0:00
Arbitraging human nature is the last sustainable edge.
Search this video
0:04
When we looked at how do we find an information edge, the obvious place was artificial intelligence. You're the pioneer here.
0:11
You're on the frontier of this.
0:14
These are mathematical functions that have layers to them, and they extract data.
0:22
Machine learning for asset managers is going to move from a "nice to have" to, a "need to have" Deep reinforcement learning really is going to revolutionize and is revolutionizing now.
0:32
The problem is that we were all basically using the same data sets.
0:37
They're equations, they're algorithms, they're only as good as the data that they have underneath them.
0:42
Every other thing we've invented or discovered, is a tool.
0:47
Everybody needs to be responsible.
0:48
The race here is really all about compute.
0:52
There's gonna be huge transformations in our lives, and it's happening more rapidly, I think, than our grandparents.
0:58
"The treasure you seek is in the cave you fear to enter" Wake up and look for the joy in all circumstances.
1:05
Jim O'Shaughnessy: Hi, I'm Jim O'Shaughnessy and welcome to Infinite Loops.
1:05
Sometimes we get caught up in what feel like infinite loops when trying to figure things out.
1:05
Markets go up and down, research is presented and then refuted, and we find ourselves right back where we started.
1:05
The goal of this podcast is to learn how we can reset our thinking on issues that hopefully leaves us with a better understanding as to why we think the way we think and how we might be able to change that to avoid going in infinite loops of thought.
1:06
We hope to offer our listeners a fresh perspective on a variety of issues and look at them through a multifaceted lens — including history, philosophy, art, science, linguistics, and yes, also through quantitative analysis.
1:06
And through these discussions help you not only become a better investor, but also become a more nuanced thinker.
1:06
With each episode we hope to bring you along with us as we learn together.
1:06
Thanks for joining us, now please enjoy this episode of Infinite Loops.
1:07
Disclaimer: Jim O'Shaughnessy is chairman and Co-Chief Investment Officer of O'Shaughnessy Asset Management, where Jamie Catherwood is an associate.
1:18
All opinions expressed by Jim, Jamie and podcast guests are solely their own opinions and do not reflect the opinions of O'Shaughnessy Asset Management.
1:23
This podcast is for informational purposes only, and should not be relied upon as a basis for investment decisions.
1:33
Clients of O'Shaughnessy Asset Management may maintain positions in the securities discussed in this podcast.
1:37
Jim O'Shaughnessy: Well, hello everyone, it's Jim O'Shaughnessy with yet another edition of Infinite Loops.
1:47
Today we are going to learn, my guest, Julie Bonafede, I love that last name, and yes, she did tell me how to pronounce it, is the co-founder of Rosetta Analytics, which is using artificial intelligence and machine learning in a unique and interesting way.
2:06
Some of you are obviously aware that I too am very interested in AI after I announced that I had invested in Stability AI and became chairman of the board over there, so I think that Julie and I will have a lot to talk about.
2:22
Now also, what a killer resume you have.
2:22
24 years at Wilshire, president of Wilshire Consulting, you're on the board of directors on the investment committee.
2:34
You've got an MBA from USC and a CFA as well.
2:34
I feel like, okay, I am credential-less other than a kind of a BS BA. Julie, welcome.
2:41
Julia Bonafede: Thank you.
2:45
Thanks for having me Jim, and we all know how accomplished you are, O'Shaughnessy Asset Management.
2:49
Jim O'Shaughnessy: Well, so obviously this is a topic near and dear to my heart.
2:51
People get unusually frightened sometimes when I'm talking to them about what excites me about machine learning and artificial intelligence.
3:06
And I was just thinking even the term itself is a new term, right?
3:11
John McCarthy coined the term artificial intelligence at the Dartmouth Summer Research Project in 1956, and then you've got my friend Jeremiah Lowin.
3:16
I'm also an investor in his company Prefect, who says, "Listen, calm down about AI because it's neither artificial, nor is it intelligence, it's equations.
3:28
And the equations are meant to be good tools for we humans to make better decisions."
3:34
So what I'd love you to do, because many of our listeners and viewers are not going to be familiar with artificial intelligence and machine learning specifically as applied to financial markets, if you could give me the background about what led you to decide...
3:50
because you were a pioneer in factor investing, you've been a real pioneer.
3:57
We are very simpatico in many, many ways.
3:57
And I find it interesting that I too now am looking to AI and machine learning for new ability in active portfolio creation.
4:01
But please walk us through how you changed your view, or how your view evolved to get you into being the co-founder of Rosetta Analytics.
4:17
Julia Bonafede: That's an interesting question and it certainly was an evolution because the roles that I had at Wilshire, two very different roles, but very interesting.
4:27
The early part I worked in what was...
4:33
they called it institutional services.
4:33
But as you know, Wilshire was a pioneer in portfolio analytics software, and truly it was Wilshire and Barra that created applications, at that point time we called it software, for portfolio analytics and really risk models.
4:46
And that was the advent of all of the research that was being done trying to quantitatively discern relationships from underlying market data.
5:01
Wilshire was the one to first commercially sell betas.
5:08
They had beta books, that preceded me, but that was the origin of Wilshire.
5:15
And if you think about it, the whole concept of alpha and beta really had its heyday in the mid to late '90s, and so to be able to actually calculate it and use it as a data point, to constructing a portfolio, or buying a stock, or putting any underlying portfolio that you were going to look at was really unique.
5:37
And that whole body of research evolved through time, we all had our foundations in CAPM and modern portfolio theory, but if you think about what the limitations of alpha and beta were I like to show this chart where the early part everything was skill, right?
6:00
And then once common factor risk was identified and could be measured, then the concept of alpha, or something residual, that was skill.
6:07
And then after further research by some of the innovators in factor research you saw additional factors being deployed, exotic beta, in a very simple form, took that to be that our measurement tools became better in terms of identifying relationships and data.
6:31
So in terms of how that evolved and then looking at working with asset managers and pension funds that had big internal shops on their analytics, and then moving over to run consulting where obviously you take a much larger lens and it's all about asset allocation.
6:51
But then the saying was, every consultant has a funnel.
6:51
And so the whole- Jim O'Shaughnessy: I've gotten caught up in those funnels often, so I know all about them. Julia Bonafede: Exactly.
7:04
And at every manager had a funnel so you would screen for certain characteristics.
7:04
But one thing that was always really clear when you look through data, and I don't mean to knock any firms out there, but active management has been a tough road to have any consistent outperformance of a market benchmark or whatever your target is through time.
7:28
There always seems to be a cycle where active management does better.
7:34
This might be one that we're in right now, let's hope and see what happens.
7:40
But a lot of, I think what's happening in the evolution, I mean we always used to watch these charts of the convergence of active versus passive.
7:44
And so the essential element in Jack Bogle, he introduced this concept that you want to pay the least amount of beta that you possibly can.
7:59
And so, having a well-constructed portfolio, or if you have a market that's clearly, there isn't an information edge, you're going to want to index that portfolio.
8:12
So everybody developed these core satellite approaches and then that basically takes the economic set of active asset management, unless you're highly specialized and you have the ability to access the very best managers.
8:21
But Wilshire always did an active management study to look at really the core asset classes.
8:27
And to this day there's really is no, maybe outside of small-cap, there's no persistent value-add over time, persistent and consistent over consecutive time periods to active management. And then there's fees.
8:48
So as I was looking at what Julia 2.
8:48
0 is going to be when I left Wilshire, it was, there's a lot of opportunity in asset management in this business, but at the time, factor based investing had become ubiquitous.
9:03
And so what was a measurement tool became a more of a smart index.
9:13
So we knew that the fundamental indexing that research affiliates was doing and you saw even more indexes adopt that.
9:18
And then there was even businesses where you produce as many factors as you possibly can.
9:25
I think there's one large bank out there that's producing 80 different factors.
9:30
Well, at that point you put that many factors together and you have an index fund or something quite close to it and the correlations where when you look at the behavior of these strategies, really not that different.
9:41
And so you think about how you replicate an investment strategy, you can do that by putting all different kinds together, but you're still not going to get anything cheaper than a cap weighted index.
9:52
And low volatility has had its day, there's been so many flavors of this.
10:00
So looking at this, what's the tool, what's the technology out there to give an information edge?
10:06
If you're going to look at quantitative investing, I'm going to put fundamental investing aside.
10:11
That's a whole different animal.
10:11
And also struggling, I believe.
10:11
I think assets, there's a lot of inertia and asset management.
10:18
So it takes a long time for this migration, especially from active to passive.
10:24
And passive is already surpassed.
10:24
So when we looked at how do we find an information edge, the obvious place was artificial intelligence.
10:37
And as you know, artificial intelligence is a big umbrella term for any type of model that is simulating human intelligence.
10:44
And then underneath that you have machine learning.
10:50
And under machine learning you still have, it's been rebranded, factor investing.
10:56
The underlying algorithms are machine learning.
10:56
They're classical linear regression or PCA or whatever types of models that you're using.
11:03
And then underneath that you start to get into neural networks.
11:08
So you have deep learning.
11:08
And then there's another discipline called reinforcement learning.
11:13
And then the middle of that taxonomy is deep reinforcement learning.
11:17
And that's where we decided to really spend our research dollars, is to see if we used a neural network, could we identify an information edge that could be consistent or at least dynamic?
11:34
Jim O'Shaughnessy: And I would love it because I've spent the last, I don't know, six months doing a massive deep dive on artificial intelligence.
11:45
And you are absolutely correct that there's a lot there in terms of understanding, because sometimes I think the general public misunderstands exactly, as you did a great job of branching it down, to the deep reinforcement learning.
11:59
So if you would, again, for our listeners and viewers who might not be familiar with the idea of deep reinforcement learning, could you give us a definition and how it's used? Julia Bonafede: Sure.
12:15
So deep learning uses a neural network and the models are really trying to minimize a loss function.
12:25
And that can be complicated.
12:25
But think about back to a high school algebra 2, where you were creating linear systems or linear algebra.
12:33
And that's basically the foundation of multifactor risk models.
12:40
You're going to identify or slice up your beta.
12:40
The initial equation was always Y = MX + B, right?
12:46
That's how you draw a straight line through data points.
12:46
And the straight line is the point where you've minimized the ability to explain all of the data.
12:54
So it's your line of minimizing that distance between the data.
13:06
And so neural networks are nonlinear, by nature.
13:12
And so they are looking at the curves around the plane.
13:12
So if you get a data set and you want to try to glean as much information as possible, these are mathematical functions that have layers to them and they extract data, they're excellent feature extractors, from data sets or inputs.
13:35
And so if you're looking to determine relationships in data, they're excellent at identifying relationships.
13:43
And as long as they're trained properly, then those relationships are then generalized.
13:50
So that means that the memory banks of the neural network will store these generalized relationships that it identifies in the data.
13:58
And then when it is presented with prediction, so if you're trying to classify or predict, in our case, we try to directionally predict a particular market using those models.
14:11
The neural network will mathematically reweight all of the underlying relationships that it sees until it finds the optimal or the closest model to describe the current data set that it's ingesting to produce the prediction.
14:25
Or if in a different kind of neural network, if it's not time series, there's image recognition.
14:32
Everybody has seen the data, probably, about putting in a million pixels and being able to identify a dog or a cat or a not dog, and it's really widely used in image recognition.
14:44
So think about diagnostics and medicine and there's lots of problems that need to be solved and are being solved that you weren't able to before because with the advent of compute power and the ability to have large data sets ingested or these models really are most adept at being able to reduce the dimension of that data and usefully produce an output that describes or solves the problem, depending on what that question is.
15:23
So that's a deep link learning model and there's a mathematical process, so where it reweights.
15:28
But then you move over to reinforcement learning.
15:34
That's actually a very advanced optimization framework.
15:34
It's a decision making framework where prior, at least on the investment, where we've been used to using methods like mean variance optimization, where you're really trying to still reduce distance, but you're trying to reduce the variance or the variability around whatever it is that you're trying to solve.
15:57
And so in deep reinforcement learning, now you have the ability to look at decisions that the model might make extracting the data, but these models are sequential and they are able to iterate in many steps from the distributions that they create.
16:20
So you basically have a sequential, multi-step optimization framework for making decisions that when combined with deep learning or a neural network, as the model is reweighting all of the features or the relationships that it has described, the optimization framework is taking all of those permutations and it is optimizing what the next step should be, in terms of whatever it is, like autonomous driving or search engines.
16:57
And this is the selection that you should use and this is how much distance you should have, and this is how you moderate from a robotic standpoint.
17:02
I mean, think about direction or repetitive motion or anticipating a movement.
17:07
That's where deep reinforcement learning really is going to revolutionize and is revolutionizing now.
17:14
But it's still, I believe all of this is in its adolescent, maybe child stage, in terms of where the development is and the potential is, because of all of the issues that you're seeing, its stability AI.
17:30
Jim O'Shaughnessy: And one of the things that actually had been a thesis of mine for a while was that, well, black swans, by their very definition can't be predicted, at least thus far.
17:43
We'll have to see where AI goes with that.
17:43
So one of my theses was, well, it can't be predicted, let's just take that as a given. That might be wrong.
17:52
But let's just say they're not predictable.
17:57
Is there a way that we could use machine learning and AI to confirm that a black swan has occurred?
18:03
So take for example, when oil went negative, right? Remember?
18:13
If we could have had a confirmation from an AI driven system, even a week before the "market" or the traders figured out that something really hinky was going on with oil, obviously I don't have to tell somebody like you, having a trading advantage like that would be pretty incredible.
18:32
Julia Bonafede: That's getting in crystal ball territory.
18:34
And I guess from anticipating a movement, we can discuss whether or not you can predict exogenous events like that.
18:43
As I recall, there potentially were some bad actors that were influencing that particular event, which I don't know how, outside of the NSA, the Department of Defense, I don't know how you get that kind of inside information to be able to feed into a model.
19:05
Jim O'Shaughnessy: And so I found, when I went on my quest.
19:10
I'm very lucky to have a pretty broad and deep network, one section of it is populated by deep learning, machine learning experts.
19:18
And so I went to one and I gave him the much more elaborate thesis.
19:26
And he's like, "Really interesting.
19:26
That would be really cool.
19:32
Do you have an extra $500 million lying around?"
19:32
And I said, "You know, Dan, I don't."
19:38
And then we got into compute time and what you just mentioned.
19:38
The depth of information needed for those types of either confirm it...
19:44
I would use the term confirm, we can also use the term predict.
19:52
And how that might get captured.
19:52
But I want to follow up because of the term you used that there were several bad actors, which everybody now knows about.
19:57
But the thing that interests me in that is, that isn't one of the premises of machine learning that the AI is particularly good at non-linearity, which is something that humans are really, really bad at? What do you think?
20:19
Julia Bonafede: That's an interesting question because humans, I think, can be good at non-linear thinking.
20:21
In fact, I think that's where you see a lot of this research.
20:28
And I think a lot of, maybe the uneasiness, between AI usurping a human role, which will be the case in terms of, I really think, repetitive functions that maybe those types of skills can be replaced.
20:43
But I think there's demographically, I think this is still okay for humanity because there's issues in terms of people that'll be there to actually fill these kinds of roles.
21:02
But also in terms of intuitive thinking, if you think about how someone learns.
21:02
And I mean you have children and so remember how they evolved and learned, but there was always these interesting times where they would just know something. It was intuition.
21:19
And I think that's still the case to it today.
21:26
And models, as you point out, they're equations, they're algorithms.
21:26
They're only as good as the data that they have underneath them.
21:31
And so I think there's still going to be the need for that intuition and non-linear thinking on both sides.
21:37
Now there's some cases, I mean that's how we built Rosetta, is we wanted to get the human out of it.
21:43
Because one of the big issues that I was really interested in is, if the whole world is investing using factors of some kind or measuring some kind, then if you take the other side of that trade, there has to be information left on the table that isn't being captured.
22:03
And what humans aren't good at, especially when it comes to markets, is the emotional side of losing money.
22:11
I mean the most calculated person, the best deal maker, is the one that will walk away all day long and they just won't participate in the emotional trade.
22:26
It's a true skill set to be that clinical in your thinking.
22:34
And that's where these models really, I believe, pick up on those trends.
22:40
So back to your question about bad actors and being able to predict black swan events, you may not be able to predict the event itself or there might be rumblings in the market.
22:50
You're seeing it right now.
22:56
We're sitting on, in my view, a powder keg in the markets right now in terms of leverage and all the liquidity that's being pulled back.
23:03
When those markets destabilize like they are, there are patterns in terms of how they behave.
23:12
We just went through a period of, “buy the dip”, I mean that was COVID.
23:18
But there was also some very strange behavior that happened at the close of the market every day.
23:25
You would watch it and say, " Okay, who's coming in here and changing the dynamic?"
23:31
But that kind of behavior gets picked up by these models.
23:40
One of the pieces of deep reinforcement learning that I forgot to mention that's key to it, is that these optimization frameworks, they learn and adapt.
23:44
So you can see from deep learning that they're keeping those relationships and then they're able to look at a new set of data and say, "Okay, did this behave this way in any other regime in the market?
23:57
And can this be used to describe the situation?"
24:02
But the deep reinforcement learning, those models have trained and they are adaptive because the decisions are reinforced.
24:08
So it's like your infinite loops but with a reward and penalty function.
24:13
So you have good decisions that the model makes, that are mathematically rewarded.
24:20
And then less productive decisions are penalized, mathematically.
24:26
And so the optimization framework or the network learns from its mistakes and its rewards and it's looking at maximizing that outcome.
24:33
So it's maximizing a reward.
24:33
So when you can get at the best information edge and maximize the outcome of using that information, that is the key to deep reinforcement learning.
24:46
And I think that's the key to seeing when there are changes in regimes, you get to it much quicker than humans and those traditional linear models, like factors.
25:03
Because factor investing, and I don't know if you would agree with this, but a lot of it is, when can you time those factors? Jim O'Shaughnessy: Yeah.
25:13
So boy, there's so much there that I want to follow up on.
25:13
The saying I'm probably best known for is that, "Arbitraging human nature is the last sustainable edge."
25:20
Because essentially of what you just detailed.
25:28
And actually, one of the things that I am proudest of over my career as a empirically factor based investor is not overriding models ever, including in the great financial crisis, including the dot com, et cetera.
25:41
And that of course, it's like Odysseus latching himself to the mast because he wants to hear the sirens sing.
25:48
That is very easy to say and it is almost impossible to do consistently.
25:55
So I got asked once in an interview because I'm retiring from O'Shaughnessy Asset Management at the end of this year, and I have a new venture that I'm moving onto.
26:06
But when I was asked about it, I think, the fact that I never overrode a model was pretty cool.
26:12
As far as the timing factor, you're probably smarter than we are because we haven't approached it from machine learning since.
26:19
But as you say that, that's like Homer Simpson, "D'oh!"
26:26
Of course that would be the way to approach the factor timing.
26:33
But the other observation that I want to ask you about regime change.
26:33
So one of the things that we found by using...
26:40
And I think by the way, one of the weaknesses of the traditional quantitative approach that I championed in my book, ‘What Works On Wall Street’ and you at Wilshire and Cliff Asness at AQR, the problem is that we were all basically using the same data sets. Julia Bonafede: Yes.
26:59
Jim O'Shaughnessy: And you do get to a point, I got to a point, where it was just like, one of the projects that we were working on before we actually decided to focus more on our custom indexing software that's called Canvas, was, I was actually creating a data set from Moody's Manuals that went back to 1900.
27:20
And we were lining them up with the CRISP data set, but that's all by hand.
27:28
We were doing it offshore.
27:28
And so, in that quest to say, "Oh, wouldn't it be great to have an additional 40 years of data?"
27:37
Well we stopped that because we went on to both focus on Canvas and then ultimately sell our company to Franklin Templeton.
27:44
And so, back to the concentrated findings, LSV which is a well known quant with the value tilt, AQR, OSAM, name the quant with a few exceptions, we're all operating off highly similar data sets.
28:05
And quite frankly, we're also operating on highly similar models, because we're inferring from the data, pretty much the same thing.
28:12
One of the things that we did notice when we did a study of what I termed severe bear markets, i. e.
28:18
I think our litmus there was a loss of 40% or more, was that some factors inverted.
28:26
So momentum, if you use the CRISP data set and you go back to '27, momentum is a reasonably good factor, particularly in smaller names.
28:35
We looked at it from '27 to now.
28:45
We found that in general, momentum works very well.
28:45
Poor momentum is a great identifier of stocks to avoid.
28:51
I wouldn't go so far as to say short because the peculiarities, but we found that momentum inverted after the market bottom.
28:57
So, let's take the most recent big dip, which was the end of '09, momentum inverts and the decile with the worst momentum moves to the top of the stack in terms of return, whereas the highest momentum decile does the poorest.
29:18
It seems to me intuitive... Back to that word.
29:18
It seems intuitive that these deep learning models would be able to pick something like that up much faster.
29:24
What else do you think they can pick up faster than a human, even a super smart human or a group of humans, would infer?
29:37
Julia Bonafede: Well, I think, again, the black box is probably the issue that people most point to, in terms of these models and being able to interpret them.
29:50
But the trading behavior, and the way that the signal, the strength of the signal, and as you know, there's different models and how they give their signal.
30:08
And some of the signals are stronger than others, especially in particular markets.
30:08
Where we see our models be the most active in terms of trading, is when there's distress.
30:15
And it's intuitively that there's been this distress before, when the relationships in the data that are identified, you see the same behavior to distress.
30:31
And so, when you think about the great financial crisis, that was a crisis of many things.
30:41
But leverage always seems to be at the bottom of all of these.
30:48
Jim O'Shaughnessy: Always.
30:50
Julia Bonafede: And so, at some point, there is a point in the market where something that you didn't anticipate...
30:54
Like what happened last week in LDI, these models tend to be constructed over time periods that have gone a certain way, and then the relationships that are traded, it work, when you say there's a momentum inversion.
31:18
Well, the non-linearity that is present...
31:18
So, it's not that these models may not be picking up on momentum, or whatever you would call it in a feature language, that some trend, or whether it's a growth value, size, whatever the underlying principle is that's causing the pricing to change in the underlying data for the instrument that you're trying to predict.
31:36
It's the behavior of how the market's responding to it.
31:42
And so, that I think is the key piece, is getting to whatever the relationship is to the other relationships that the behavior introduces.
31:58
Because one of the key memories I have of March, 2020, sure that was a COVID story, but that was an oil story, that was a pipes freezing story, the repo market, everything at one point, one day, there was some real fragility that went on.
32:17
And that fragility is probably very similar to the great financial crisis.
32:23
To think about one event that defies logic is December, 2018, when Jerome Powell said a few things that disturbed the market, and everybody panicked.
32:41
All the fish went to the side of the bowl.
32:41
Jim O'Shaughnessy: Exactly.
32:42
Julia Bonafede: So those things, those are in the data, and the data is ingested, and those relationships continue to be refreshed.
32:52
And that's where these models, because they're continuous models, I remember back in the day, when we used to build covariance matrices, underlying the Wilshire risk models.
33:03
Well, there wasn't enough compute power at that point to quickly deliver these models.
33:03
First you had to wait for the data to come in, then it had to go through the production process, and it could take two or three days to produce one.
33:17
So, the obvious answer to that, was to shorten the data series.
33:23
And so you went from, clients would ask, "Well, why don't we have the whole history?"
33:28
Well, five years of history should be plenty of monthly data, or daily data, or whatever it was we're using daily.
33:35
Then these VAR models came, and there was a different statistical waiting mechanisms, to be able to highlight recent data versus historical data.
33:48
And it's all constraints that are being put on models and data, to help inform whatever the question's being asked.
33:59
So, what problem are you trying to solve?
33:59
And I think that's where we got ourselves to in the industry.
34:06
Just thinking about information ratios, and how active share, that was something that came after information ratio.
34:16
And I think about the low volatility that was, well this is how you're going to reduce your volatility so you can compound more efficiently, because don't want to be in equities, they're too volatile.
34:29
Look what they did to us in 2009.
34:29
It's always a mathematical response, to put a boundary around an outcome.
34:34
And I think that true alpha, you have to swing for the fences, and take it for what it is.
34:42
Right now the conversation is, we need uncorrelated return streams.
34:48
Everybody's out looking for an uncorrelated return stream. Well, why is that?
34:53
Well this is the first persistent time in anyone's investment knowledge that has a shorter career history that's seen stocks and bonds move together.
34:59
And since the 1980s, the markets have...
35:08
At least the treasury markets, have only had a downward trend in yield. That's over. So, what do you do now?
35:15
And that's where adaptive models are going to need to be in place in some form, maybe not...
35:21
Think about your use cases.
35:21
There's many use cases that you can use for a deep learning, a deep reinforcement learning model.
35:28
It doesn't have to take over your whole investment process.
35:33
We decided to go to the very frontier, because we wanted to see if we could do it.
35:38
And it turns out we could, but it's slowly proliferating into investment shops.
35:38
But there is still this barrier to wholesale adoption.
35:46
Jim O'Shaughnessy: That's what I want to ask you about, because one of the things when I was looking into this was...
35:55
And I always try to get the views of a wide variety of people, with a wide variety of takes on things, just because I need to remind myself how often that I am probably wrong.
36:03
And so, it would be helpful for error correction to get as diverse a group of people asking me these questions.
36:16
And one of the things that I found when I was talking with them about my thoughts about AI, and machine learning being the next place where this is going to actually...
36:32
I'm of the opinion that machine learning for asset managers, is going to move from a nice to have to a need to have.
36:39
And because of Stability AI, I know what the stack looks like, and I know what releases are coming, and they're really cool.
36:47
But here's the thing that I want to ask you, because you actually have experience in this as an active manager.
36:52
How do you overcome the almost unique human need for ‘why?
37:01
’ And one of the things that I found when I was trying to...
37:01
I luckily am just constructed...
37:10
Maybe it's because I've always been an algorithmic guy and equations guy, but I'm constructed so that, I'm quite happy if I have a huge and high base rate on a model's prediction success.
37:22
And it tells me what and when, but it doesn't tell me why.
37:27
And one of the things that I found with our client base was...
37:27
And obviously, it's probably similar to your client base.
37:34
And yes, I want to get into that.
37:34
Actually, it might not be similar.
37:38
But the hurdle that I really fought hard to reframe for people, was the ‘why.
37:46
’ And to be glib, you can say, "Well, we invent why when we're talking to our clients, because they like a story and that's human OS.
37:52
And we're not going to try to deviate from that."
37:52
But what has your experience been with your client base?
37:59
And if you wouldn't mind, if that's open information, do you have a particular style client base that differs from other asset managers?
38:11
Julia Bonafede: Well, we're an emerging firm, so we're still out there beating the pavement for lots of clients.
38:13
But our main client is Verger Capital Management, which is the Wake Forest endowment portfolio.
38:19
So Jim Dunn, he was the former CIO, and he's built a great business.
38:26
So, he's seated our strategies.
38:26
So, when we just came up on our five year track record, and he was the reason that we went the direction we did, in terms of what markets we were going to begin this in.
38:39
And so, that's how we landed with trading the S&P.
38:39
And so, this is the classic question, how do you get investors, clients comfortable with something that you really can't say, this is why.
38:54
Now obviously, I have a background in analytics and in consulting.
39:02
And so, the ‘why’ has always been a big part of this is how you interpret results.
39:09
So, we've built a pretty elaborate analytics system that we feed all of our strategies into, so that we can put it in an investment framework, and help people understand at least the typical metrics.
39:23
And they don't necessarily look the same, because we don't use leverage.
39:23
We're not large enough yet to take that risk.
39:29
If a particular client wanted it, that would be fine.
39:29
They're actually built to leverage.
39:36
But if you think about a lot of the returns that are generated by CTAs, there's leverage that it's being employed to amplify certain bets that they're taking.
39:42
And so, that's what I think is so interesting about what we've produced, is this is pure signal we're trading, there's no window dressing around it. And so, when we have...
39:56
One of our strategies is really outperforming the S&P significantly, plus 50% right now.
40:03
But the last couple of years, we're coming up on a five year track record.
40:13
But from March, 2020, we saw these in incredible outsized returns, that we believed was a proof case.
40:22
Because all of the back testing you see, great financial crisis.
40:28
Everybody can produce a back test that looks great on that.
40:28
But then, to be able to see it in live data.
40:36
And March, 2020 was our moment when, oh okay, this is where it starts really picking up on this.
40:42
But at that point, we were only three years into it.
40:42
And so, every once in a while you need a second case.
40:50
Well, now we're having that second case, and the return streams are uncorrelated.
40:55
And so then, you have to say, "Okay, well why are they uncorrelated?"
41:00
But if someone presented you, what are the characteristics of an uncorrelated return, what would you say?
41:06
Jim O'Shaughnessy: What are the characteristics of an uncorrelated return?
41:08
I would say that the underlying asset class, and/or factors that describe that asset class are very different or the behavior is truly, there's a zero correlation between the two asset classes.
41:24
And that somebody had built a very clever model, that took advantage of all of the knowledge that we have about the correlations of things.
41:34
And as you know, most financial assets are highly correlated, and it gets really difficult to find one that's not.
41:40
The closest we came with is, our microcap approach.
41:46
But even that, it tends to move out of sync with the general market.
41:46
And yet, when we have situations like today, the correlations go to one.
41:53
And everyone throws the risk assets out, they rush to the safety of riskless assets.
42:01
And that's why when I was looking at your stuff, that's why I was so terribly intrigued.
42:07
Because you really are uncorrelated, which is talk about a black swan.
42:12
Julia Bonafede: So, think about some of the very famous tail risk hedging strategies that are out there.
42:16
And you really, it's expensive to put that on so that it's there when you need it.
42:25
So, you're going to pay a lot for it when you don't.
42:32
And then, it's going to be great when you don't...
42:32
Or some board is going to get upset that it's that costing that much, and unwind it.
42:39
It's human nature to rationalize where you are, rebalance, I'll call it. Jim O'Shaughnessy: Sure. That's very kind of you.
42:51
Excuse me, we're over here panicking, but let's call it a rebalance.
42:59
Julia Bonafede: And so from our standpoint, we have these very...
43:00
When you have an orthogonal return, but the markets are going up exponentially, a lockstep over and over, like you saw in 2020 and 2021.
43:14
You get lulled into complacency, but you don't necessarily want that insurance policy.
43:14
So, it's really explaining that you're going to get additional return to your beta, but sometimes your beta is out of whack.
43:30
And so, think about on average, which we don't like averages, but at some point, when you think about what...
43:37
If you're going to invest in an asset class, say equities, and you aren't expecting the long term average return, or whatever the underlying risk premia is that you develop, whether it's four or 5%, then that asset class, you shouldn't be in that asset class, and you shouldn't expect outsize returns in your portfolio.
43:58
And so, really it's how you think about risk, how long your investment horizon is.
44:03
I sound like a financial planner.
44:08
But it's true, there are maxims to investing, that the shorter term your horizon, the more that you're going to probably get burned.
44:17
Jim O'Shaughnessy: And or be just completely constrained. Julia Bonafede: Exactly.
44:24
And now, with 4% yields on short term treasury securities, we're back to at least an environment where savers are going to hopefully be able to be rewarded for lending their assets.
44:38
I think that's what's been so out of whack for so long.
44:43
So, now you can put an uncorrelated return stream in your portfolio.
44:43
That's diversification as long as you realize that what I was getting at is sometimes uncorrelated return streams, are negative too for time periods.
44:55
Jim O'Shaughnessy: Oh, yeah. For sure.
44:57
Julia Bonafede: There's no perfectly positive asset class outside of cash, that you're going to.
44:58
And even that, has been...
45:04
Depending on where you live, has been a challenge globally.
45:04
So, when you're paying an investor to hold your money, when you lend your money to an investor that's...
45:11
You're paying them, that something's completely out of whack.
45:16
So, this is where I think our types of models can shine.
45:24
Hopefully, our clients and our investors, and potential investors, see that there's a surprise element to it, which I think is most important. It's a trust element. Do we trust the model?
45:30
Well, you did it before, but past performance doesn't guarantee future performance.
45:37
So, that's where we're getting to with these models.
45:42
And what I worry about is that, the fact that they're opaque, that either through regulation, or through some fiduciary responsibility, that there won't be an appetite because of the somewhat unpredictable behavior that these models have.
46:03
And it really depends on the type.
46:03
Our deep reinforcement learning models, just have lower beta, but they're equity-like.
46:08
They're more equity replacement tools, because the maximizing of the roar means that the model is trying to maximize the efficiency of the compounding of that portfolio return, which is really important, because what you're at the end of the day trying to do, is mitigate losses.
46:28
Because if you have permanent losses of capital, you can't efficiently compound a portfolio.
46:33
Jim O'Shaughnessy: Yeah, I agree.
46:37
I ended up solving for this by, as I mentioned, investing in what I think is one of the best up and coming full stack AI company.
46:43
Because I agree with you on the, how are you going to speak human, so to speak, and deal with human operating system, which hasn't changed very much in millennia.
46:57
And yet, also try to explain in ways that regular intelligent people, obviously your clients are going to be intelligent, can grasp and understand.
47:06
I personally have a view, which is why I decided to solve for this by buying the interest in Stability AI, as opposed to trying to develop Grey Swan, which is the name of my subsidiary, that we're going to look at using machine learning.
47:22
And I think that ultimately, this is going to have to be widely adopted in asset management, simply because of performance, simply because of edge.
47:37
And I learned a lot more since joining the board of Stability AI, and having all of those geniuses explain things to me.
47:45
But really, the race here within the AI itself, not its applications for us, or for medicine, or learning, or any of those.
47:53
But the race here is really all about compute.
48:01
In addition to being about compute...
48:01
And here I am using terms that maybe our audience will scold me for, because it's jargon of AI.
48:07
Compute is just how many super computer and they're not super...
48:14
Well, some are super computers.
48:14
But how many of those chained computers, do you have together?
48:19
And that determines your ultimate compute power.
48:19
In other words, how quickly you can run the model through its cycles. And then, scaling.
48:30
And then, the idea that everything is one model being forked in a new line of code, coming out like transformers with GPT-3.
48:37
And so I think, that this is going to become a must have, but I'm not sure that it will ever become a focal point, unless folks like you break through.
48:47
And so, where you're breaking through... So, help me out here.
48:54
If I was a potential client of yours, what would the pitch be?
49:02
Would the pitch be, let's say I'm all in on AI, I love AI, is the pitch non-correlated assets, is the pitch risk minimization? What is it?
49:14
Julia Bonafede: Well, for the deep learning models, it's uncorrelated return.
49:17
So, we have enough of a track record to measure that that's the case.
49:25
And also, to be able to demonstrate that the model is going to pick up without outsized returns, in these periods of stress, as it captures a lot of elements in the market.
49:38
On the deep reinforcement learning side, it's really going to be an efficient compounding of returns.
49:43
So, in terms of when you think about portfolio construction, you typically aren't putting all your eggs in one basket, unless it's an index fund.
49:47
And so, putting together a portfolio where you can have the most efficient mix of, in this case equity assets, that can help you smooth out that downside risk, but then also help you continue to grow through time, it's that efficient compounding.
50:09
So, we're really looking at really strong risk adjusted returns through time, and it's just unique in terms of its edge of how it plays well with beta.
50:22
So, it's a portfolio structuring, and that's a little bit more nuanced, I think, from an investment standpoint.
50:28
But if you think about how the portfolios are designed, typically it was by asset class buckets, which nobody likes.
50:35
But it's really more about how do the different strategies that you have in your portfolio work together to achieve your ultimate objective from an investment standpoint.
50:48
And so, we believe on the deeper reinforcement learning, that these models really play well.
50:53
They don't have a typical profile because the underlying structure of the portfolio can change through time as the models finding new regimes.
50:58
So that's how they fit, and it's really ...
51:03
you have to sit down and really do the analysis of how that would work within your investment structure.
51:08
And that's what we work with clients to help them see those relationships.
51:13
Jim O'Shaughnessy: Yeah, it seems to me, listening to you, that I could see you as a absolute natural fit for a Wilshire, for a big consultant or a Mercer or whomever, because the explanations here, almost out of necessity, are pretty complex in and of themselves.
51:33
And so I'm always just trying to figure out, who would be most interested in this?
51:41
And obviously institutions spring to mind.
51:41
Anybody who has a process in place for active asset allocation should be very interested in this.
51:49
But I wonder if you did some kind of partnership with a bigger provider, whether this couldn't also be packaged almost into an active ETF. What do you think? Julia Bonafede: Yeah.
52:05
I think that's where it's interesting, in terms of creating the well diversified portfolio where you can play well within a structure, but it's not based on a static allocation, because you're going to see those relationships change through time.
52:21
And a lot of times, as we've talked about earlier, the factor relationship doesn't exist.
52:28
But in traditional models, in order to change those factors, you have to completely change the models.
52:35
So you only have one lens to look through.
52:41
With the neural network, those relationships are constantly being re-parameterized and so to develop that optimal system, and that's what is so unique, is you're not stuck with four factors or five factors or seven or whatever the number is.
53:01
And you don't have to go back to the boardroom and think about, "Okay, well how else are we going to look at this market now and describe it now that it's changed?"
53:06
That's what happened in March 2020.
53:11
All of these models, they fell apart.
53:11
And so what do you do then?
53:11
Our model didn't do that.
53:11
It just kept ingesting data and kept adapting through it.
53:20
And that's where I think ultimately you have to get, at least from a systematic approach.
53:26
Jim O'Shaughnessy: And where would you be on the price structure?
53:29
If we did do that ETF that could be available both to institutions and individual investors, what kind of expense ratio would you be looking at?
53:42
Julia Bonafede: Well, I think that it depends on what you're pairing it with and so what the overall allocation is.
53:45
But I think if you think about anything from an index fund where you're getting six to 10 basis points in terms of your expense ratio, that's too low.
54:01
And so from that active alpha place, I think you're somewhere at the 1% range.
54:07
You're not going to get the carry that you would in a hedge fund, but it's more of a scale. Jim O'Shaughnessy: Yeah.
54:14
Actually, that leads to my next question, which is why didn't you just do that was going to be what you are, you weren't going to think about ETFs, you weren't going to think about any other use cases?
54:24
Julia Bonafede: Oh, okay. Well, you have to adapt.
54:24
You have to get adoption.
54:27
It's interesting, recently in the news has been a large systematic investor that I think uses some sort of networks in their models.
54:33
They must be capacity constrained because they just went to three and 45, or- Jim O'Shaughnessy: I saw that as well.
54:43
Julia Bonafede: And so we're a startup.
54:47
We're going to try to price ourselves in a place where if you can get your mind around the technology, then the price point is palatable as well.
54:59
Jim O'Shaughnessy: I agree.
55:01
And that leads me to my next non-finance question, but which has been one I've really been struggling with because of my new role at Stability AI.
55:07
First off, I love science fiction.
55:14
I have read virtually all of the classics, the Foundation, the entire series, Dune, you name it.
55:22
And it seems to me that science fiction has done those of us interested in AI a disservice in that it has consistently ...
55:30
What's the great John le Carré quote, "The cat sat on a mat is not a story.
55:40
A cat sat on the other cat's mat, that's a story." And so, I get it.
55:40
The stories have to be fun and maybe scary and/or inspiring.
55:47
So this idea of equating AGI, that's artificial general intelligence, which I don't think is happening much, as fast as many other people think.
56:00
But anyway, it seems mostly when I've been talking to people after the announcement that of my involvement with Stability AI, of course, what do you think the first thing they bring up is? Skynet and terminators.
56:12
What do you think, do you think we're going to get more positive stories about AI and machine learning because of where we are actually now with the innovations in the technology?
56:25
Julia Bonafede: Well, the more that you see it or at least you're aware of how it's used positively.
56:26
So diagnostics, that's definitely going to impact.
56:37
But I don't know how much the everyday person is paying attention to really how dramatic the ability to detect cancer has become, for example, and how it's really assisting radiology.
56:51
And in all of those, it's just truly transformative.
56:56
Self-driving cars, that has had some bad press, right?
57:02
Jim O'Shaughnessy: A little bit. A little bit.
57:07
Julia Bonafede: It needs to be more widespread.
57:07
But there's so much noise out there and you have science fiction and there's a lot of ...
57:15
Some of the research I just read, something about the marriage of the soldier and the artificial networks and how much more powerful that'll make their decision making, which is great on the ground in a combat situation, but in an adversarial situation that maybe doesn't sound so great to the general population, right?
57:33
Jim O'Shaughnessy: Right.
57:35
Julia Bonafede: And there are some entities out there globally that are speaking about how the ability to chip a human and how we're going to be some sort of cyborg is ...
57:46
that's not positive, And there are limits to, I think, the responsible use of technology, but everybody needs to be responsible.
58:00
And so think about what's happened in cryptocurrency, in blockchain where there's been some very spectacular news about fraud.
58:05
And when that happens, the same thing happens in any industry.
58:11
We can talk a lot about medicine and what's been going on.
58:17
And I think any time that you suppress thought or opinion from wherever it comes from, you dilute the positive aspects of how something evolves because not everybody that wanted to have a voice is able to have a voice.
58:34
And if I could contribute anything to the world, it would be why don't we just start having really good discourse?
58:41
And if we don't like opinion, then sit and listen to it anyway because maybe it will inform.
58:49
And I think AI is a product of that because you had one narrative going on for a while of how it works.
58:55
And it's always, as you point out with John le Carré is, those are spy novels so there's lots of intrigue and it's governments against governments and it's everything that we've been had a front row seat to.
59:07
Now from an information standpoint, the amount of information that a human can absorb at any given time is, it's the- Jim O'Shaughnessy: Minuscule. Julia Bonafede: Right. It's minuscule.
59:21
And so I have a parking lot theory that we all have a parking garage in our brain and we have a certain amount of cars that can park in there.
59:27
And to put another one in, one has to come out.
59:27
So that, I think, is where we are.
59:37
This is technology, it's going to evolve.
59:44
I mean, we've had these transformative times that we've just seen and the bringing the internet, the advancement in technology on all levels, think about your career and what you've used, the tools that you've used through that.
59:59
And you just talked about your canvas and hand coding data.
1:00:03
I mean, now, nobody would put it in paper form, those Moody documents, they just wouldn't.
1:00:12
And so now everything's very available.
1:00:12
I think it's also the responsible use of data too that will help it.
1:00:20
There's just a lot of issues everybody's grappling with right now.
1:00:24
Jim O'Shaughnessy: Yeah, I agree with that.
1:00:24
And I thought you expressed it beautifully with the idea of if you suppress thought and opinion, you're diluting your own outcome.
1:00:32
So you're intentionally hobbling the outcome that would be the optimal outcome.
1:00:42
To that, I say amen, which is one of the reasons why I am an investor in Stability AI.
1:00:48
I think that open verse closed responsibility though becomes very important, as we talked about prior to actually recording.
1:00:54
My personal view is that AI, like every other thing we've invented or discovered is a tool. It itself is neutral.
1:01:02
The use that human beings put it to will be either good or bad, determined by that human being and their use of that particular tool.
1:01:12
And like fire, for example, when we discovered fire, and I prefer the Promethean myth of him bringing it down and then getting punished by the gods.
1:01:26
But when we discovered fire, like ...
1:01:26
arguably that's the reason modern humans, us, the homo sapiens, smart, smart human are here because cooked food allowed our brains to massively add on.
1:01:37
And great thing for humans, great thing for our species, but also fire is one of the most destructive things in the world.
1:01:45
And therefore, we develop fire departments, fire alarms, fire doors, fire signals.
1:01:51
And that's the idea of error correction on an ongoing basis.
1:01:58
And I don't want to steal the thunder of our CEO who's going to be announcing these things in several weeks, but we are definitely going to be highly involved in the ethical use cases, in the deep fake detection software.
1:02:13
All of the things that we can prepare for, knowing, of course, that ...
1:02:22
I'm a huge fan of David Deutsch who wrote the book The Beginning of Infinity, which he really ...
1:02:29
he's a quantum physicist in the UK, and he builds such a wonderful scaffolding for how we got to where we got and why.
1:02:42
One of the things that he points out is that one of the easiest traps for we humans to fall in is this idea that we believe we know the future.
1:02:51
We believe that we know what innovations are going to happen. And of course, we don't.
1:02:58
And he puts it in a very clever way in the book.
1:02:58
He's like, "What were physicists in 1900 saying and doing about the internet and deep learning?
1:03:11
Well, they weren't doing or saying anything about it at all because they didn't know about it."
1:03:17
So it's very easy, and especially if you are operating on the fear element of our base programming, it's very easy to say, "Oh no, no, no, no, we have to follow the precautionary principle."
1:03:31
Imagine if we were that frightened of new technologies and when they came out with really great carving knives and we said, "Can't release that knife, manufacturer, unless you can prove that it will never be used to stab a human being," and we can't do that. Reaction? Julia Bonafede: Right.
1:03:56
No, you're exactly right.
1:03:56
To me, the tipping point, at least in the US was in the 70s.
1:04:06
And you think about how societies change, and there's a great book called The Fourth Turning that puts it kind of in- Jim O'Shaughnessy: Yeah, I've read it.
1:04:15
Julia Bonafede: But think about just the freedoms that you had at that point before the advent of ...
1:04:17
well, maybe it was even through the 70s and 80s, the advent of 24 news.
1:04:26
I mean, think how dramatically that changed how we lived our lives and the fear that it brought in, because every bad thing that happened in the world was instantaneously delivered.
1:04:40
And you had to watch 24 hour news.
1:04:40
I mean, CNN used to loop the same thing over and over again.
1:04:49
Jim O'Shaughnessy: Headline news, on their headline news substation. Julia Bonafede: Yeah.
1:04:54
And the freedoms that you had to make mistakes or that people had accidents and they died. And it's a bad thing.
1:05:02
It's sad for humans, but taking risks, being able to rise to some idea, to see it to fruition even if it spectacularly fails, you learn something from that cycle or you learn from just evolving through it.
1:05:22
And how many times in the quest or stumbling on fire, or just think about just the modern conveniences that we have today.
1:05:36
It changes the freedom that you have to do something else when you reach that pinnacle of success.
1:05:42
I mean, societally, not personally.
1:05:42
It's just you have these transformative events when humans used to spend all of their time trying to just feed themselves and find shelter, to now, that nobody really thinks of that.
1:05:55
I mean, there's still destitution, which is a real thing, but true poverty in the United States, even at the saddest level is not what it was 200 years ago, 300 years ago, 400 years ago.
1:06:08
So you just have to pull back the lens and look at how we've transformed ourselves.
1:06:15
And AI is going to provide some of those conveniences where people are in repetitive jobs that don't stimulate them.
1:06:23
And it's going to help in education too.
1:06:30
Jim O'Shaughnessy: Very much.
1:06:31
Julia Bonafede: Education is an area where really, I think we're in a rut and that needs to completely change.
1:06:36
So I'm excited for the future, for where we're going.
1:06:45
There's going to be huge transformations in our lives though.
1:06:45
And it's happening more rapidly, I think, than our grandparents.
1:06:50
Jim O'Shaughnessy: Well, this has been a lot of fun because I look at my own thing and I always know when I'm in a flow state with a guest, when an hour and a half goes by and I feel like it's been one minute.
1:07:03
This has been absolutely terrific. I wish you very well.
1:07:03
I think that this is the future of asset management and that it'll come in a little bit at a time.
1:07:10
You are the pioneer here.
1:07:17
You're on the frontier of this.
1:07:17
I am very attracted to it.
1:07:23
Because I sold my company to Franklin Templeton, I personally cannot be part of a company that manages other people's money for another for five years.
1:07:28
I think that's quite fair.
1:07:28
But that also, what's nice about that is that it gives me the freedom to experiment with all of just my own money.
1:07:41
So if I'm a catastrophic failure, I have only myself and my models to blame.
1:07:47
So we always end the podcast with kind of fun question, and that is, we're going to wave a wand and make you the ruler of the world, but only in one way only.
1:07:54
You can't execute anyone or put anyone in a re-education camp.
1:08:03
However, we're going to hand you a magic microphone and you are going to speak two things into it.
1:08:09
And every human on the planet, whenever they wake up, whatever their next day is, is going to think These two things were their own ideas, and more importantly, they're going to start acting on them.
1:08:24
What two things would you incept in the human race to make the world a better place?
1:08:31
Julia Bonafede: I would inject in the morning thinking of everyone, everything, to wake up and to look for the joy in all circumstances.
1:08:41
And then the other piece would be fear not, because we have to take risks to advance, and overcoming fear is the biggest barrier to success.
1:08:55
Jim O'Shaughnessy: No kidding. I love both of those.
1:08:58
Joseph Campbell's great quip that, "The treasure you seek is in the cave you fear to enter."
1:09:04
And so I could not agree more with both of those, especially with the joy, because we also tend to find what we're looking for, so to speak, if we're giving all of our attention to the woes of the world, we're going to find nothing but woe.
1:09:17
And if we're giving all of our attention to, for example, you mentioned poverty, one billion people worldwide emerged from poverty over the last decade.
1:09:28
That's beyond extraordinary as far as I'm concerned.
1:09:34
Julia Bonafede: [inaudible].
1:09:36
Jim O'Shaughnessy: And so, look for better things, listeners.
1:09:38
Well, listen, this has been wonderful.
1:09:38
Where can we find you online, on Twitter, et cetera?
1:09:43
Julia Bonafede: Well, you can find me on LinkedIn.
1:09:44
Jim O'Shaughnessy: LinkedIn.
1:09:46
Julia Bonafede: I spend most of my time there.
1:09:47
Jim O'Shaughnessy: Okay, wonderful.
1:09:51
Well, thank you, and I wish you the very best.
1:09:55
Julia Bonafede: Thank you, Jim. Great conversation.