound.com
Corey Hoffstein 00:49
My guest in this episode is Sandrine Gauri, head of cross asset quantitative research at sock Chen. Sandrine cut her teeth in the industry as a fixed income pricing quant but made her way over to sell side investment quant research in 2006. Her early research focused on credit and macro, but since 2012, has been heavily focused on equity and alternative risk premia. Our conversation begins with equity factors, and Sandrine provides insight both into how factor construction has evolved over the last decade, as well as her thoughts into where the field is headed. we broaden our discussion to include alternative risk premia and Sandrine provides a useful mental map for categorizing this broad range of strategies. We discussed the risks of crowding, late and beta risk and levered factors and the influence of macro economics. More recently, Sandrine has focused her research in the application of machine learning and strategy construction. We discuss one particular example the application of recurrent neural networks and trend following and Sandrine shares her views as to how machine learning might affect factor investing going forward. Sandrine also shares some interesting ideas about where the future of risk premium might emerge from, but you’ll have to tune in to hear please enjoy my conversation with Sandrine and Gauri Sandrine, welcome to the podcast, really excited to have you here coming all the way from France getting quite a bit of a time difference here and recording but super excited to chat. I know that before we do you have a little bit of a compliance memo you need to read for the listener. So why don’t we get that out of the way? Hi, co
Sandrine Ungari 02:35
Yes, so I was just wanted to say that us express here I’m a personal views and neither solicitation your role. Now its subsidiaries affiliates accept any responsibility for liability arising from the use of all or any part of the material from this interview. That’s all I had to say curry.
Corey Hoffstein 02:57
Wonderful. Well, that tees it up nicely, because it’s clear here that you do work for SOC Gen. And my suspicion is that there’s a large number of my listeners who unfortunately, don’t have access to your wonderful research and may not know precisely who you are. So why don’t we tee it off with your background? Maybe you can tell us a little bit about how you got into the industry and came to the role you’re in now.
Sandrine Ungari 03:21
Yes, sure. Thanks a lot for the question. So I guess I started to for quite some time ago in in financial industry, I started to work as a quant as a price or quant. So originally, I was designing pricing models for interest rate derivatives. And then I felt into the bucket of investment costs. So I started to work for selection in 2006. In the role that I’m still on at the moment when I started to look at markets, financial markets through the lenses of models using older models that I had, carefully, painfully, maybe studied before. And that’s how I started my journey into investment strategies. Back in 2006, we were mainly concentrating on relative value trades from a discretionary background and either credit or interest rates. And when 2008 crisis hit, while credit went much out of fashion. And we started to get more interested into systematic strategy and deploying our models for systematically trade markets. And that’s how I guess starting in 2011 2012, something like that. They started trading to be interested in systematic strategy and using back testing frameworks to test the validity of models. And that’s where my journey into risk premia investing, get it all started. And I guess that brings me in front of you today just to discuss what systematic investing and risk premia strategy is all about. That
Corey Hoffstein 04:59
I am Definitely excited to dive into that. One of the things I wanted to ask you about though, was this podcast definitely has a bias towards buy side investors, a lot of the folks I end up speaking to are asset managers. I would love to get your perspective as someone coming from the sell side, what do you see as being the primary differences between buy side and sell side quant researcher?
Sandrine Ungari 05:22
Yeah, that’s why I was looking at all the guests that you hosted so far. And I was quite flattered to see that I was the first sales psych phone. So thanks a lot for that. I guess there. First of all, in on the sell side, you find a lot of quotes, but there are more pricing quants. With modeling quants, they have a role in the bank as to help the bank to hedge products or design products. They’re very much productor and hated specialists. So what I’m doing is much more in line with what my clients do on a lie site. So that’s something that I would call investment quotes. So as investment quants, we’re interested in to the dynamics the markets, but not not the dynamics that allows you to price structure and complex products, but more into the dynamics of the market that will allow you to make money at certain points. So in that sense, although I’m working on the sell side, my job is very similar to what buyside quant would be interested in. So I’m very much focusing on statistical learning how to use statistical models to calibrate systematic strategies, how do you think about portfolio location while taking into account historical correlations, left tails, behavior, distributions, historical distribution, etc. So I say an investment course is, if you put it in a modeling language, an investment quant would be very much interested into modeling the historical distribution on the historical measure, while cell cycles more interested into pricing would be more interesting in modeling the risk neutral measure. So in my work, I’m more interested interesting into the historical measure. Even though my job will resemble very much the job of a base, I just tell differences there, because I’m not managing money. And I guess that’s the main point. And as I stated, to start with, I’m not my views are not linked to any company’s views or any in AUM, basically. So as such a more guest my role with respect to my clients is more of a consulting role Investment Advisors role, investment advisory. So in this perspective, gives me the opportunity and quite unique chance of looking at a wide range of investments use cases, because I’m not working for only one investment managers, I’m working for potentially all our clients. And that gives me exposure to a lot of different business cases, every asset manager out there have their own institutional mandates, their own investment constraints. And the great thing about being a CSI quant is that you’re you get a little bit more exposure about what the clients are potentially doing.
Corey Hoffstein 08:15
That is one of the things that I noticed when reading research from sellside investment quants is that it does tend to span a very wide breadth of research agendas. It’s not just focused on how do we make value investing work, it’s all sorts of different applications. Whereas on the buy side, you obviously tend to see much more of a focus around whatever product that firm is making available in the market, their research naturally tends to focus much more heavily in that area. Given that open ended mandate, how do you think about structuring your research agenda? It seems like you’ve got a wide field to play in, how do you know where to go?
Sandrine Ungari 08:53
That’s a very good question. It gets me scared sometimes. Because you can go so many different directions. And at the end of the day, you still have to make yourself even to your clients. So even though you have the freedom of choices in a way, you’re still linked to basically staying relevant to your clan base. And I think that’s, that’s what’s guiding in very loose terms. What is guiding my, my research agenda with how do I stay relevant to my own clients? What are the topics that we’ll still be discussing a year, two years, three years times, and in a way because we’re unlucky enough to discuss with a lot of people. So whenever you have a new trend emerging, you sort of if you’re careful enough, and if you know how to listen enough, you can catch that. And then I guess it’s how all that research on risk premia invest invested good started in 2012 is when some sales guy and some came to see me and they said, Oh, you know in the equity world At the time, I was like fixed income quants, I was very much interested into swaptions religious strategies using interest rates derivatives very much into trying to measure correlations in financial markets. So nothing to do with factor investing is permanent. And when they came in and they say, oh, but you know, lots of people are talking about factor investing in equities and risk premia, investing in equities. And from my fixed income background, I said, Yeah, always promotes it says it’s scary, right? It’s like you’re selling a 10 year bonds and you’re getting some duration premium info, that’s risk premium. Say, No, no, it’s not that it’s about trend following quality investing. And then I started to dig up. And well, I guess, seven years down the line, if that topic is still relevant. So the time we had to make a decision to start to research about that. And I guess that’s how I tried to get my research agenda to back in 2010. We had some colleagues coming back, they went to do some studies in at Stanford and they studied with Huston Tibshirani, the guys who thought the elements of statistical learnings, which to me is like the very first book you need to read if you want to do machine learning. And back in 2010, no one finance would talk about machine learning was just a concept. And we said, oh, this is it, the elements of statistical learning is exactly what we’re trying to do. We’re trying to learn statistically, from financial markets, let’s try to apply all those techniques there. We started to develop a lot of models using their books in the business cases that we were facing. And I guess that’s how little by little is relevant because that topic became so much so much popular.
Corey Hoffstein 11:48
And just for clarity, I probably should have asked earlier who are your clients
Sandrine Ungari 11:52
that many institutional investors. So all the range of ensuing distance I can find out there ranging from pensions, to sovereign wealth funds, private wealth managers, we also took a lot to asset managers, institutional asset managers, and also hedge funds. So it’s very much folder. We’re covering pretty much all institution clients of the bank.
Corey Hoffstein 12:18
So as you mentioned now, a few times that sort of cross asset risk premia and equity factor investing really seemed to take hold in the industry around 2012 2013 really accelerated and I can speak particularly from a US perspective here in the US, at least, really seemed to accelerate 2013 through 2015. I would love to get your perspective now that we are closing on on eight to 10 years of real adoption in the space. What do you see as the future being?
Sandrine Ungari 12:47
I guess that’s the good thing about quant investing in general. And whisper investing in particular is that’s an area of innovation. So it started in the 70s, with fama and French, all the equity guys and equity factors. It’s started to develop in other markets but that’s that’s an area where really innovation is focusing on that reminds me when I was working as a quant in 2000 2005 innovations in banks was lying into pricing models like all the HGM BGM Heston’s stochastic volatility models, whatever you name it, and that now that innovation, it’s not the field of innovation anymore in a way, because the models are there, they’re very well set up or not so well set up but they’re there and then you Deputy us. So there’s no more new developments there. And the new developments are really up happening in those investment strategies. That is that falls under the labeling of risk premia is much broader nowadays and risk premia. You can extend that to quantitative strategies. So in the jargon, we call that Q is quantitative investment strategies, that groups together risk premium strategies, but also other strategies that relates to hedging, for example, systematic hedging, or that relates to some execution algorithms. So I think there’s a lot of innovation happening in the field going forward beyond the risk premia concepts, the future of all that is really trying to trade systematically markets to provide investors, institutional investors, but maybe down the line retail investors also are the tools they can use to better manage their portfolio. There’s maybe two discussions there. This is the classical risk premia discussions, which is how you do generate alpha in general, and that’s the really, really challenging one. Then there’s this other discussions which is how can you use this ematic strategies to arrive to better pot for outcome, because you’re doing things systematically because you’re putting in place hedging overly, because you’re doing some very clever risk management of your portfolios. So a lot of opportunities there.
Corey Hoffstein 15:18
In one of our prior conversations, one of the really interesting comments that I sort of kept in the back of my head was this idea of risk premia going from not just something that exists in the market, you can capture but potentially something you can transfer between parties. So this idea of saying, well, the bank for perhaps regulatory reasons needs to offload some type of risk to an institution, that institution might be interested in holding that risk because it might have an a premium with it, but it’s just something the bank no longer wants to hold. How much of your focus is on that sort of approach to the future of risk premia
Sandrine Ungari 15:56
part of it. So that concept is very interesting because it’s the concept of what I call alternative what I called trading premium. So in that whole spectrum of risk premia, you have the academic premia, which are fat, the fama French type of premium momentum, value quality, growth, profitability, with which a lot of investors have made a fortune, right, we are sort of seeing the returns livering out from those tiles on those premium, juicy value investing, for example, is facing great challenges. Louisville investing has faced quite a few challenges in the past crisis. So all those premium have have their own challenges. And what are we seeing is that getting aware of those Academy premiere, you have a new range of Premiere, which I imagine which are called trading premium, which are more or less academic in the sense that there for some reason, that Academy have been less studying those strategies or those premium because of data access because of very fine practitioner knowledge. But there still exists. And one example of that, as you mentioned, is result the regulations, the banks have been constraints and the balance sheets to do some form of activities. So there, nowadays, if a bank wants to do structured products have limits, and very exotic things like correlations, like SKU like Volvo, like recovery rating credit, for example, there’s limits on all those implied parameters, once bank have done a lot of structure products as parameters may be saturated, which means that the banks cannot do the business anymore. So there’s a way to, and there’s an interest for the bank to recycle that risk to final investor who would be interesting in, in, in hosting the risk and taking the risk on behalf of the bank. And that sorts of activities, it was was very prepare before the COVID crisis. It’s slowing down now. But it will come back eventually, because there is a real economic rationale behind it. On the one side, the bank needs to offload that risk to be able to make more business and then the part of fits activity. And the final investors is happy to carry that risk because he gets paid for it. So there are a few strategies like that, that were challenged with that is to do the to be able to package them in a in a systematic way. Few of them are doing AdWords ads doing that things like I’ve worked on sure those like crypto arbitrage, for example, strategies allowing you to get exposure to the slope of the cost of funding, which is a way to provide funding and the very long guns to investors who need it. Or think like cabbage in arbitrage on dividends, for example is another example. Although dividends are a little bit more challenged at the moment for whatever reasons, but yes, it’s alternative alternative risk premia alternative squared if yield.
Corey Hoffstein 19:12
So I’ll return us back to maybe the more traditional alternative risk premia. One of the things I have noticed is that a lot of the early academic papers within the world of factor investing were utilized very naive sorts in portfolio construction. A lot of the initial products that came to market were very simple in their construction. Over time, a lot of the research has evolved in the application of how these factors should be incorporated into a portfolio and how that especially on the equity side, how it should actually be constructed. What is your perspective? What have you seen in the evolution of factor construction over the last decade,
Sandrine Ungari 19:53
the market that is the most mature in terms of factors is equity factors, so I think we’ll have the opportunity to get in Back to other type of factors, but let’s focus for a moment on equity factors. Equity factors have had quite a good time, I’d say until 2010, something like that, the discussion around the construction process of equity factor wasn’t so, so relevant. Because when you have performance, you’re less concerned about the contraction, the construction of your factors. So it was more like you rank stocks on a given criterion and then give an indicator. So that can indicate I can be anything can be the past momentum of stocks, it can be the past mean reversion, it can be the level of debt, it can be the average and resentment. So you take all sorts of metrics for your stocks, you rank them from the best to the worse. And then you construct the top quintile, versus the bottom quintile, the top quintile, versus the bottom quintile, everyone, every quad has his own tastes. And originally, the way that it was done, it was just equally weighted, so you would just buy the first contest sell the worst quartile, and you will construct an equally weighted basket of long stocks as short stocks. So a lot of people are doing that. Now, quite earlier on in our research, where we say that you shouldn’t be doing that you shouldn’t be shorting the bottom content of your indicators, because let’s say you have a very good indicator to identify value stocks. So that means that you have something take the P E ratio, for example, the pitch, GDP ratio tells you which stocks are very much undervalued. That’s indicator, now, it might not be a good indicator for auto value stocks. So very early on we say well, you should we should not really do long the stocks versus short the stocks, you should be doing long the stocks versus short the market. So short form of market cap index, now, you should do that, you start to see some biases there, because your long basket of stock is equally weighted is not in Sector neutral to your market cap index. So you start to introduce a lot of biases that are not necessarily desirable, like sector biases, basis and other metrics geographical basis. So while we had done quite some times back, we had made developed a way to allocate two stocks changing the equal weighting scheme to a scheme that would help us to cancel out the basis. So we would have a long value stocks that would really be values doctor would be no doubt biases, no market cap, no size bias, no, no sector bias so that whenever a sector is over performing versus versus another one, we don’t get that exposure. So that’s I think, in terms of factor construction is the evolution that we’re seeing how that has happened. So people have been more thoughtful about how do you wait, you do your stocks in each basket? Do you short basket of stocks? Or do you shelter benchmark and all those sorts of things? The next step for me is that what more and more people are realizing now is what we’ve been doing as equity quants is data mining exercise is a massive data mining is society’s going back to Fama, French, which is to say, let’s, let’s look at sets of indicator, let’s rank the stocks according to this indicator, and let’s see if it makes sense. And you can make money out of it. So in those days, more and more people are realizing that you want to do data mining, it’s good. I mean, that’s what we all do. Data mining is not a bad word. It’s a good concept. But you should do it in a modern fashion. And to do it in a modern fashion, we have modern tools to do that, which brings back to machine learning. So you can learn, you can use statistical learning, or you can use more advanced that go even deep learning why not, but you can do things that will allow you to identify what factor is efficient to trade the stock market. And by doing so, you you have a construction process that is robust, not only come to the market cycle, but it’s also robust if you change regime, for example, if some factors are working based off some factor, working the soil.
Corey Hoffstein 24:36
So one of the things that just struck me as you were talking about this evolution of factor investing in early days when everyone had maybe and still does perhaps have these unintended bets, you see more dispersion among factor investors that two people implementing the same value portfolio maybe would have not as much overlap in what their whole Holding, is there a risk that as investors move towards more and more pure implementation, there’s greater convergence and crowding in these trades that could either lead to either greater disruption in the existence of the risk premia, or a greater risk of crashing?
Sandrine Ungari 25:18
Yeah, that’s the usual pushback that you get for the underperformance of equity factors. Sometimes there’s a risk of that. Now, if compared to where we were in 2007, for example, in the big quant crash, there’s a lot less leverage in the market. So hedge funds, equity, neutral hedge funds are much less leverage. So yes, you have a real crowding in some stocks, but you have probably less of according than you used to be before the total leverage was controlled by regulations. That the answer to that, in a way this machine learning can be an answer to that because if you see more and more people trading the same factor, then you’ll start to see that factor underperforming. And then if you’re adaptive enough to capture that, then you can deliver that factor in going on to another factor. So I guess the as returns have to all those factors have collapsed or have leveled out, you can do a much greater job and monitoring the performance of the sector and trying to add a timing dimension to those factors which can be brought into into the picture of machine learning us weather or weather timing mechanism, basically.
Corey Hoffstein 26:32
So on the same sort of line of thought is crash risk and maybe liquidity sensitivity. One of the arguments that you often hear about alternative risk premia, so maybe go and stepping outside of just equity factors. Now, broadening the scope to the full collection of alternative risk premia is that all the diversification seems to be there precisely when you don’t need it when equity markets are doing well. But in tail events, they all seem to be ultimately susceptible to the same liquidity margin collateral risk, how would you respond to this criticism?
Sandrine Ungari 27:06
Yeah, I hear that criticism very often. And also hear that you have a lot of hidden beta in tourist permit strategies. So to that, I want to say two things. First, is that we are in the in a world of long shots, long shots, press shot village. So most of those funds, most of those strategies are try to be neutral in beta, but one way or another, they’re either long or short volatility. So yes, they have some beta exposure, but they have some beta exposure through the volatility channel. So it’s not direct exposure, it’s because volatility correlates negatively with the stock market. So once you’ve said that, you’ve said it all. So what you need to manage in those funds is your risk, your volatility risk, premium that we had in 2017, that will start in 2017. But it also was the case in 2018, and 2019, is that yields were so low and so compressed, that a lot of investors engaged into some form of volatility selling and taking only one side of the trade, which, you know, is very ill designed portfolio construction process, because you’re, you’re not trying to diversify away your shoulder at risk into some long volatility components. But that’s where is that we know, we’ve been highlighting in our research for the past two years, at least, in our baby portfolio that we have, which are pure paper based. I’ve always had these sorts of tail risk hedge strategy to cover up the tail risk or volatility risk that you have in this portfolio, which manifests themselves, obviously, only when you have a crash in volatility, like February 2018, to some extent, December 2018. So all those events were Yes, you have Peteris keen to spell through that because you have volatility risk. The other thing that I wanted to mention too, is that you had the China had learning and then early 2016, you had rates starting keeping on going down so you have a massive invest of massively going into low vol and quality. And then in the summer 2016, you had rates going higher up that you had to speak rotation between Louisville quality into value that created a lot of losses in equity factors.
Corey Hoffstein 29:42
I mean, one of the things that quants I feel like quants often ignores these macro economic factors, we almost tend to focus again so solely on things like measures like value or low vol or quality and look at our sorts and create our long short portfolios. And I think maybe cry is our fingers and hope that via diversification, we’re avoiding these macroeconomic factors. But clearly they have some influence. So from your perspective, how should we be thinking about that influence when it comes to portfolio construction,
Sandrine Ungari 30:13
that’s a good point. And I like to go back to the definition of the macro factors. So I guess that two ways to define the macro factors one way, which is very quantity, so imagine that you have all the assets in your markets, and you’re doing a form of principal component analysis. So you’re trying to identify the latent factors, which are driving the market dynamics, and you end up having like a growth factor, a currency factor, a systemic risk factor, with monetary and fiscal policy factors. Most of cases with Premier strategies are very much neutral to those factors, because that’s the way they’re constructed, they constructed as long short versions of those assets. Apart from short volatility strategies, for example, because far too short variety strategies have this correlation to the growth factor through their volatility and the creation between markets and for So, if you look at market factors define like that switch PMS strategy tends to be relatively neutral. But if look at long term cycles, look, long term macro cycles we fill in, if you look, think about recession, expansion, contraction cetera. Or if you think about monetary cycle, which are long term cycles. Or if you think about price cycles, like how we secure GMO low risk regime, then you start to see differentiation in the strategies. So value, for example, tends to perform well, in the risk current environment intends to underperforming the risk of environment. Carry strategy tends to perform well in an accommodative monetary policy environment and tends to underperform in a in a tightening environment. So you do have to look at the long term pictures of those factors, you start to see emerging some form of patterns, which in turn, you can use to drive out the investments, depending on where you stand in the cycle. Now, there’s a big debate out there, which is timing versus non timing factors, I think is famous quote, from an asset manager and a guy from a research company, I’d say, both are right and wrong. There’s truth and knowledge to be good to have this from listening to a lot of different people. But what I’ve noticed is that in a world where all your strategies are performing really well, you don’t care about timing, because if she let’s say you have a Sharpe ratio strategy that has a Sharpe ratio of one, and you mix a bunch of strategy, which have a Sharpe ratio of one, then by the virtue of diversification, you’ll get to some fantastic outcome. In a world where premium Mac compressed, and your arrest strategy may return something like two three persons while it was returning maybe six or seven person before you get your potential losses that are much higher. So you can get a much better, you can do a much better job at trying to time your left tail, basically, because you have more currency. So that left tail, we’re in a world where premium are compressed returns are challenged. For whatever reason, being crowdedness being central banks being other reasons, then knowing the macro behavior of these factors, and adding a sense of timing in the platform construction process is very important and becomes more and more relevant. I think there’s a great deal of things to be looked at when you look at those factors across cycles, even if the measure is not perfect, even if the cycle is different to the next. Definitely the cycle that we’re in is different. But at least you get some sort of knowledge of what can be happening in the future.
Corey Hoffstein 34:11
When we look at the full sort of palette and spectrum of alternative risk premia, and we consider the how they exist across different asset classes, equity indices, currencies, commodities rates, we think of different styles that can be applied. And think of the things like when they apply within a given macro economic or monetary cycle. It all gets a little overwhelming, to be quite honest. Do you have a framework that you use to sort of think about how these different risk premia strategies are categorized and perhaps how they fit into an institution’s portfolio?
Sandrine Ungari 34:51
Yeah, well, do. We do a lot of work on that? So it’s, so it’s going to be a bit challenging to do it But an audio only because one of our greatest tool to describe HPMS strategy is what we call a minimum spanning tree. Which is imagine you see the London tube map? Yep, I’ve got exactly the same thing. But instead of London tube station, there are all different strategies. So that’s how the minimum spanning tree looks like. Obviously, there’s some simplification meaning behind distance in the minimum spanning tree. So whenever two points, so imagine two tube stations are very close to each other. That means that very well, they’re very well correlated. And that we went that distance, they’re very decorated or they don’t resemble to each other. If you want a minimum spanning tree that sorts of visualization technique. On to diversify portfolio of risk premia strategies, you’ll see appearing three different buckets for three different main groups of strategies, why don’t is quite obvious, there are mainly all carry strategies, although short volatility strategies, there are those strategies that are called risk on bond. So typic, typically the strategy that will do well, when risk appetite is high, when interest rates are rising, and they will do not so well in the opposite situation, then you have another group that I called risk of bond earn, which is quite clearly the opposite. So strategy that will do well. When risk appetite is low, risk aversion is high. And when rates are going lower, there are things like quality investing any form of defensive strategies, hedging strategies belong there to some form of carry strategy and rates to here. And then you have a third back Ed, I called risk on off, and they tend to perform well, either when you have a strong risk appetite, or when yields are going down, going up at some point. So typically, value investing is of that sort. So why value investing will underperform if you have a lot of risk aversion that will tend to perform well, if you have a rising rate environment. So when the bonds are underperforming. So once you’ve got these two, three big buckets, then you sort of start to see the sort of strategy mix that you want to achieve. Or you start to see, okay, if my portfolio is ultra defensive, I want to have more risk on strategy and vice versa. So there are tools to categorize strategies to know their risk profiles and to know what to add in a portfolio. And then the other thing that we’re doing is, we are looking, we are running all sorts of analysis to see, take a portfolio, let’s say, and because an investor will have its own style, maybe a portfolio manager will be more like Warren Buffett value investor, some other managers might be more like Nassim Taleb and might be more conservative and prudence. So you can run tools to try to define what factors exposure a portfolio manager would have. And then you can use this analysis to intend to do some form of completion to this portfolio to broaden up the scope of the portfolio or to to add some form of premium that can bring diversification and they can bring alpha. So there’s all sorts of analysis that we can do,
Corey Hoffstein 38:39
given the broad breadth of style premia that have now been defined both sort of traditional and non traditional. And it feels like so much effort now has gone into the evolution of portfolio construction. I would love to know do you think there’s still risk premia out there to be discovered? Or do you think we’ve sort of discovered everything at this point? And now it’s about sort of sharpening our focus and how we implement them?
Sandrine Ungari 39:04
That’s a good question. I think there’s a lot of things to be done in fixed income, there has been very few research and very few investment solution in the bond market. So now that market is very challenging, you have central banks, activity trading, that market to liquidity is relatively low. But in terms of research, it’s it’s a fantastic ground of discoveries. And it’s very different to equity factors, because that dynamic is completely different to equity dynamics. So I think I’m seeing great research coming out of asset managers, in particular, on that topic. That’s clearly a field that is an expansion that’s application, way beyond only the risk premia investing but also, you could see like smart BJTs for example, using bonds of funds from asset managers. So definitely Fixed Income is an area of discovery in a way. And then you have what we’ve mentioned earlier on is anything related to trading premium. So those form of alternative carry, that are being created by some flowing balances in the financial world. So you have some retail guys laying some structure product on one side of the planet, in Asia, typically, some banks in the US or in Europe having to hedge their structured product books because of the retail business in Asia. And that creates some distortion in the market in the pricing parameters of the market, which allows investors to in turn, harvest that premier by providing liquidity where that implied parameter is depressed. And I think that’s an area of innovation and discovery there.
Corey Hoffstein 40:54
We’ve danced a bit around sort of the evolution of factors talked about, sort of historically, they’ve been naive, linear sorts, and you’ve brought up machine learning a couple of times now. So I’d have to focus in on that a little. What impact do you think machine learning is going to have on factor definitions going forward?
Sandrine Ungari 41:12
Well, machine learning can be used as a tool to better construct and better discover the factors. So as I was mentioning earlier, you could put all your indicators that you want to use to trade an asset class into your machine learning, and just lets to the algorithm to decide on what would be the best indicator to best forecast the Sharpe ratio or best manage the Sortino ratio or best reduce the Jorah. And then you can introduce a new reward function in the machine learning Django that you want, so that you create the the algorithm that best fit your your need. So that’s something that we’ve been doing an equity factors with some successes, Mercy. So we’ve launched the strategy back in one year and a half ago, actually. So where we are, we have an index on Bloomberg ticking, and we have some TF a women that and machine or the the model learns what are the indicators that are best forecasting the next month return of the stocks, and that model perform relatively reasonably in the past two year and a half, in a period where that was really much challenged for factors for equity factors, I think the algorithm is slightly up here to date and is up since we launched the algorithm. So and compare, if we did when we did the peer to peer analysis, we’re like to second ones in the peer to peer analysis. So you do have some fact some possible construction improvement to when using this machine learning methods. And that gives you also a framework to think about your factors. So one example of that could be involved if you’re trying to discover new factors, to look at a wide range of indicators. And you could want the same type of algorithm,
Corey Hoffstein 43:12
I want to dive into some specific machine learning research I know you’ve been doing lately, I was able to find a paper that was published by some of your peers, and you were thanked in that paper. So I know you had at least had a little bit of hand in it, at least in reviewing it. But it was the paper was on the application of recurrent neural networks and trend following which I thought was really interesting. And I know this is something you’ve been researching internally and trend following is something near and dear to my heart. So I definitely wanted to bring it up. What in your opinion, do recurrent neural networks bring to the table when it comes to trend following that isn’t already achieved with existing trend models? Why look to machine learning.
Sandrine Ungari 43:52
And this work? Well, I was trying to find applications for neural network. Because I guess as everyone in finance, I’m a bit. I’m observing what’s happening in the other areas with neural networks. And you can see that it’s a major disruptive technique in a lots of Fourier. So that’s coming in to finance but we just need to find a way to make it happen. So I started with trend following because we had to start somewhere. And the idea was to say, Okay, we have all those traditional ways of estimating trends. So in our in house model, we’re using a form of measure the trends of our civil time windows and then several past 10 windows and then we mix them based on some risk matrix and then we read them and then finally get to get a signal. I guess the way that we do it is very standard. Our this was industry, sort of measuring and constituents and constructing agenda, city type of model. It was like It’s basically based on common sense There’s a bit of math behind to justify the parameters, but it’s more or less common sense. So okay, so now we have the common sense Bayes model that performs as we know, and we know that transferring is being challenged. And the idea was to say, Okay, now let’s try to use a modern technique to detect trends. Now, for trend Trend detection, you have a big problem, which is the labeling of your data. So when you want to learn an algorithm, you need to label your data. So in the picture, it’s very easy, because you know, a cat is a cat and a duck is a duck, you can get to human operator to label the data for you. Now, what is the trend in financial markets? Is it plus 10%? Of affirmance? Or no is like plus 20% over a year? Is it like going up and down in a smooth manner. So visually, your brain is geared towards you, cognate, you’re recognizing patterns. And as a human being, we think we can recognize trends. But when you ask people to identify when there is a trend, it’s very challenging. So we, what we’ve done in this work is that we said, Okay, we’re going to solve the problem of labeling by simulating data. And this is also the new tendency that you see happening in the literature. And it’s a very much a topic in finance at the moment, how do you simulate realistic data? So we said, Okay, we don’t know how to label trends in financial markets, but we know when we simulate a time series, if there is a trend in that time series or not, because you can simulate a changing process. So we simulated 1000s, and 1000s of processes, somewhere upward trending some downward trending some ways, no trends and trends was like breaking at random times. And so effectively, we had simulated 1000s and 1000s of possible assets. And then we say, Okay, we’re good. Now, we had this big data set that is properly labeled, because we knew how we’ve simulated and we’re gonna feed that into a neural network. And we tried several architectures to speed debate about what is the best architecture to be using in what in some contexts. So some argue with CNN convolution neural networks, among other arguments RNN recurrent neural network. So here we use, we converging to recruit RNNs recurrent neural networks. And so we had this neural network that learns how to recognize a trend based on simulated data. And the way that we were able to train it is because when the algorithm was wrong, we told him no, you’re wrong. We know the label that trade that data is trendy, so you should be recognizing a trend. So we showed in this paper that you mentioned my colleagues showed that they had a fantastic hit ratio using recurrent neural network compared to more traditional way of detecting trends on simulated data. And have you see where the when we test that given we were using a new set of simulated data, right, so like out of sample type of simulated data. So now we have this object, a very complex mathematical object that is called a recurrent neural network, which has learned to detect trends and simulated data, properly labeled data. And we said, Okay, now we’re going to put into that algorithm, some financial market data. And then guess what, I was quite sceptical in the way because that neural network had never seen anything close to financial data before. And actually when we fed, so we fed like the algorithm with all the possible futures time series that we could find, like 60 Time series, or something like that for commodities, futures rates, futures, fixed currencies, etc. And then we run the transferring algo. And then at the end, we had something that was performing almost as well as our initial model,
Corey Hoffstein 49:17
a lot of work for almost as well.
Sandrine Ungari 49:21
It’s smarter for thought and experience. But it’s quite fascinating because that algorithm didn’t know anything about financial data. And you knew about simulated data as we had built, we have built the chance for you to transfer with no chances of overfitting because your chances of overfitting and guess what, over the recent past, that guy wasn’t performing either. Which is another way to say that for trend falling in particular due to lack of performance of other risks. And she has is not so much due to overcrowding or things like that is much more due to price patterns. And arguably central banks. And you can find all sorts of explanations, but more than overcrowding, because even a guy that doesn’t know anything about financial data, or risk of overfitting fitting doesn’t manage to recognize trends. So I
Corey Hoffstein 50:21
want to stick with that concept of crowding for a second, because it always it’s always struck me that as systematic managers, the more people that we convert to our style of investing, intrinsically, the more crowded our style of investing becomes, and therefore challenges our ability to potentially generate returns going forward, you almost want to convince everyone that it’s real so that you can gather AUM, but do so in a way that doesn’t affect your outfits, this ultimate scale versus alpha problem. How do you think about these sorts of risks of factor crowding going forward?
Sandrine Ungari 50:57
Factor crowding is true in some cases. So for example, back in spring 2016, you had a massive rush from investors into level quality that led into the big rotation to seek clarification of the summer 2016, from quality level into value. But you could see that in the prices, you could see the factor prices, you could see this acceleration in price, that was a manifestation of crowding, one of the indicators that I’m following, and when I’m looking at factors is the relative performance of a factor versus its peer. And whenever you see a factor, return accelerating, versus it’s pure on a cross sectional basis, this is most of the cases, the sign of overcrowding, oh, sign of rash of investors into this factor, like a bubble basic. And that happens that happens regularly. If it’s matter of fact, I don’t believe so. In a way, I believe it’s movement in euros, because at some point, lots of investors are gonna get interested into growth and go into golf. Well, at the moment, it’s a big rush into growth and doesn’t doesn’t. And there’s a big rush into into some US stocks. Doesn’t seem to end but at some point that things reverse normally. So just to say that factors in general are overcrowded, I think it’s an exaggeration, at least, so you need to look at factor by factor. For example, trends. I don’t think that at the moment is no cloudiness in chance. There was a problem in December 2017, February 2018, after that massive rally in stocks in the US. So stay tuned for our massively long and overleveraged equities, stocks. And that lead into that helps the seller that didn’t help them that made the sale of worse in February 2018. But at the moment, if you look at the positioning of transfers, they’re rather neutral. So the market can go either way that will not have an impact. So it really depends on the leverage of the strategy, the amount of AUM that is in the strategy, and how many investors have been investing into the strategy.
Corey Hoffstein 53:30
All of that factor. Crowding discussions are cross sectional. But when you talk about something like trend following, there is that systematic beta component and there have been some researchers that have argued that certain systematic trading strategies, such as levered ETFs, CTAs vol, targeting strategies, and even just rebalancing itself have started to create predictable and overwhelming flows in the market that potentially are actually a systemic risk to market operations. What are your thoughts on this topic?
Sandrine Ungari 54:05
Okay, it’s really depends on the size of those players versus the global market. So you should look at trend following back in 2018, January 2018, there was a massive of our leverage of equity transfers there so that there was a risk there. She look at risk parity funds before, roughly the same time before the COVID crisis, same thing, the leverage and risk parity was very high. So all those strategies that are levering up some assets are creating some distortion. But knights needs to be monitored because it’s not always the case. Now in terms of leverage of risk parity, for example, we came back down to two levels that are two to three times lower to where we were before the COVID. So it’s scary, less of a risk. But it’s true that in certain market environments and Because lots of funds, lots of strategies, lots of retail product or using some form of volatility targeting mechanism, that as a form of impact on markets and one of the most obvious impact is probably no compression of volatility that we’ve seen. Because when you you’re doing some volatility, targeting of volatility selling in delta hedging, that that tends to suppress volatility. So that’s something that every investor should wait already. There are ways to benefit from that. So one thing that we’ve been studying is intraday trading patterns. And the market microstructure and an equity market. One thing that you notice is that in some market configuration, you tend to have trends at the end of the day. So you tend to have to start to see transforming, and that tends to happen, whenever the overall on the markets to the market participants are what we call in the conjugal negative gamma. Which means that imagine you’re an option hedger, sure negative mag gamma, as your stock is going higher, your delta is going lower, and you need to buy a little bit more stocks. So as your stock is going higher, you will have to buy more stocks. So if you’re negative gamma, you the market is nedic indicative gamma, the market participants hedging this having this gamma position will be forced into buying more stocks in an upward market and we’d be forced to sell more stocks in a downward market. And generally it used to smack it that sevens there, maybe option option hedges and banks, that we need to report zero, at the end of the day, there might be a levered ETFs, who would need to value their funds at the end of the day. So they’re forced into buying or selling more or less at the end of the day does that create that tends to create those two sorts of introductions such as so now, I’m sure some investors that try to monetize that what you can do is to construct trend following position during the day. So say the market starts to go down. So you progressively sell the market following the trends. And at the end of the day, you net short the market and you provide liquidity to those who wants to sell even more. And then you act as liquidity provider the close of the day, and you’re getting paid for that. It’s a very subtle other source of premium that you can do. And that’s the way that’s the way to benefit from all those things. If you want to try to position yourself as a liquidity provider too slow to force to, to execute because of that strategy. There’s ways of monetizing or you know, making, positioning your strategies if you don’t profitably, even though you have to the distortions.
Corey Hoffstein 58:01
Same Sandrine, it has been an absolute pleasure chatting with you. The last question I want to ask I know 2020 has been a weird and difficult year for a lot of people. So I’m trying to make sure we end these podcasts on a positive note. So I would love to know, going forward, what are you really excited about? And they can be research ideas or something personal, just what is something you’re really looking forward to?
Sandrine Ungari 58:23
At the moment? I really look I’m really looking forward to my holidays in Croatia, I’m gonna be sailing for three weeks.
Corey Hoffstein 58:30
That sounds wonderful.
Sandrine Ungari 58:33
proficiently speaking, I guess in terms of research, I’m really exciting by a project that I’m considering at the moment, which is trying to compare convex optimization with neural network again, but this time is convolutional neural network. And to see which one of those guys will give me the best outcome in terms of allocation. So that’s my, that’s my research project. And I guess in terms of investment strategies, we’ve had quite a lot of success with hedging strategies. And that’s something that I’m still actively working on, because I think that we’re, we’re still in in difficult and very environments. So I think there’s lots of work to be done again. That’s exciting.
Corey Hoffstein 59:19
Well, Sandra, and I can’t thank you enough for joining. I look forward to reading that research in the future. I’m sure it’ll be fascinating.
Sandrine Ungari 59:26
Thank you, Carrie. Thanks.