BlackRock’s Jeff Shen on Systematic Edge and AI

break out. I'm I'm pretty surprised how well value has held up, and I'm also surprised how the defensive strategies just have not worked, things like low volatility, things like quality. They're not getting killed, but they're not getting that safe haven bid that you would think you would get when

you have a geopolitical event like this. So I think you could you could kind of interpret it as the market shrugging office war or looking past it, but we have not seen the safe haven flows internally that you would expect.

And I think the other thing that I think maybe in terms of misconception, is also that historically there was certainly a bit of a feeling of systematic strategies could be a bit of a black box. People don't exactly understand their all these labeling, but people don't exactly what's going on. And to the point that you know, a minimum volatility strategy may not be as defensive as one thinks.

I actually think that that when there's an inflation story, there's also an AI story that may, you know, maybe underneath a hood of that particular conundrum. But I think I think the exciting thing on a forward looking basis, given some of a new technology, is that I think you can also understand these strategies in a much more transparent and humanly understandable way. So I think that's probably one of the misconceptions people say, Oh, I understand the stock,

but I'm not social about the factors. I'm certainly not social about quantitative methods. With the aid of AI and large language model, some of these transparencies may also be evolving a forarlooking basis as well.

are as fundamentally different, you know. I think about back in nineteen eighty five when we first started trading to be able to get a price to book, price to earning and give it and yield information for five hundred stocks in S and P five hundred. That was big data machine learning back in nineteen eighty five. But fast forward today, I think on this point of evolution, it's certainly about the amount of data you can access, the amount of learners that you have in your system, and

the type of execution you're going to have. So I also that maybe the core source of alpha is certainly anchored around innovation and an innovation today. I think it's also one of the most exciting times, if not the most exciting times over the last forty fifty years, is that the many innovations you can go for data learners execution. So I'm sure we'll dig into it, but I think it's it certainly has evolved quite a bit over the years.

O can you time the market? So I guess I'll ask you, like, do you think factor exposures could be timed in a somewhat reliable way? And what, if any does factor timing play in your process.

Some of the blow ups there certainly caused people to think about think about this quality metric in a very careful way. And I would say that maybe right after the financial crisis, that's probably there is a bit of a watershed moment in terms of think about factor and factor timing financial crisis, a lot of these factoris actually had severe draw downs, and they all actually draw down together.

So I would say that around that time there was a you know, twenty ten twenty eleven, there was a big debate in the industry that you know, are the old factor is going to come back and you just need to time them a little bit more, or their additional new new things that you should be looking at. We sort of took the second route, where we think that there were a lot of new innovations alternative data, machine learning AI that you can use for systematic investments.

So we went for that sort of a route in a significant fashion. But when we look back at the you know the factor lend if you will. I think these are becoming very well known. Uh you know, drivers of across sectional return. They're pretty persistent drivers of return, even though the time variation to a point is becoming much more pronounced. And I think this attempt to time them is certainly I think it's you want to do it carefully in the sense that you want to identify

the factor in a very careful fashion. Depending the definition, it can vary quite a bit. You want to InCor put it not only bottom up information, but you also want to look at top down information how the regime can change the factory rotation. And then last but not the least is also you've got to think about how to manage the risk of this kind of timing because by definition, it's a low breath type of strategy. It's a low information type of rotation. So you want to

make sure that you manage the risk accordingly. And if you do this carefully, I think it's certain. We've certainly done you know this type of strategy in different form, and we think there's actually some promise, but it's not nearly as easy as people think it is.

like every quarter. You know, I'm not asking for secret sauce here, but like do you have do you incorporate other factors that rebounce maybe more frequently or even less frequently to that kind of more traditional cadence.

liberal arts. You also need economics and finance and understanding of policy and the geopolitics to really make that uh, to deliver consistent set of returns. So I think the nature of the game is actually quite different, and you know, in the high frequency or shorter horizon, you actually don't really need too much of a sensibility per se or economic intuition versus in the intermediate horizon to the you know, whether it's a value or momentum or minimum volatility, there

is actually some economic story associated with it. So I think in that sense, I think the the big price I think in my mind on a forward looking basis is that, if you will, there are certain firms kind of sawt of the little bit of the high frequency trading game, and it's still very competitive and keep on evolving, But there's actually a set of very successful large incumbents.

Now in this kind of medium horizon, I think the space is reasonably open, uh, discretionary, systematic, and you know maybe old fashion systematic, modern fashion, uh, you know, modern version of systematic. So that's a bit of a wide space that I think we'll see a bit more evolution forwardlooking basis.

actually this concept of collaboration. Now it's a little bit overused words to a certain extent. But I think you know what we did was, actually we provide a tremendous amount of transparency for folks who join us.

you really need to step up. So but that level of transparency really fostered the level of collaboration, making sure that there's actually a lot of disciplines involved in solving hard problems, and that there's actually a lot of people trying to help each other. So it's actually one big platform really leverage that. So so that that part, I

think from a cultural perspective, really help help that. I think the second part that I think is also probably equally important, this is I'm going to speak a little bit from the San Francisco Silicon Value perspective is that this idea that the world can be a better place and whatever is being done, uh it's all good, but there will be something better and more and different. Uh. So if you will this kind of outlier type of thinking,

I think it's actually very central to our culture. It's not just to say, oh, everybody's actually done this kind of things before and we're just going to execute a little bit better.

five you know many different ways. You're so kind of building on these questions, so you know, there's these well known factors that I call open secrets and finance that people know the value moments and quality low of all you know, generally like what percent didch what you say of your research is involved in those factors, trying to refine them, trying to get the best combination, trying to even time them, versus trying to find new things that

are bespoke that people might not know about or are not well known to add value.

We're not going to do quality version seventeen. And you know, while there are other people who may want to do that, we're going to go on a completely different route, which is to look for new signals, new ideas that nobody has actually really thought about before. You know, simple things like you know, if you look at the job posting data in the United States, there are you know, millions of these job posting data. You know, that's a you

know on companies website, on aggregator's database. And if you have that data set, you can essentially get a sense of Okay, is this company hiring and what type of skill set are they hiring? Modern skill sets versus old fashioned skill set. We have this signal idea which is jokingly called lung Python short Excel. You know, people who are hiring more Python oriented skill sets are more ready for the machine learning AI world and versus actually quite

like Excel. But we're bet against the company that's hiring heavily into Excel oriented skill set. So you can actually have a new data set and come up with new factors, if you will, new signals and asking a bunch of questions you never thought about asking. And so we've been on that journey for sixteen seventeen years and fast for today it is you know, one thy five hundred signals or factors if you will, far far beyond the traditional,

open secret type of traditional factors. And that really allow us to you know, really tap into the data engineering piece, the different learner piece, the some of the new new things in terms of you know, natural language processing, and fast forward today large language model. So I think once you open it up, you do realize the world is very big and there are many many places to go.

And I think given the development of AI today, I'll see that that number is going to go skyrocket even more on our forelooking basis, just given the technology is getting a lot better.

So I think that data explosion is going to continue. On the other side of it, I think given the more and more data that's available. I think it's also extraordinarily important to think about what are the interesting questions we're asking to this particular data set. To your point, extract more information from a particular data set. And I think that super exciting area because I think the quality of the questions in this kind of a big data

machine learning era is very very important. And this ability to ask interesting questions I think now can also empowered by generative AI, so you can actually, you know, we have an interesting research which is actually essentially co working with large large language model try to extract more interesting questions that we have not thought about asking for a particular data set. So it is actually data to answer your question is actually broader but also deeper, more data sets.

But it's also trying to make sure that we ask deeper questions to each individual data sets. And I think this is we can talk more about it. But I think this concept of scale in alpha generation, I think it's becoming increasingly critical. You need scale in data, you need scale in inside generation, you need scale in compute,

you need scale in implementation. So how do you leverage the scale for innovation for alpha I think it's really going to be a bit of a defining theme in the years to come, you know.

let's focus on top down. So, like, if you can give us maybe just an idea of like the general things that maybe you could use the time factors, like is it the momentum of the factor, the evaluation differences, is some kind of economic regime, anything you could share that would be very valuable.

it's very comforting to see. But ACCENTI, how do you define a regime and how do you link the regime to the factor performance in a rotational sense? So the interesting work that we've done I think I'm happy to share is actually try to define the regime in a

way that that's much more machine learning driven. Is to say, let's look at everything going on in the world, you know, different asset prices, different economic variables, and you can actually use a machine learning algorithm to essentially define what type of environment that is given all the you know, if you will time serious data that you observe in a particular period, and then you can actually you look at sort of you know, we use a sort of fancy

Euclidian distance term, but you can look at how similar this regime is to the history that we've actually seen. Some people say, well, if all your price is going to so this way, it looks like nineteen seventies, or if there's actually a private credit issue, this may look like a Carbinger of you know, two thousand and seven,

two thousand and eight mortgage crisis. So whatever the talk that may be, I think you can actually verify it with a more machine learning algorithm to say, okay, let me actually look at the Euclidan distance between this and to see how is that defining a particular regime in a systematic fashion and then you can link that to the factory performance to say, this is a regime, how

did this factor do? Allow the machine to learn the factory performance in different regimes once you define it, and then you know, once the market price start to move, you can identify the regime and then in relation to

that factory performance from a timing perspective. So that's maybe one specific example that I think is very intuitive, but it's actually what human we as a human well look at market do anyway, but it's actually in this where you can quantify it, you can systematize it, and you can also do it in a bit of unbiased way.

there's a bit of efficiency gain aspect of it. And the second part, executly to your point is also I do think that from a performance perspective is that once you have a big stack in terms of data, in terms of learners, in terms of assystems and secution, it is actually possible to drive the diversification. From alpha perspective. You know, clearly multi strategy hedge fund have been very successful in general in the hedge fund industry a quarter

of the assets and growing very rapidly. I think that's also there's a lesson learned there, which is really if you deploy scale across different parts, across different technologies. We have a little bit different way of looking at it, but it's ultimately it is about driving that scale in the type of strategies horizon insights, and that's where you can drive both the consistency uh, but also the differentiation UH.

And I think that's if you will, the the from an investment outcome perspective, also, the consistency and differentiation are probably two most important, UH, two important metrics you want to you know, all perform consistently from a sharper information racial perspective, but you also want to make sure that when there's a market warble, when there's all your price spike, or when there's certain things happening in the market that's

impacting active managers at large, that you have a differentiated way of looking at the world, that's the strategy is actually performing better than some of the competitors. So I think that's yes. So the short answer to that is, so we've certainly seen that, you know, not only this year, but also over the years, and I think this idea that you can build scale underneath it, and you can build that consistency and differentiation even more on a forelooking

basis using these modern tools. That's certainly a very high confidence view that I have.

work value should work. And then I see other people say, you know, if you find the signal that's statistically significant, that's robust, but you're not really sure why it's working, you could still use it. What is What is your opinion on that?

still very much well in the life. At the same time, I think, you know, we've been working with some of the AI related the professors and at Stamford, at Berkeley and other universities, and I think, you know, there's this professor Stephen Boyd who at Stamford who works with us.

Back fifteen sixteen years ago, he told me, he said, maybe you want to let the data to speak a bit more, and this idea that when we use machine learning, it is actually open minded learning using machine and that data intelligence should come into the process without too much of a human.

has happened for us. So the answer is both. There are plenty of things that we do that has a very strong economic inquation and reason, But then there are stuff that's actually a little bit more machine intelligence that we set it up such that we intentionally want to get data, let data to shine through the in this of an AI machine learning world, and I think that's

especially important. You know, we talked about horizon earlier. If I were to do stuff at the millisecond high frequency space, the economic intuition.

also very important. So I think, yeah, we are we are a little bit of a school of we got to use both and.

as well. So I think you can almost think about min variance optimization as your basic benchmark, and it has its own pluses, it has its own minuses, and some of its modern techniques using machine learning at least, you know for us has you're proven to be very additive above and beyond, and once you get into that world.

You also realize that there's a lot of flexibility in using some of these modern tools, and the search space is a lot larger, you know mein variance optimization from a grading descent perspective sometimes can narrow you into a corner solution, a local maximum pretty quickly, and the search

space may not be as large as flexible. And once you adopt, you know, some of these modern tools, I think a more simulation based approach also becomes a lot more important daily To answer your question is like, you can do this, you know certain methods, a certain way of doing it, but how do you know it is a best and a much more simulation oriented approach. Using these modern methods allow us to really to get a bit of a sense of what the search space is.

So I would say that the application of a machine learning A lot of people think about this essentially it's a fancy large language model and that's about it. But I would say that this is actually what I would like to call it, a big umbrella AI. AI can be applied not only to the large language model, which is certainly very exciting, but it can also apply to

a profolio construction risk management simulation. Many things. So that's so we certainly want to use a full stack of AI as opposed to just one narrow application.

actually we're trying to narrow that particular factor lens. It's still you know, in our legacy, in our history. We certainly want to look at it that way. A lot of market participants certainly look at it that way. But if you open up some of its potential for IDROO syncretic you know, security selection that's above and beyond factors. Once you open that up, you realize that there's actually a lot more interesting things above and beyond the factors, and some of them may carry it to your point,

some factory loading to it. You may want to, you know, model it out and making sure you hedge it and make sure you don't you know, you don't you're getting the purefied exposure. But I see that's the overall risk management is certainly think about the factor risk exposure carefully and try to moderate it. For us, it's actually trying to be differentiated. It's actually just basically not have too much in this kind of factory loading and factory risk. But then it opens up this new world of IDEO

syncretic risk taking. You know, this idea of you can be broad and deep. So this you know, factories is probably only broad but not very deep. And maybe the fundamental discretion investors are only deep but not necessarily broad. And we're trying to do both. And that space, in my mind is a rich opportunity set from an alpha generation perspective.

fashioned systematic quantitative investors. And I think this concept of being modern, I think in the world of AI development is extraordinarily important. So I think at the from allocator perspective, I'll say that's a try to get active managers that are actually moderan on both front. Now, I mean, eventually is there going to be a you know, more AI adoptions for the discretionary managers and more AI you know

adoption in the systematic managers. And therefore this concept of being both broad and h and deep all converge into one thing. There could be but I think this is how modern you are. I think it's a pretty important concept from allocator perspective. I think one more lens, given you know what Christopher was talking about earlier. I also think that this factor versus pure alpha piece is also

very important. What you realize is that when you have multiple strategies, when you put them together, I using credit, alpha sometimes actually gets diversified away and what you are left with is actually a bit of a compounded factory exposure. This is a you know, very important concept from a central desk management, from a multi strategy of the perspective, but I think from allocator perspective it's also very important.

You don't want to just hire ten managers, but all you have left is essentially momentum bed and that's a very expensive way to get momentum exposure. So how do you make sure that the portfolio from a CAT perspectives been driven by idosyncratic security selection in a very diversified fashion and be constraining some of the factory exposure and knowing some of the all alativity associated with it. I think that's a very important question for the educator as well.

to to be cognizant of. At the same time, I think I'm actually reasonably optimistic from a capacity from a you know, diversity of the strategy perspective, in the sense that I think the modern methods and the availability of the data and how you know, you can mean we were talking about how to use this data, it really becomes much more of an open ended question. Uh. It's actually I think it's time for creativity. It's a time for uh, if you will, liberal arts thinking on this.

So if you can do that, there's actually a lot of diversity thinking that you can actually get to drive alpha stream that could be you know, much less uh, you know, get you get you into some of these crowded trades and can open up capacity, and so I think the if a scale player can play the creativity game, I actually think that the capacity will be large, ELPA will be differentiated and you can actually avoid some of

the crowded traits. Now, to build that, I think you you really need to build a large scale platform that is ready to innovate at that scale. So innovation at scale is easily said than done. But I do think that for the platform that I can do that. I think on a forward looking basis, this is actually a very exciting period, if not the most exciting over the last forty fifty years, and that's going to drive longer

term durability to the alpha generation. You deliver that consistency but also differentiation.