Jay Rajamony – Beyond Factors: Reimagining Quant Equity for the Modern Era (S7E23)

All right, Jay, are you ready?

>> Yes.

>> All right, three, two, one, let's jam.

[Music] >> 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.

[Music] Cory Hoffstein is the co-founder and chief investment officer of Newfound Research.

Due to industry regulations, he will not discuss any of Newf Found Research's 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 forformational 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.

[Music] In this episode, I speak with Jay Rajamani, director of alternatives at Man Numeric.

Jay has been with the firm since 2004, giving him a front row seat to the evolution of quant equity.

From simple factor models and broad signals to today's world of alternative data, model ensembles, and the human machine collaboration.

We start with the history.

What's changed in quant over the last two decades? why the 2007 quake still matters and how the definition of alpha has shifted alongside new tools and data.

From there, we explore the interplay between factors and macro regimes, how sparse data sets are reshaping the research process, and what it means to manage risk in a world where your models don't always line up with reality.

Jay also offers a compelling perspective on how modern quant investing isn't just about signal breadth anymore.

It's about firm breadth, organizational design, and knowing when to lean in and override the machine.

Please enjoy my conversation with Jay Raja.

[Music] Jay, welcome to the show.

Excited to have you here. It's been a while since I've done an episode purely on quant equity, and I feel like you're a great guest to bring us up to speed on how this space has evolved throughout the years.

So, I really appreciate you taking the time. Thank you for joining me.

Thank you so much for having me.

Looking forward to chatting about the evolution of the space and so on.

>> Now Jay, you've been with Man Numeric going back a little over 20 years now, since 2004, which really gives you a front row seat as a practitioner to several of the major eras of quant equity.

Can you walk us through how the quant equity landscape has changed over say the last two decades from back in the early 2000s with simpler factor portfolios to a lot of the modern complexities that we see and face in portfolios today?

>> Absolutely. In fact, if you go back to around the time I started in the industry or even just prior to that in the '90s and so on, it was a very different industry back then compared to today.

It might seem like simpler times.

One of the most difficult things back in those days was simply gathering data in a timely manner, getting that into a systematic process and so on.

In retrospect, if any idea today for example tests exceptionally well back in those days, that's simply because the way we look at these data, the way we understand these data, that would not have been possible in real time back in those days.

As we progress forward from the '90s into the 2000s, the quant world was recovering from maybe the first schooling it had received in value investing, which was of course the technology, the TMT bubble.

And what followed after the tech bubble was probably the first golden age of quans.

That was a time of extraordinary good performance for many quans, a period of wider adoption of quant equity investing.

And it was also a time when quant investing went into new universes and newer markets. Similar trades or similar ideas may have been used by multiple quants at that time.

The capacity of these ideas may not have it well understood but it was a time of boom and there was also a time when extension strategies 1330 and so on were taking wing and some of the risks of value in nursing were clearly known by that time because the tech bubble had just happened but perhaps alpha risk contain risk management and so on were not fully understood or may not have been given enough thought and as we progress from there into the 2007 quant crowding episode and then further into the global financial crisis those two periods were instructive in multiple ways.

First of course was that it was a very good education in terms of crowding for the whole crowd and the second is that it was a renewed instruction on value and there was a distinction between how value performed back in the TMP bubble and of course during the global financial crisis. This was then followed by a period of very increased interest into alternative data and tail risk management and the intentional search of ideas that were orthogonal to more common ideas such as value and momentum to build that strategic depth into which one could recede if these commonly known ideas once again prove their worth during difficult times.

So that was the lay of the land by the time 2008 had happened. And the betrayal of 2008 was especially harsh and I call that a betrayal only because when the market failed or when the market fell is also when quant alpha fell and that was very different from of course what had happened in TMT bubble. This was an especially difficult time period for the whole industry to have worked through and as we continue to move forward from there towards the present time we find ourselves of course today in this world of elevated levels of competition increased supply and demand of highly specialized talent.

So I believe for example the MIT masters in finance program was started perhaps in 2008 or 2009.

So it was only post those episodes that we saw some highly specialized talent come forth. We live now of course in a world of enormous data and increasing use of machine learning and AI and so on. So if you look again once again at the last two to two and a half decades these three periods stand out the TMT bubble the global financial crisis and at the value winter of 2018 through 2020 with that whiplash around the co vaccine time thrown in into that different quan firms have changed in different ways through each of these episodes and I'd say that the competent ones are the ones that did not waste any one of those learning opportunities and as a result of those learnings maybe we have arrived D at a time of high complexity but you can also think of that as an improved reason to believe in the robustness going forward.

One point I do want to make when I look back at this time period is that as a whole we have seen the quan space increase its appetite for deploying new sources of data, new sources of return.

It's become less reliant on what we can call common factors.

It's become more tail risk conscious and so on. But each of these moves to the extent that they're commonly practiced may well be walking us into a new era of crowding if we don't watch out. So that is something that we have to be aware of.

What hasn't changed of course is that at any point in time alpha was hard. So in retrospect it may look like some of these earlier ideas may have been easy or maybe we were in a time period when computer wasn't hard or we were in a time period when ideas needed to be broad.

But at that time these were difficult things to practice and that hasn't changed.

Although today's practitioners, you can cut them some slack if they claim producing alpha today is harder than back it was in the old days.

>> I want to talk about the last two points you brought up there. One about this idea of crowding and second about the way in which compute and data has changed over time. But let's start with the former.

That early 2000s period seemed to be a time when the risk of crowding was low and a lot of people ended up unintentionally walking into the same trades. At least maybe perhaps the perceived risk of crowding was low and a lot of people walked into the same trades and that resulted in that quant quake of 2007 which I don't think unless you're a quant you're necessarily aware of the history of but that was a pretty devastating event.

Why do you think the industry didn't really internalize those lessons of the Quanquake of 2007 either pre207 or even it seemed like into 2008 it didn't get fully digested >> some of this of course goes simply back to how we as human beings behave and what incentives we need to actually correct our course but I believe around the mid 2000s ahead of 2007 clearly there was some awareness of potential crowding people were aware of increasing assets that's in similar trades may may not have been very widespread concerns.

But thinking back to those days, clearly there were voices articulating some of these concerns of lowered alpha opportunities which then necessitates higher leverage to produce the returns that people seek. And practitioners had noticed for example how intraday price moves would happen reacting to financial statement data that were just breaking.

So that is clearly an indicator of how other people are also trading very similar ideas at that time.

These concerns existed but beyond saying that well we ought to be doing something different there was not an easy way out so it wasn't very clear what else needed to be practiced but perhaps in good times even if concerns exist they were not have so much of an incentive to do something different because the money kept printing itself at that time but I do want to point out that there were practitioners who were raising the alarm back in the day solutions weren't obvious but clearly this planted the seeds of alternative data in the collective minds people were thinking of what to do next it wasn't that easy to find.

Of course, there was also an increasing awareness that as quans, we need to keep an eye on what our fellow Quans are practicing. That is, you may think you have an idea that has a large capacity, but if everybody else is practicing the same idea, well, then maybe a lot of that capacity is used up, even if not by you, but by others.

These kinds of ideas were percolating in people's minds.

But of course, nothing drives home a point as much as a period of actual difficulty. So, the industry collectively learned this lesson only post 2008.

So even the 07 episode may have been dismissed as oh that was just a week or that was just a few days.

The episode of 2008 is what drove home some of these points clearly and caused actual behavior to change on the ground.

>> At the end of your first answer you mentioned this idea started to allude to it at least that in the early 2000s it wasn't necessarily compute that was hard as it relates to generating this alpha.

There was some limited data.

The ideas needed to be broad to be impactful.

That seems very different than the era we're in today.

The tool set that's required to be a skilled equity quant might be very different than the tools that were applied back then. And I think it's easy to look back with rosecolored glasses and say alpha was easier. It had its own limits to arbitrage back then that people have just picked off. But can you talk a little bit about maybe how the skill set has changed as we look from what it took to be a quant in the early 2000s versus what it takes to be an equity quant today? As you correctly point to, it's only when we judge by today's standards that some of those earlier times seem easy.

There were clearly a different set of difficulties back in those days. For example, data quality.

How do you compare pre and post euro adoption in those days because that was within recent history at that time.

There were wider differences in accounting standards for example or even timely delivery of financial statement data once they became published.

Well, that was also not that straightforward and the quality of data that was being obtained was questionable. One had to make sure it was clean before you started using it and so on.

The question you had is about what is different in the skill set or toolkit of a modern quantum compared to what we had in those days.

And the first thing to say is what is common and then we can talk about what is different. And what is common of course is nothing has changed in terms of the industry's thirst for talent that is creative, that is driven, curious, innovative.

What has also not changed is the need for the pace of discovery being quick and the need for speed in terms of bringing those ideas to life in the portfolio.

And there is also of course all the usual needs and concerns around risk and execution. All of those things remain very much the same. But what has changed is that the voraciousness of consumption of data. How quickly can you get your hands on the exploding amount of data?

How quickly can you build and monetize the signals that you build from these data?

that has changed very much in terms of scale in terms of just the pace of that effort and the feverishness surrounding all of that. There is an increased need for tech muscle orders of magnitude higher today than what it was back in the day. A modern quant team is actually a very good mix of investment and technology.

It's not just a bunch of quants anymore.

The toolkit therefore is about building an organizational ability to bring tech and investments together.

And I think that is critical because quants and technologists don't always think alike.

And the culture must exist to bring the best out of the collaboration between the two.

We have found for example that in our own experience such a culture leads to more innovation and it helps us become faster to the market in terms of how the toolkit is different. I would say it's not just about technical differences in the toolkit but some of these are organizational as well. So for example, simply setting up a firm where the engineers are allowed to see the entirety of the intellectual property that simply brings everybody along on the same journey. That door between tech and quant that needs to be kept open for easy back and forth of ideas and even of people.

Everybody's interest change over time.

And if you have that opportunity to be on either side of that line, which is a highly blurred line, that just brings the best out of everybody.

And it simply puts a spring in your step as a quant firm to be able to bring these two critical things together. We have talked more than once about data, but there almost a paddle track just like there is new data, there is a paddle track of new techniques.

I would even go to the exit of saying that many of the new techniques we've developed simply would not have been possible without this close collaboration between investment and tech.

So that I think is a big way in which the modern quantum differs from how things used to be back in the day.

I want to talk about the traditional quant factors and here I think you could probably talk about value, momentum and quality being the most agreed upon three and then maybe to a lesser extent some people actually include growth some don't.

How relevant do you think that these ideas and concepts are still today in this modern environment?

And how much of the struggles of these traditional factors do you attribute to a permanent structural decay versus them being cyclally out of favor? It's a good point that built and traded the old way that is just you define a price to earnings or a six-month momentum and you build it in the same way that you could have built it back then or you trade it the same way that you could have traded it back then.

Well, in that definition, these factors clearly have become more risky and they've become less return friendly.

Part of this could be crowding and part of this like you say is the large cyclical up and down swings that you've seen.

Even the oldest definition practiced the oldest way cannot be written off.

It's just that the volatility level has become so elevated that they don't seem to find a place in the modern portfolio.

The concepts however are relevant. It's hard to argue that whether it's value or momentum or quality or growth, none of these things have become any less relevant economically today than they used to be.

So there is an element of structural changes here that have led to these classical definitions falling out of favor.

This could simply be because there is now advanced leakage of the data via other sources. So what a company's earnings may be tomorrow.

Maybe today we have more than one data source that might help us take a sneak peek at that earnings unlike what used to be the case. That may be better ways of trading them, better ways of implementing and there could be some changing relevance of specific metrics.

For example, intangibles clearly just the way these companies have evolved, they've become more important in terms of understanding the value of a company.

increased crowding clearly has made them more volatile.

There is different ways of addressing it. Of course, you can change the way in which they are measured.

For example, if you look further up in an income statement, you see items like sales and so on, which may be better anchors of value in growthy times.

Items further down might be better anchors of value in value times and so on. So that you could have different definitions to help you through this period of changing relevance of some of these metrics.

Not to mention things like intangibles and so on.

Another approach is simply try to be first in the data whether through the use of alternative data or simply capturing the traditional data as soon as it breaks. That is also a way in which you can continue to use some of these older definitions but perhaps in a modern application way. Finally, I want to highlight this concept which certainly is not unique to us but something that has taken hold across the industry which is factor timing as opposed to factor squatting.

So when we're talking about value and momentum and quality at least in the early days was thought of as well you simply occupy these spaces and the returns will follow and you just have to hold tight through the ups and downs. I think that has given way to thinking more in terms of timing these factors and there's very good reasons for doing so and very good amount of caution that needs to be practiced.

Beyond that I'd say the decreased relevance of the old school factors may have come from two other reasons.

one is closely tied to the factor squatting versus factor timing that I'm talking about which is there is an increased understanding of how these factors are simply reflective of the macroeconomic winds that bow across the equity landscape.

So even if they're profitable in the long term that is less willingness to tolerate these long periods of poor performance from these factors.

That's one and the second is there is an increased appreciation of bioaccumulation.

And what I mean by that is if you are an asset allocator and you farm out money to multiple managers and if they all give you 50% common factor counted and 50% let's say unique alpha pretty soon by the time you reach the fifth or seventh manager a vast majority of the allocators book becomes concentrated in these common factor as the source of risk. So that also simply makes them less relevant irrespective of how they performed. I want to touch on that idea and I'm going to use a quote you said in our pre-call which is that factors are the way macro projects itself onto stocks. So this idea that there's a macro factor connection or mapping.

So hoping you could unpack that idea a little bit further for us.

What is the mechanism at least in your view by which macro regimes show up in factor behavior?

I'd say that's probably a couple ways to think about it.

First is that macro consoles almost by definition to the extent that they're macro that is they must affect a whole set of stocks whether it's an entire style category an entire industry or maybe even a set of industries or a few countries and so on.

So by definition macro concerns affect a basket of stocks. So from that perspective simply given the breadth of the impact and the definition of the basket you can say that is a clear way in which macro concerns are intimately tied to style factors. Another way is to think about the classical discounted cash flow setup. That is a numerator in terms of what the company earns and that is a denominator in terms of how it's discounted.

And macro concerns can affect either of these in very clear ways.

Either the earnings or expense of a very broad set of stocks whether it's an industry or a group might be affected by macro concerns such as taxes or oil prices or tariffs and so on.

or the growth prospects of these industries may be affected or the risk of these industries can also be affected.

We've all seen these intuitive explanations of how changes in interest rates may affect simply the payoff of value. I mean it's almost a mechanical algebraic explanation.

I know that is empirically disputed but those are the kinds of intuitions that can lead to this connection between macroeconomic concerns and directly two factors whether it's value or low volatility which you can think of as some sort of a bond substitute or what have you or is leverage that's also a very intuitive connection that we can understand or the impact of interest rate regimes on bankruptcy risk which also we can think of that as a factor so those are the kinds of connections that I see between macro and factors and stocks. So I do think of that characterization as macro concerns exist and then the prism of these factors projects it directly onto a whole bunch of stocks. The most direct way or the most immediately experienced way is of course through momentum.

I think almost every practicing quant has an understanding of how quickly momentum tends to pick up on these concerns that span across industries across any kind of buckets and almost picks up an organically evolving definition of macroeconomics as it goes along.

that exists too.

Of course, >> if we take what you've argued as given and true, does it imply that we should be intentionally and proactively adjusting factor exposures as a function of a shift of macroeconomic condition?

Should equity quants be focused more on the macro or does that just simply introduce more model risk into the picture?

>> The answer is clearly both.

Of course, it is an opportunity. So to the extent that it is an opportunity to exploit, quant should pay attention. But it is a risky proposition.

Even merely sitting on a factor is risky. And now you're talking about timing it. That just adds an additional element of risk.

Once again, it is a classic quan game of how do you balance that opportunity and that risk.

Controlling common factor risk is one approach to preventing macro risk from seeping into your portfolio.

So forget about exploiting it.

This at least gives you some handle on how you can take charge of that macro risk in the portfolio.

But as you say if you could adjust these factors with over time that is a second way in which you can bring in some of these opportunities and now I believe there's two ways of doing it.

One is to do so with an understanding of macro that is you change your stripes because you think a certain macro environment is currently underway or you think something else is going to happen tomorrow in a macro sense and the other is simply to say look I don't know anything about macro but I'm still going to take advantage of whatever patterns I've seen in factors and so on. Clearly between the two the first seems like the most evolved approach that is take some stance on factors based on an understanding of macro and that of course also addresses one can almost call it a blind spot that many quans must have had or at least quans had collectively in history which is portfolios are often built to be all weather but of course repeatedly they proved themselves not to be all weather because when a large macroeconomic change happens or an event happens the quan strategies collectively were found to have dropped the ball. Now that can be addressed of course if you try to make the connection between macro and factors and learn to evolve with the macro situation.

Factor timing brings its own kinds of risk. Of course you have to know what the breadth is.

I think even anybody who invests for example in a universe of three five or 10,000 stocks knows that the full breadth of that universe is certainly not that large. It may be limited.

Likewise even with factors however many factors you have at your disposal it is a limited number of bets that you have.

That just means that is an element of understanding of risk that is needed.

And partly in terms of understanding the tail risk that can be brought about by factors.

At least to me it seems like a good way to do this might be to think about what is the biggest hit you can tolerate in your portfolio and then work backwards from that into what is the largest factor exposure you're willing to tolerate in that portfolio so that you may size these bets appropriately.

But if you are willing to take these kinds of approaches to sizing a bet, this is clearly an opportunity and that connection between factors and macroeconomics actually is a doorway that we should walk through.

>> So you talked through the history of quant equity sort of the 2010s one of the major emphases was alternative data and so I want to talk a little bit about that era now and how alternative data changed the way quants had to think.

The first question I want to ask is sort of a compare and contrast. Right?

In the old quant world here, I'm using air quotes for all the listeners, early 2000s, there's this idea of wanting long histories and global breadth and all this data which we could apply these ideas cross-sectionally.

In the modern alternative data world, you often can get a data set that might only have a year or two or three of data, and it might be incredibly niche.

Maybe it's just a handful of Japanese consumer stocks, for example.

How does your framework for evaluating signal viability have to change with that data sparity?

>> You're absolutely right that while we did speak of some of the earlier periods, simpler times or some such, what was available in terms of the luxury of data was the history.

That is the data may not have been clean or easily timely available but long histories were certainly available.

Today that is very different.

Many data sets as you say come with short histories.

The way we think about it is first you should have a hypothesis in mind when you go after these data.

That's of course easier said than done.

You cannot always do that. But to the extent possible, being sure of the hypothesis and questioning whether that hypothesis makes sense and whether you can connect the dots of economic causality between the data and the returns that you hope it'll eventually produce.

Checking that is step one in the days of limited history.

Connecting those dots of economic causality also gives you an additional way to test.

So what you make up for in terms of history is by trying to see for example if it was Japanese consumer data as you said you could see if it predicts some of the intermediate fundamentals whether it's the sales or earnings or what have you before you jump forward into testing whether it predicts the returns that I think produces now an additional thing to test for an extra layer of comfort before you start using the data even if it's a very limited amount of history now of course that can be entirely different kinds of data it needn't be consumer it could be flow data uh something about one's peers or fellow investors intermediate markers may not be available easy in this case for us to test but even here simply having that first hypothesis is very important and I think that is a way one can get around the lack of history no matter what the data promises so a second way to approach a problem beyond having this hypothesis and testing for other elements is that no matter what the data promises you always make sure you limit the allocation so look at the end of the day you're walking into a situation with limited history, there can only be so much faith you can put behind the return promise that it shows. In other words, you make up for the lack of history by simply looking for vastly more amounts of data that is different from each other, but then you spread your bets among all of those. Given that these data sets can be so niche and they can be short periods of time, you end up with this almost exponential explosion of data alternative data sets that have come to market.

Given this proliferation, how do you think about designing a scalable research process to ingest and explore them?

>> There's thousands of data sets today, if not tens of thousands of data sets.

So the question is do you just go for all of them with a scalable process like you mentioned and consume as much of them as possible as quickly as you can or do you try to pick and choose and if so what are the metrics based on which you might pick some of them we see the benefit in terms of if we can put in a process of scouting and filtering before we just consume all of them that is something that we have found beneficial so for now we start the process by asking a few questions about the data for example is it very similar to what we already have.

Are we looking the same kinds of data?

For example, going back to your consumer thing that you cited earlier, is once again a mining of the same Japanese consumers.

That is one way to begin the exploration.

The second is to say do these data exist for what question that we may be trying to answer for whatever gaps that we may see in our process.

What data exists in that spectrum?

What can we go after? What is a conviction level?

Does it provide a concept level diversification?

Does it provide a regional diversification?

And finally the flip side of the regional diversification is there are some data sources that are very specific to particular regions could be a country could be some sector in a country that is not available elsewhere.

Those kinds of data sets are also useful.

So going through what are some of the themes and then scouting along the dimensions of those themes is something that we have found useful.

And in fact I would say that one of the benefits of doing it this way is that not every data set is advertised on platforms. There are of course large data platforms that serve the needs of the quant community but not every useful data set is available on these platforms. Now it's an entirely different question as to why the data provider does not put themselves on these platforms but there is some expertise that also comes in handy in discovering these data that are not easy to find.

I'd say our primary goal is not to build a process that is as scalable as possible and consume every data set out there.

Of course, we do have that focus on being more efficient in terms of monetizing what we find, getting it mapped, getting it verified, exploiting it and so on. But there's an element of selection.

This is not to say, of course, that those who go after the scalable consumption of data, that may be an approach, too. But I believe we found some value in going after the scouting and filtering first.

>> Talking about alternative data was highly in vogue in the mid to late 2010s.

it feels like it fell out of favor and at least in my private discussions I feel like a lot of quants saw a tremendous amount of promise in alternative data sets but ended up hitting frictions and struggling and roadblocks and I'm curious in your experience what have been some of the biggest problems or issues you've seen in utilizing integrating or evaluating alternative data sets >> so we can talk about it as two ways we can talk about simply what are some technical issues in terms of using alternative data but we can also talk about philosophy ical issues as to why people don't talk about it as much as they used to as you alluded to.

So the technical issues that immediately come to mind are how does a vendor define point in time data that is simply not the same across all data vendors.

This also means that many data vendors may be restating their data immediately bringing to question the ukity of any back test that you may be able to run.

Then of course in the live process there may be a delay in delivery that was very different from what you understood in the research process. That can also be a problem.

Product upgrades on the part of the data vendor or the changing of the data structure.

All of that can lead to a lot of extra work for technologist to maintain and continue to deliver that data into the production process.

These are some of the more common issues that span multiple data providers that come to mind when I think of issues with alternative data.

The philosophical points with alternative data that I like to make are to the extent that I've taken a look at these data.

One thing that I find very interesting is earnings season continues to be just as relevant in terms of price discovery.

One argument that you can make is that suppose alternative data were truly giving you an advanced preview into what might be happening under the hood of a company.

Well, then earnings periods ought to be a lot less exciting or at least somewhat less exciting because you would have known ahead of time.

So if the entire industry were collectively using alternative data in a clever way then one may hope to see some sort of a decreasing importance of either analyst estimates or earnings release itself.

I don't think I've seen anything of that nature.

In fact when you go look for the relevance of earning seasons whether it's through volume or other kinds of metrics that you can think of volatility or the benefit of advanced knowledge of earnings and so on they all continue to have just as much value today as they used to have. Of course, none of these tests are bulletproof. Where I'm going with that point is that alternative data is not some magic bullet. It is in some sense no different from any of the data.

It's available to everybody.

Anybody who's willing to pay the price can receive that data and then it's up to individual practitioners to make sure that they make the best use of it.

So, in some sense, alternative data is no different from simply knowing what is the cash flow of a company last year, for example.

It's just a matter of how you use it. So that is one reason why some of that initial sheen may have worn off and there is no magic here unlike what may have been initially expected.

That's one possibility. The second of course is that widespread use of the same data clearly if you're an alternative data vendor you would like it to be consumed by more people and widespread use of that data can lead to crowding no different from what was seen earlier and in fact I believe you may have seen a couple of signatures that even in the recent past of potential crowding in some areas of the alternative data space so I think these are some of the more philosophical reasons as to why this is not talked of as be all end all of investing.

Can you talk a little bit about the criteria or hurdles with which by an alternative data signal is integrated or even sunset from your process here?

Again, I'm thinking if history is thin, how do you develop confidence and manage the turnover of ideas in the live portfolio?

>> I would go back to what I mentioned earlier about the hypothesis and the prediction of the intermediate fundamentals.

Intermediate meaning standing between the alternative data and the returns. If you were to think of the turnover of these ideas, the first thing to check might be not just if the returns have stopped coming but to see if there is any change in the reason why you think returns have stopped happening which could be understood by looking at the intermediate data for example.

So that is one approach and the more alternative data that you consume the more it makes sense to have a statistical tools by which you can evaluate them in a near continuous manner rather than look at them once every so often. The more data sets that you consume there should be a continuous evaluation in a single sense of the utility so that these decisions can be taken more quickly. there is more data and what that gives you the luxury to weed out the old ones and put new ones in and that needn't always be in terms of what doesn't work or what works because some of that can also be in terms of has a new data set made its appearance on the scene that actually does the job better than what the old one was doing. Keep the same hypothesis but simply sunset one data set in terms of another because maybe the new one is more comprehensive, maybe it is delivered in a more clean way, maybe it is more timely. So for any one of those reasons you might want to move forward at sunsetting that old data set and of course all the other reasons like if there is a reason to believe that the data set has become unstable or is badly mapped and so on and once in a while you might have to sunset a hypothesis itself.

If for example the flow data that you're going after or the sales segment that you are capturing that has become less relevant. If the market fundamentals themselves are changed in a way that causes you to retire the hypothesis, that is then a different reason as to why you might want to do away with the data set. But I would just return to that idea of always having a clear view on why you're using a certain data set.

Not just because it's available, but maybe there was some question that it was answering for you.

Having that clarity of why you're using it, that all makes it that much easier to decide what to keep and what to sunset.

When you have a limited history of data, the correlation matrix or coariance matrix you develop can get really noisy really quickly. I want to talk a little bit here about the portfolio construction side.

And my question to you is how do you make sure that something as important as your coariance matrix remains numerically stable when you're adding all these new signals or data sets that have a really shallow overlap with your existing data.

So I have a very engineering approach to this question.

First I would say look there may be so many different ways of statistically addressing a coariance matrix and shrinking it and so on.

But at a minimum I think we all should bring to this battle the common sense of not believing too low a correlation or negative correlation numbers because in practice we know that this diversification is never to be found when you need it the most. So I think we should just be wary of especially given the limited history like you say we should not fall into the trap of false precision with some of these correlation numbers and again limiting your allocation to any one data source.

That is another way of simply using judgment to make sure that there is not an over reliance on the correlation numbers but you sort of enforce a forced diversification which is I think a very important skill that the portfolio manager even in the quant universe has to bring to the table. There is an element of judgment here that goes beyond simply looking at the numbers and reading the correlation structures and that judgment is critical in the quant process and many times you could simply set aside what was measured and use something that is expected that is typically if you see among these ideas a correlation no less than.5 even if two data sets give you some absurdly low correlation such as 0.1 well don't use that use what you think is more appropriate for these economic ideas and that might be a much better way to think about this.

Also, you think of your risk allocation that ought to be some sort of a measure of expected sharp that you have from the process. At the end of the day, whatever methods you've used, simply look at the risk allocation in your portfolio across various ideas in a correlation adjusted way. And if that seems to map to an absurdly high sharp expectation from any one idea, whether that's a data source or what have you, you ought to correct yourself.

So you have touched on what is a favorite topic of mine which is that a portfolio manager really ought to sit down with what were the risk allocations they have made either explicitly or implicitly and keep on questioning those because this layer of judgment that an investor can bring even in the systematic or quantitative arena is critical because at the end of the day diversification is your only true friend in this endeavor and you have to enforce that. I want to talk about the idea of diversification but perhaps more at an organizational level.

One of the ideas you brought up in our preparatory call for this episode was that firm structure has sort of changed over time in that today you need this idea of a broad firm that supports very deep quant effort versus historically it was all about a wider quant effort.

I was hoping you could explain what you meant by that.

How has the organizational structure that has to support quant equities today changed as the landscape has changed?

>> In the early part of the call, we touched on how we may have evolved from what appears to have been a simpler time to the times of modern complexity.

And you're correct in that as we go from then to now to support this increased level of complexity, you need a firm that is more broad. So in some sense it's almost as if that old paradigm now either is inverted or works both ways and what I mean by the old paradigm is simply that quants need breadth that's what we always say and that used to mean that quants need breadth of stocks a broad universe of ideas but it is also equally true that a modern quantum needs to sit on a very broad platform without which I believe chance of success is highly limited and in fact touching back on some of the alternative data questions that you and I just talked through in some sense you can say that broader the firm the more capable it will be to exploit some of these narrow breath narrow universe ideas that is you can't go after these highly specialized data sets without the breath of a modern firm and partly look if you think about the multiple areas of risk and data execution and technology that surround the activities in a systematic firm we can just touch on for example data and technology just to see why the breath is needed goes without saying that equally true for risk and execution as well.

But in data, we talked already about the need for hypothesis in understanding data.

But you can imagine that as the number of data sets explodes, there is a very high level of due diligence that is needed.

Multiple calls have to be had with the data vendors. There is a need to look deeply into the data libraries, scouting activities, onboarding, mapping, verifying and so on.

All of this requires a large team and it is in fact once you cross a certain number of data sets it becomes an industrial operation.

You do need a way to constantly bring on new ones even if you're not consuming every data set out there.

That is a need for processing these things at scale and simply maintaining these to keep them going in the live process also requires a lot of effort.

Now I'm moving from the arena of data into technology. So now I'm talking about things like hardware and storage that is needed. Whether it's simply sorting large data sets or getting a quick answer to multiple queries that are going back and forth. All of this can happen only if you have a very large tech stack.

And that requires scale.

Once again, we haven't spent much time talking about machine learning or AI, which is also, by the way, one of the reasons why people don't talk about all data as much anymore because they need time to talk about machine learning and AI.

That requires a large amount of GPUs.

It requires libraries and resources so that researchers who may not be experts in machine learning themselves are able to use machine learning as a tool which you simply cannot build all of these out without a large scale.

Talking about execution for example there is a need for development for maintenance and use of trading algorithms. You need not just good technologies but you also need large amount of good technologies to make all of this happen and it is simply not possible by throwing a lot of researchers at this issue. So you need low latency, high throughput systems which simply cannot happen if you were a threadbear quant operation just run by a few people and to get that tech oomph and so on you do need that very broad scale of the organization for a modern quant firm.

I want to talk a little bit about modern risk management. And one of the first questions that comes to mind here after all this discussion about acquiring alternative data sets is actually the operational risk management of having all of these third-party data vendors.

The risk becomes a data vendor stops providing a particular data set that you're reliant on or a data vendor goes out of business or they start charging an excessive amount of money.

How do you balance this need for using a third party data vendor managing the risk of that vendor themselves versus trying to in-house the data acquisition and the cost of doing that? How do you balance that risk?

>> The way we balance it at least is by taking advantage of the number of data providers.

So the greater the number of data providers and the more you have spread your bets among them, the risk of any one of them affecting your process is limited.

For example, you should think of all the things that you mentioned, but there is also an element of what are the data are simply wrong and they continue to come in the way that they used to. You may not catch it on day one, maybe day two when you catch it, for example. For all these reasons, I think I keep sounding like a stuck record here in terms of going back to the need for diversification.

But I think when it comes to this almative data and so on or any kind of data for that matter, the broader your footprint, the easier you will find to wade your way through this. And you do have to think not just about anything wrong with the data but an evolving situation for example might make a certain data set less relevant and that also means that you should be able to pull that out perhaps replace it with something else.

One of the big debates that exists among quant practitioners is the idea of discretionary intervention and you mentioned on our pre-call this idea lean forward risk management where managers and portfolio managers will proactively step in when the world starts to diverge from what models expect. Can you talk a little bit about what this intervention looks like in practice at Man Numeric and if you're comfortable with it provide some specific examples of how lean forward risk management has worked in the past.

>> The purpose of these interventions or the lean forward risk management is to make sure that the process is still consistent with the hypothesis we had when the investment system is built.

So for example, if a country were to ban shorting, which is something that has repeatedly happened over the last 10 to 15 years, the immediate question is how do you deal with it?

There's multiple options of course. One is you freeze your shorts so that you then have the ability to keep the longs going or you could say no, I need to vacate the shorts and that might also involve unwinding some of the longs in a market neutral context for example.

or any shorting related data that you might have used as part of your alpha process may may not become less relevant in the situations.

This is an example of a very obvious case of intervention that is you can't imagine that a systematic process can run entirely on autopilot because the world changes and you have to react.

It may seem like a simple example but even beyond that walking forward to more complicated examples would be sometimes the market itself will present with you an opportunity to understand risk that a risk model cannot. So for example think of the US elections in 2024.

Of course all portfolio managers would have been keen to know what kind of risk they carry in the portfolio whether it's prodemocrat or pro-reublican.

But no risk model could have told you that in a clean way.

But if you were to look at how the market behaved on some days that you could take it canonically as either pro-republican or pro-democrat perhaps around debate time period or when one candidate exited and another entered those then provide you with some opportunities of being able to measure risk in the portfolio in a very direct way and then make any adjustments you want.

that adjustments may have primarily been in the direction of limiting risk but at a minimum even if it did not lead to an intervention.

This is the kind of lean forward risk management that I think makes a lot of sense that is not simply relying on backward-looking models or backward-looking risk models but trying to take advantage of every information that is out there. Intervention to me can mean something like what we just talked about in this election situation which is macro related. So for example, even during co even before the vaccine came about and produce a big market move in the November of 2020, one could imagine what the arrival of a vaccine might do and then query the portfolio as to whether you're able to withstand that shock.

That would be another example of a lean forward risk management.

When Russia invaded Ukraine, for example, how did that change the climate related concepts?

What happened to those ideas?

Those kinds of questions can be asked even before for example the invasion unfolded.

So these are some of the ways in which one might think of interventions.

You can have interventions related to data problems like you mentioned or some of these could be based on how the strategy itself performs. So for example if the performance pattern of a certain idea is extraordinarily good or bad or if the volatility is unexpected despite the best efforts if the product is running anti or pro-inflation bets.

These are all reasons why you might want to get in there and do something about the whatever it is that you see and fix it in the portfolio. The point I want to make is that most of these interventions that I've talked about or all of them are really about making sure the process sticks to the hypothesis you had in mind when you first built them.

That's the idea. But I also don't want to make it sound like every intervention is only along risk dimension. There is also a very exciting opportunity dimension to it.

Valuation stretch is something we all understand the distance between cheap and expensive companies.

And of course when that spread widens you can imagine that is an opportunity for you to intervene and step into value.

Another example is meme stocks.

When these kinds of phenomenon happen in the marketplace they do provide an opportunity for whatever limited amount of risk you're willing to devote to it.

Here is an opportunity for you to harvest returns in a systematic way but again not possible if you are not lean forward.

>> How is that structured operationally?

I can imagine you can try to redte team this and proactively explore different ideas and some may have to be reactive.

So how do you think about dedicating resources to I think what we're effectively talking about is emergent risk factors that may become very impactful to the risk of your portfolio and then decay in impact over time.

to the example you gave, Republican versus Democrat might be able to be measured during certain periods in the presidential cycle and then may be less relevant for years after or totally changing characteristic.

So I'm curious again how you think about dedicating resources to exploring, developing and maintaining a library of these emerging factors and when you let them go.

>> This is a good point. There is no correct way to do this. To me it is less a question of dedicating resources to doing these things and more a question of having the right people in the right seats who are naturally inclined to think like this. So you need people who are very awake to the market.

They are market animals. So even though it's a systematic firm, you need folks who are constantly looking at what's happening in the market not because it's part of their job but because they simply love doing so.

And that is when you get the best out of these kinds of possibilities.

That is they notice something is happening in the market that is an echo of that in the portfolio and then the question becomes how are these connected or they can imagine something happening to the market or in the macroeconomic scenario they can imagine what implications might unfold in the portfolio. This will then lead these individuals to come up with these kinds of definitions. So the most exciting part of this is that none of these are being capable of predefined.

So just because you happen to navigate one pandemic and one election cycle through use of a couple of ad hoc measures of risks and use that to gauge your portfolio doesn't mean you can put that in the library and then use it for the next election or the next pandemic.

That will going to be an entirely different ballgame.

Rather than thinking of devoting resources, I think it is more a game of finding the market animal type individuals and having them be on your side of the table as you navigate all these difficulties.

>> When you do ultimately intervene in a portfolio, say it's reducing exposure to some specific industry, how do you balance that judgment against your quant mandate?

At what point does a discretionary overlay ultimately undermine what is supposed to be a systematic process?

>> That is a line here beyond which capriccious interventions will certainly be wrong.

So I will begin by saying that look at the end of the day the mandate is quant like you said you're correct but the primary responsibility of the portf manager is simply to be a good steward of the assets and that is the driving force behind these interventions.

If you have reason to believe that whatever risks are unfolding in the marketplace if it's not well understood or even misunderstood by the current setup that we have in the portfolio or in the models that is the reason to intervene and of course likewise as I said if there is an opportunity that presents itself and you have some systematic way of capturing it in a risk control manner well you should go for it rather than saying this is not part of the process as of now. So a lot of this comes simply from not sitting on your hands and watching performance as it happens.

This is not a spectator sport.

You have to get in there.

And the question is not of making a market or a sector or an industry call. It's more a case of making up for the backward-looking nature of literally everything that a systematic process will do.

You are just understanding that the situation today it will take a certain passage of time before the system can digest what is currently unfolding.

And if it makes sense in the sense of being prudent about risks then well you should intervene because that is what it means to take care of those assets under management. So long as there is that level of clarity in terms of why we are intervening I think that is nothing wrong in intervening from that perspective.

Uh capriccious overlays of course is certainly off the table and that's not the kind of interventions I have in mind.

I'd say there's two things to watch out for.

So like you said you can't take it beyond an extreme.

One is any excessive tendency to intervene can be walked back simply by asking yourself not only why you're intervening but also under what circumstances will you stop intervening.

If it's a factual intervention we try to lay out exactly here are the conditions that need to be met at which point we will undo the intervention.

That discipline needs to be there and if you're unable to write that down maybe you should not intervene. That's a very good discipline to have. And the second is that in some very good way today's interventions will become tomorrow's systematized models because the more you think about why you're doing what you're doing and the more you're able to codify it and say here are the conditions in the market because of which I'm doing such and such well you could simply make that a systematic part of your systematic process tomorrow and in some ways that is really how you keep moving the process going forward.

Let's talk about as a final question that process moving forward because if I were to blanket summarize this entire episode, it's been an evolution where both the alpha and risk management has gone from breadth to depth over time.

We started in the early 2000s measuring both risk and reward across a very broad spectrum.

And then from the 2000s to the 2010s into the 2020s, the story of quant equity has been going increasingly deep, increasingly deep in data sets and increasingly deep in emergent risk factors as to how we measure what's happening in the portfolio. And so my question to you is going forward, do you think that that is a trend that will continue to persist? Where is man numeric putting its efforts on a go forward basis and trying to send the leading edge of quant equity >> the move from breadth to depth may well describe the journey of quant so far but as we stand where we stand today and look to the future I don't see these as one or the other kinds of choices on the one hand data will continue to take us deeper that'll be more specified or highly specialized data sets that tell us more and more about an individual company and perhaps Some of these machine learning kinds of techniques will help us in that ability to deal with those kinds of questions at scale.

But at the same time, there is an explosion of techniques that address the breath.

It's not as if all of the broad ideas in the quan space have already been mined.

Many nonlinear techniques are coming to the four these days.

In fact, you could think back to the factor timing ideas that we talked about.

Many of those are broad full universe yet they produce some new source of return just like there exists a dimension of diversification between for example regions or across time horizons.

You know you have investment ideas that span a week to a month to a year.

There are also maybe a philosophical dimension of diversification that is available between broad ideas and deep ideas and they produce very different kinds of performance patterns for you.

Whereas broad ideas may be more susceptible one can say to macro, deep ideas may not be.

So that is just peeling back one more layer on that potential diversification there.

But at the end of the day it is an arms race on trying to build out more and more sources of alpha as we look forward from where we are. For us it is about two things. One, it is about focusing on making sure that we force that burden of return generation across as many different alpha sources as possible because that is the best way to play offense against any future difficulties in the quantum arena.

And the second is to focus more and more on delivering idiosyncratic content that is content that we at least believe is peculiar and unique to numeric.

Moving forward along these dimensions I believe will bring us the best outcome going forward.

Okay, Jay, we have come to the end of the episode and I want to ask you the same question that I've asked every guest this season, which is, what is something outside of the realm of your work that you are currently obsessed with?

This could be an idea, a hobby, music, a book, a show. What is something that you are currently obsessed with right now?

I am obsessed curry with the origin of words, ethmology really in multiple languages.

I'm just fascinated by how and why some of these sounds that the human body is capable of producing became associated with the meanings that they carry.

There is poetry in this and understanding how some words came about because the words literally describe what the object is or in some cases it is so far-fetched that it is almost comical that these words carry the meaning that they do. But this is an area that I'm fascinated by these days.

>> I love asking this question because every guest gives me such a unique answer and it's such a often unexpected answer.

That's a great one and a fascinating one.

Thank you so much for joining me.

It's really been a pleasure.

>> Thanks so much for having me.