0:00
where's all the talent in data science and actually it's online it's not necessarily on Wall Street it's all around the world what if you could give away the data but obfuscated so they have no idea what they're predicting on
Search this video
0:16
it's to a machine learning person that description is not strange they can just model the abstract data if I had my way there would be no stock names there would be no tickers there would just be numbers everyone looked at me like I was
0:28
insane the cool thing about nimurai is you can build your model however you want we don't know what model you've got because you haven't given us your model if my model does badly you can destroy my state and if it does well my stake will do well they discovered that taking the average of all people's guesses on a particular thing and found that the average was much closer to the actual weight it worked very well when the
0:59
contestant had to put money in so if you join and you stake a hundred thousand dollars someone else joins and Stakes ten thousand dollars we will give you a 10 times higher weight in the final meta model and if they do lupus it's fine because of portfolio is constrained enough and then they'll get better because hey we just burned a bunch of models foreign [Music] with yet another infinite Loops my guest today I feel so simpatica with them
1:38
you're gonna You're Gonna Learn out why that is uh it is Richard Crabb who is the founder and Lead portfolio manager for numerai which is a fund that is incredibly attractive to me and you'll get it and we'll get into that uh as well he's been named to the Forbes 30 under 30 in 2017 in finance is a ba from Cornell in pure mathematics well done I love that um and you have a very unique hedge fund which is using data scientists from
2:15
around the world to predict Equity prices we're going to get into how that works and why I think it's such a cool uh way to set things up I had a similar idea uh your performance has been ridiculously good especially since you guys are playing on hard mode you are playing with a market neutral fund again doth my tap to you I tried and failed to come up with lots of Market neutral funds when I was still in the asset management game
2:45
uh since 2019 your numerai One Fund is up 49 and I think you've just added like a lot of money in 2023 because you did really really well last year which was a very difficult year to do well in uh Richard welcome I can't wait to dig
3:06
into this well thank you that's a that was a great introduction so so let's let's get started uh I'll play the uh Jeremy Irons character on uh Margin Call and say uh Richard can you explain this to me as if I were a golden
3:25
retriever why don't you uh give an explanation that uh our listeners and Watchers uh can can grasp because you know this has been one of my uh passion projects AI for a long time uh and we sometimes get lost in the jargon so if we could if you could just spell it out for us first off what gave you the idea um and then and then how the fund works yeah well back in the day uh it was uh around 2014 I was working as a Quant uh
3:59
at my first job and using machine learning uh to to model data um and you know every single Quant fund out there is in some form you know Gathering a lot of data modeling it what is that what is modeling mean to a golden retriever well it meetings you know Finding patents in the historical data and the trick is there's a lot of patterns and a lot of them don't repeat themselves most of them don't repeat themselves in the future and so the goal
4:33
of a machine learning model uh is to find patterns that will generalize into the future and in the best Quant funds uh have managed to do that and numerate as unique idea uh which is what I had when I was had my first job was well where's all the talent in data science and actually it's on online uh it's not necessarily on Wall Street it's all around the world and data scientists were meeting up on websites like kaggle or playing in the
5:07
Netflix prize and I just thought and it hit me very hard when I thought about it I was like there needs to be at least one open hedge fund and hopefully just one um where it's like open source it gathers the best people they build amazing software um and and that's what numerate did for hedge funds um and and the reason that uh I'm so simpatico is when I was leaving O'Shaughnessy Asset Management after selling it to Franklin Templeton the my
5:40
passion project was a thing I called gray Swan and it was going to do exactly what you guys have done listen uh except we were my thesis was uh okay so if we accept the definition of the Black Swan that it can't be
5:58
predicted that doesn't mean that we have to accept that the idea that it can't be confer and much like you've done I set up the uh the system almost exactly it's going to be worldwide we were going to invite data scientists in we were going to
6:15
stake them we were going to give them I was going to spin up an LLC that did the trades and they would get money from those particular trades uh but then a good friend of mine who's a genius in AI I I tested the thesis on it and going through it he's nodding he's nodding because one of my other points is look I think we've we've mined the data the existing traditional Quant data as far as we can like it's not a big surprise
6:47
if you're looking at traditional quants like uh osam or aqr or LSB um like our models are very similar because we were using historical data usually the copies that some people have additional data and my dream right
7:04
around this same time that you had it and actually executed against it against it I thank you um was we need to do we need new uh data we need a completely different category of uh person uh and cognitive ability to to make sense of it
7:25
um so I I definitely love the idea by the way he told me at the end of it uh oh by the way Jim do you have about 500 million dollars lying around for the compute that this is going to cost and I'm like um sadly I do not so I ended up investing in a open source AI company uh stability AI instead um but but that leads me to just a ton of questions because I I love the idea I love what you're doing what what gave you the notion to to
7:56
bring open source to Wall Street which is a notoriously closed Source kind of place it is it is um I had my first boss was a was amazing uh and he was uh he helped me with in so many ways but there was one discussion we had that was uh kind of interesting to me which was I had I was quite good at uh decision tree based algorithms uh like XG boost random Forest these types of algorithms and that's what I've been doing back in back in 2013 2014.
8:29
um but I had a friend who was studying his Masters uh and the bet the best person I knew at neural networks at the time so I said to my boss you know I know a guy he can help us he's a student he can help us uh all we need to do is just give him the data and my boss said well if you give him the data you're gonna lose your job uh immediately and you're being breach of all the confidentiality movements in your employment contracts having your
9:01
rep at uh and I said well wait a second he's gonna help the performance like the reason we're doing this to help her performance like what if he runs with the data a lot of the IP is in the data what if he runs off with the data and so that was also a legitimate that was a legitimate thing so the the really Innovative thing Newberry did was what if you could give away the data but obfuscated so you see a feature in the data which
9:30
is some some characteristic of a stop like p e ratio but you don't say you know this feature is the p e ratio uh you just say this is feature nine and then you sort of mess up the feature to a large extent that uh it doesn't even look like it's like a p e ratio um and and then you can share it with anybody and anybody with data science background who might not know anything about value investing or warrant Buffett or piggy ratios or anything
9:59
they can just get going um because you've given them this clean obfuscated data set and there's no stock IDs in the data either so they have no idea what they're predicting on um but that to a machine learning person that description is not strange you have it's Google uh they can translate say from Zoo to uh to French but they might not have on the Google translate team a single Zulu speaker former linguist or anything but they
10:33
have the best translation for that language and that's because they don't care about the underlying data or the domain of of of uh of language they can just uh they can just model the abstract data and and that is the genius here I think
10:54
we we were going to do a very similar thing where we're going to obfuscate the data uh because uh talking to all the data scientists I knew they're like yeah that's why that doesn't make any difference uh and you know I I used to
11:09
make people's heads explode back in like the late 90s and early 2000s what I would suggest that if I had my way there would be no stock names there would be no tickers there would just be numbers and everyone looked at me like I was insane yeah it is it's counterintuitive to take uh take it out of the the world that it's in and move it into an abstract space and and and that's one of the the challenges is that you know semantically
11:39
we are so tied to language that uh you really with people's perceptions who aren't data scientists right uh because you've seen the idiocy of what happened like it might be happening now with AI but uh back during the.com anyone any company no matter what they did I think the classic example was a company on Long Island that made iced tea and they changed their name to longislandicet.com and they went up like 100 percent
12:08
it's happened in the blockchain VM as well yeah energy companies to blockchain companies yeah um so so walk me through the process okay so they get the uh uh data that is obscured um and and then what's the next step I know that you have more than 5 000 models um and they get fed into your main model but but take me through how do you disambiguinate um you know models that as you mentioned at the opening uh you know patterns that
12:46
existed once and are no longer existing how do you how do you make the distinction between ones that will persist versus those that will dissipate yeah um well first you know the data we give we definitely have to analyze like we
13:03
have quants and people who understand Finance who are building the data so that the people who model it don't have to and I've heard it from one of uh my investors Howard Howard Morgan who was early at Renaissance uh he said you know
13:20
that's that's actually quite that is kind of how uh some corn funds work and Renaissance probably one of them uh where you have a separation of like this the data team who maybe know a few things and then there's the modeling team that deal with more abstract data um so that's sort of the the setup when you download the data the cool thing about numera is you you can build your model however you want you can take
13:50
our data share it with other people do whatever but you could also choose any Computing environment to use so you might spin up a really expensive AWS server that trains your model or you might use your own custom language or library or training you don't have to use pipe then you have to use it you could do anything and then you start submitting to us uh your predictions now when you submitted predictions another interesting thing is
14:18
we need to know what model you've got because you haven't given us your model um you keep your model wherever you want so we don't know how you trained your model but we do know you used our data and that's cool for the users and sort of the open source Spirit where you can keep the IP uh that yourself you're not handing over the IP to anybody and that's one of the reasons we we grew um but for sure there's plenty of users
14:46
who uh who who sign up and then get confused I mean they expected it to be like a cont thing and they find out there's nothing nothing for them to do uh they they it's just for machine learning people um and very good ones um
15:03
so yeah getting to uh to you know how to how do they get how do they not overfit first thing is you know it's a little bit our responsibility we could put in very bad features for example features that on point in time maybe the earnings
15:18
announcement to one day lag and you give it give it at the wrong day and then suddenly everybody thinks that's a great feature but it's really just to look ahead bias um but then all the extensive techniques in machine learning like what part of machine learning is you're going to use a non-linear model that might overfit if you don't train it carefully and so you could so a lot of the users do things like cross validation different types of
15:42
tests uh in the in in the historical data to prove that they can trade a model on you know everything before 2010 and it works for 10 years out of sample and that's a very good clue that you're not overfitting and and so tell me about the process uh first off actually uh what has been the biggest impediment to open source like regulations how how at the out of your uh Masters at the SEC uh if you are registered uh how how have they
16:20
responded to this uh we are registered investment advisor um I probably shouldn't spend you know all that much time speaking about the regulations uh in the U.S uh because they do change and uh uh but what what we're doing is really about
16:38
uh getting people to model abstract data no one is giving us trading ideas that are modeling our data and giving us signals we can kind of see all of our contributors as data vendors and they're not uh like we have not like
16:54
we have different portfolio managers all around the world we just have different data vendors giving us signals um and so that abstraction also means you know they're not for example providing US Financial advice or something like that they're actually modeling our data we're in charge of what we end up doing with the signals that they they provide us okay excellent so walk us through that end you you've you've got all of these
17:24
data scientists sending you their signals what what happens within numeri uh and what what do the what does the team there do to to make it a tradable uh event well yeah it uh all this all I've all that I've said you know uh
17:45
worked for a while we got all these users they were building data they were submitting models and actually it wasn't working on live data we couldn't trust the models because here's what you do if you're smart and a little bit uh naughty
18:02
uh you download the data and then you make a thousand models and then you make a thousand accounts on numeri one of them is gonna work and win some money uh so Uber I we had to solve that problem and the way we solved that problem is the way Millennium the big multi-strate solve the problem you get it skid in the game you get the user to put up some capital um and and the the what they're what they're doing when they're doing that
18:35
and call it staking um they're saying if my model does badly you can destroy my state and if it does well my stake will do well so that limits the attack surface it's like the only thing you can give us that will do well for you is a real model that works on live data and ever since we came up with staking which uses our own cryptocurrency for staking um things got a lot better very quickly you know no one was making a thousand pounds
19:09
and hoping to get lucky because they would have to stake all of them and they would lose on more than half of them and then they wouldn't uh they wouldn't do that attack so once we got that right uh things got a lot better and the way we end up combining all the models today is by taking the state weighted average so if you join and you stake a hundred thousand dollars someone else joins and Stakes ten thousand dollars
19:37
your model we will assume your models better because you're staking more on it you believe it will generalize more and we will give you a 10 times higher weight in the final meta model uh than the than the next guy and uh when we combine all these uh with the stake weighted average that final signal is called The Meta model signal and that's just a signal it's not a set of trade recommendations we then use our own internal Optimizer to say
20:07
okay how can we turn that signal into portfolio that's neutral to every factor in the bar risk model and other risk models we've found along the way and within only expresses the core Alpha of the model um and and that is why you know if a
20:27
factor does badly numerize unlikely um to do badly because we're we're running neutral to uh to all the factors that's absolutely brilliant because you're bringing together uh skin in the game which has been repeatedly
20:45
demonstrated to be a better um you know when when uh they discovered that you know taking the average of all people's guesses on a particular thing I think it was the weight of a cow that got it going uh and found that the
21:02
average was much closer to the actual weight and in fact often closer than the winning single entry uh one of the things that I was interested in when I was reading about it was it worked very well when you when the contestant had to
21:19
put money in uh whereas if it was free and they could just take a guess it did not work nearly as well so I I think that's quite brilliant how how do you then transfer so let's say let's make it me and I I come up with through O'Shaughnessy Ventures I've got a team here that is doing exactly this and I think ah you know what I'm just gonna send it over to newerai uh and we stake it for uh the hundred thousand that you mentioned
21:52
um and and it works how how do I receive a payment from that so if it works the way the way we score you is based on how good your model is at predicting residual returns so we don't want to pay you if you bet on the
22:07
factor the factor does well we consider that luck and you get nothing sorry but if you can come up with an alpha that's a proper Alpha that's like neutral to all these things and then that's really good so you might submit a five five
22:23
thousand predictions to us and what tends to be quite good is about a three percent correlation with the target so it's very low right but three percent of the stock market is very uh monetizable um so we evaluate your signal it takes about a month for us to know how well you did we evaluate you over a month of Horizon and then we say well look you you have a three percent correlation uh with the target so we're gonna grow your stake by
22:53
by three percent so you're taking a hundred thousand dollars you'll now have a hundred and three thousand dollars but it will be all in our cryptocurrency got it so um what is the output uh and first off so many questions uh like have you ever uh seen the output from The Meta model and just scratched your headed bot oh my God I I it's gonna be really hard for me to pull the trigger on on this trade or or or are you just so comfortable with
23:25
the way you're doing it that like my advocating you know it should just be numbers and not tickers that you you don't even notice we don't really at this stage we feel very confident with the process because it's been going a while it's working um the also we're buying a lot of stocks I think our core fund numerator one has um thousand positions about 500 long 500 short so the fact that it's so many positions
23:55
um they're all quite small weights and there's no fat exposure kind of makes you feel like okay well ex you could tell what this type of portfolio will do you can get an expectation of the variance of volatility of that portfolio and then you do I just kind of trust the trust the process um because you know if if thousands of data scientists you know are staking these models and they together believe after modeling the data
24:29
for years that it's going to work kind of makes me sleep well at night uh that they are have a lot to lose if they provide us bad signals and if they do lose it's fine because a portfolio is constrained enough that it won't be
24:45
shouldn't be too bad might be someday at times bad but shouldn't be too bad and then they'll get better because hey we just burned a bunch of models you'll if you have that negative five percent correlation will burn five thousand
24:59
dollars from your hundred thousand dollar stake and then you have a lower weight in The Meta model so it's kind of constantly improving it's we we don't we think about it as like the worst thing that can happen with a quad fund is strategy Decay you know great for two years and then it's flat forever afterwards and newerai in some ways is designed to be immune to that because we're always adding data and we're always adding models and new
25:27
models are coming in and out all the time trying to to update their models with the latest developments in machine learning or the latest data that numerized providing and uh is there a size limit to the amount of money through that you raised
25:44
through LPS uh that uh that you have within the fund yeah um there is and it's it's uh it's somewhat other than Target because you know a few years ago when we had a small data set only a few users we felt like
25:59
we couldn't do that much Capital against the strategy but now we're up to about 350 million in in AUM and uh and you know we think it can do quite quite a lot more uh maybe up to up to two billion for right now by the time we get there we don't end up maybe we'll have way more users way more and start thinking we can do three billion but uh the reason it's it's a constrained is also that we use a we use leverage right
26:30
this is a Quant fund we we use a lot of Leverage um on these sort of low risk portfolios and um and that leverage you know is multiplied by the nav so it's it's much we already have much more than a billion dollars of
26:47
positions just with three billion 300 million in assets and uh leverage I I've always been uh wary of too much leverage uh how do you make your leverage decisions is there a formula for that as well there it is it's related to how good we
27:09
are at um at the risk management piece um you know they will be think Millennium strategies or strategies that Renaissance with 10 times leverage or or more but it sounds a little bit more scary that it that it is um because of the fact is um you don't want to leverage things that uh that are high volatility but if something is low volatility uh you you're much safer leveraging it so when most people think of Leverage The well
27:44
you know what I think the S P 500 is going to go up so I'm going to put uh 2x leverage on the s p 500. and I say well wait a second the sap has quite a few 50 percent or 45 drawdowns and in those events you're going to lose everything because the volatility of the s p is quite High but if you can do something else like what if you were long apple and short Microsoft now these are both us companies they're both exposed to Tech risk they're both
28:19
exposed to AI they're both they sort of have some similarity so maybe that portfolio you can leverage a little bit um and uh and so that's how how it's thought about and uh what's been the reaction of uh you know the various groups The Usual Suspects uh for Market neutral funds you know institutions endowments Etc uh wealthy individuals family offices is is there one type of potential client who Embraces this more enthusiastically
28:54
if you will than others uh it's a good question I mean we had a very interesting start where nibira became quite famous because it's very unusual an idea that and we were using machine learning crowdsourcing even cryptocurrency and it's like is this even real how is this how does this work so a lot of people wanted to meet us in the early days like who can be anyone talk to the top allocators in the US but no one wanted to invest
29:30
uh because we didn't have a track record um and that is a little bit how the industry works well a lot how it works and uh the Venture of backers we had because we had VCS invest in us like Union Square Ventures they loved it because they're like hey this is gonna be in the long term this would be quite good uh but the LPS are like well let's see a few years so the story was you know after one year or so We Off Track we started to get uh quite
30:03
uh close with a with a Canadian pension fund uh to invest and they are considered you know very important Quant allocated as an allocators generally because they actually have a lot of savings over there in Canada and they're sophisticated
30:18
um so they looked at it and they said well let's try it so they came in uh for 20 billion dollars which seemed like a huge amount at the time because the fund was only about 20 million dollars and uh they took us to
30:33
that investment took us to 40 million dollars and was our First Institutional Investor and since then uh although it's continued and uh we then ended up meeting with endowments uh you know and and and uh a couple of other
30:52
European funder funds and now we are up to 350 million uh with uh with the ivy league endowment with uh with Canadian pension fund and some of the top allocators you have for Quant but it did take many years you could have a Quant
31:09
that just hated so you could have an allocator that just hated Quant altogether I don't like anything Black Box okay then you could find an allocator that logged Black Box strategies but they hated crypto and the week we were paying our users in crypto they didn't like that and uh and then you have someone who's really likes Quant but they don't even they don't like machine learning and so we had to kind of get three for three uh of these
31:37
things and it was it was very challenging um but but I'm glad that the industry you know kept up with us and looks at our performance and tells us uh they like it and I think we we have a long way to go talk a bit about you mentioned you have
31:54
your own cryptocurrency how does that work yeah well we started to actually pay our users in um with PayPal and uh we're trying to pay one of our contributors who just won three hundred dollars or something like
32:08
that and he said he was in India and that they didn't have PayPal and at this point there was no consideration that numeri would ever use cryptocurrency for payments but this user said could you please just send me Bitcoin uh and this was a 400 Bitcoin at the time and uh we're like sure it's very easy for us to send you Bitcoin if that's better for you and that's how sort of got started with playing in crypto but
32:38
the reason to make our own cryptocurrency was to do the staking I mean not many people had even heard about staking in 2017 when we started doing it um and it's the exact right technology for us for this because if you ask a user please wire me a hundred thousand dollars and I promise I won't uh steal it that's not as good as a blockchain system that says set you know stake your stake your model and uh and and like we can't touch it
33:13
because it's it's part of the stake uh and and so it really was the right thing to do the other thing is burning we want to burn your stake you can't 30 US Dollars like what are we gonna do send the user video of us
33:27
burning their money would be like you lost it doesn't make sense but burning is a natural thing in the blockchain you take some of the cryptocurrency you send it to the zero address then it's gone it's out of circulation they can tell we
33:39
didn't take it from them it we're not benefiting from anyone losing money on nimurai it just goes straight to the zero address um so it was the I thought it was the perfect use case for for crypto and we also had a very technical user base we could get their heads around all this stuff so we launched NMR in 2017. and uh and that's what we're using ever since and and now uh you you have maybe changed your stance on crypto in general
34:10
as an investment right you sold your ethereum and and I read some other stuff uh any any reason for that yeah I do uh it's a yeah dude I do hate a lot of uh crypto stuff um it bothers me uh I mean I don't like for example um I think it's quite good that maybe they're way fewer icos where you you sell cryptocurrency to the public because I do think uh there's a sense in which public might hope that the thing goes up after they
34:47
buy it uh and numerai when we created NMR we actually gave away the NMR for free to the users on the platform and that was the right thing to do because we wanted them to use it we did it really care how valuable NMR was we just wanted people to stake it for like the purposes it was designed for and I think crypto has did become quite a you know just like uh it's kind of sad when someone makes a lot of money in the.com
35:19
bubble and then loses it all uh and uses leverage or now with crypto had it all on FTX or had it all in Luna and so those types of things are quite disappointing um and I I Still You Know believe in crypto in the sense that I think it's
35:37
really easy way to pay people really good way to do staking uh but uh the other stuff is quite bad and then from the investment perspective it used to be this really uncar related weird thing uh when I bought Assyrian
35:53
back when it started and uh it was like the straight the strangest debt I had in my portfolio you know is what this will do and then it actually became like very correlated with uh with just a NASDAQ or the the emotional you know dysregulated crypto traded Traders uh so that made it less attractive as an investment um Switching gears many of our listeners probably would when they're hearing what you're doing the the first group that
36:27
might come to mind would be uh quantopian uh which ended up failing um tell me how you you're very different than they are at first the techniques that they were trying to implement yeah I mean so I like gwenzopian and I think many many people used it and tried it out um the biggest thing with them is they basically started the company before machine learning like they started around 2011 I think and uh they things weren't really clear about
37:03
machine learning being very important and a very good good way to to to model so what they did was build a community of quants that actually wanted to be able to write a hand-coded trading strategy that said I'm gonna buy Apple this day and then I'm going to be short Tech and then I'm gonna it's just like a rules-based algorithm not a learning algorithm um and you could run a back test and see wow that that rule I made worked really
37:28
well but that wasn't that wasn't like a deep result or a result that was in gonna generalize that you just happened to choose the growth factor for a few years and the growth factor by itself at a sharp of two so you seem like you have something so numero really took a much more mathematical Technical Machine learning approach and the final thing is quantopia never had staking they were also started a little bit pre-blockchain
37:56
uh and so they didn't have staking they didn't have skin in the game it didn't have machine learning and that's Together made it very hard for them to succeed and and now tell me how you you would go about improving your process uh in terms
38:15
of uh the research you're doing Etc what does that look like the biggest piece of research um the two is uh is data and optimization so data is the input to all the models if we give out bad data it's just simply
38:35
won't work there won't be any way they could be made to work um no matter how good the algorithm was so we spent a lot of time trying to figure solve the problem of well the way other Quant funds do data is they have a lot of different teams buying up different data sets and analyzing them nimurai said well we want to be way more Capital efficient well let's build something called what if you call the feature Assessor which takes in
39:05
arbitrary data that we buy um and just totally it's just really disentangles whether it's going to uh whether it's going to work and so if all we do this this year and next year and the year after that is if we double the data set and we double the number of models those two simultaneous doublings are are all we need to keep doing um to to get to be to be to be the best so that's what we try to do every year kind of for the last three years just
39:39
like we have a system that works from the incentive design standpoint let's make sure we grow the data and the people talents um the final word is optimization this very it's a very subtle thing maybe you you know maybe if you take your feature exposure down to zero you're spending a little bit too much trading costs just on the rebalancing trades could you maybe soften that without taking on more risk and there's so many
40:10
considerations with bats as well but uh but the core is uh more data more users and tell me what types of data you buy so we do have quite a skew in the data set towards data that's long I think you're right to say that things like the
40:31
basic Factor models uh of the past might be kind of in trouble for the for the future because those are basically just known risks this gigantic ETFs now that expose these risks and maybe you can't uh bet on them as if
40:46
it's the 90s but it doesn't mean you don't want those factors in your data at two because they're very important they do Drive variance in returns even if they're not predictive of returns so we give out a lot of data that goes back very far as long as we can go like usually about two decades um and covers the whole 5000 stock Universe we have and if we can conquer that universe and go as far back as possible that's when machine learning
41:16
works if you told me you had some special um credit card data that you bought from a startup something and it only starts in 2020. that's not going to be a great use case for machine learning you must know from a stability you need a lot of a lot of data to get these things to train so your stock market is kind of a small data problem that we're trying to make it into a big data problem uh by by going as far back as possible
41:49
um do you see this uh ever leaving equities and going into other asset classes we have some very good uh engineers and some of them have warm up in places that trade Futures and and other other currencies or things like that
42:08
um I've actually been quite full on sticking to to equities uh until we get to capacity for equities and I also like equities because of the breadth because of the amounts of data maybe there's only 80 Futures you can really trade
42:25
um uh with size so the you don't have the breadth and you have to come up with a whole new style of doing the forecast uh like directional trading instead of cross-sectional trading so I think it's been good at that we've stuck with equities there's plenty of complexity and equities uh to get your head around do you think that there would be any value in kind of the traditional Quant sense you mentioned and I know much more
42:53
about it now since by association with stability the the need for locked up data as opposed to its small amounts of data do you think that there would be any benefit to you know hiring a project for example uh we were looking uh at
43:09
recreating the Moody's manuals uh back to 1900 uh where right here in the New York City uh Library they have them all um would that be of any benefit do you think uh here with your uh system well I am very interested in large
43:32
language models obviously like everybody else and uh we've all you know everybody there's people at the phone that are that are familiar with it and and can build these models but I do think there might be a way to do it in a way that's
43:46
like kind of profound like yes we've got very good at understanding this sort of uh shorter term details of the market we we have prediction Horizons that are about a month long um and they they maybe aren't that good
44:04
longer than that because they don't really understand the company they understand all the features and the properties of the company but they don't know yeah they haven't integrated the whole of human knowledge uh with respect to
44:18
that company so I've actually been working with a few luberi users and thinking about this for a while and I think there might be a way to to have that take us to the the next level I want nimurai to be the absolute best at
44:34
lom's with respect to the stock market I'm not gonna it's a little bit like maybe with crypto I knew we needed to have our own crypto and do the do the crypto thing I try to focus and stick to the core business but the sometimes something comes along that's like oh this is this is now getting really exciting and we've got to dedicate some resources to it and um I think we can do a really good job turning text-based
45:03
data into incredible new features for the newer users that are very uncorrelated and different to uh what they have right now have you noticed any um inferences that you can make about um the the type of model or the type of
45:21
uh data scientists uh who is um you know uh is it just contingent on the amount of their stake that the attention you pay to them or are there any softer uh non-mathematical uh things that you look for in in the data scientists that
45:42
are that are uh helping you with this yeah it's it's been very hard for us to beat the state weighted meta model we've tried a few things like uh we we looked at you know well if this person's done extremely well for the last six months why not give them a little bit more weight and it always is right before they have a drawdown and uh actually their own assessment of what their weight should be is coming through
46:10
through their stake so you it's very hard to beat the stake weighted meta model um we do meet our data scientists I mean we know a lot of the top ones they come to our conferences or they chat with us on Discord and uh and on our forums so we get to know them and some of them have you know very interesting backgrounds one userset used to be at uh NASA jet propulsion lab and another user at CERN another user works at Nvidia on machine
46:42
learning so you have this amazing set of people but still we we don't have any better way of deciding who gets to be in the meta model aside from staping yeah and and I live I that's actually quite an elegant solution as well uh
46:59
because it it removes all of our biases Etc but uh in our human OS which I found off it is the biggest problem in uh in in in trying to to put these types of things together so uh your overall goal perhaps is you think you
47:19
might be able to get as high as 2 billion in the fun um and and you are in fact looking at other uh uh asset classes that you could potentially um uh find models that work with uh you know you had a quote which I thought was great and I want to ask you about it uh basically your master plan is to monopolize uh intelligence monopolize data and then uh money and then share it uh I'm getting that wrong maybe you give me the actual the actual
47:57
quote but but why the open source uh the identity for the open source well to to monopolize uh so that is our master plan monopolize money monopolize data um sorry monopolize intelligence first monopolize data monopolize money and then decentralize the Monopoly uh and so to monopolize intelligence that's sort of talking about the Talent how can we be really quite definitively uh the the hedge fund was by far the most Talent
48:35
and you might say well the way you do that is you have to raise maybe three billion dollars uh and somehow uh start hiring all the best talent in the world one by one and putting them inside an office but the the way we thought about it was
48:51
if you have an open hedge fund that pays a while and it's fair uh then that is the much better way to do this um and so that's that monopolize intelligence in some way the the goal we've got the furthest towards
49:07
um because I do think we are going to keep growing and and adding new talented users the second one monopolize um data you know I'd I would actually suspect uh two Sigma with all the the money they had and time have more data than us they they probably do um and that can't continue we have to we have to catch catch up eventually but um we're doing a lot per unit data that we have and uh and we'll be building this feature assessor that
49:42
can onboard data as fast as possible um but you know which with some with the data plus the intelligence you start building this sort of monopolized money plan which is if you have the best performance people will come uh and you
49:58
know for the last few years there are a number of very high profile you know machine learning funds or Quant funds that we've really definitively done quite a lot better than and so uh slowly allocators uh re-weight into us
50:15
if we can keep that out and the decentralized one final point of the master plan is just like you know if you did have a system like this which had all the data all the talent and was like an AI super intelligence allocated the capital around the world wherever it's needed and most effective then it's like a it'd be nice if somehow um people would own a share of that and uh that lost pot is being hard to do and I haven't really
50:49
um I don't know the right vehicle I've often thought well what if numer I had an ETF like Arc except with all the numerite old things and it's actually tricky because you can't do an ETF with Longs and shorts and leverage maybe we
51:02
can have a holding company or a insurance company one day but uh these are also quite tricky and expensive so but try to focus on the near term getting to that two billion public fund I I love the way you break it down what
51:18
could go wrong um I mean yeah it is this there used to be a lot of answers to their questions and then now to her uh you know we I do I do think the U.S it's it's quite hard to read whether the US wants uh the this type of company I mean I love America I immigrated to America because it's my dream to start a company here and um there is a lot of discussion about um yeah more regulations that that make it much harder to do these businesses
52:03
that it used to used to be in the past I look back and say uh Ken grissen is starting a funded in Harvard uh to home room and being successful and you think of how huge the bar it is to entry are to play this game for real
52:19
um nowadays and uh I think that's quite sad uh and memorize in some way low as those batteries right because suddenly you don't need to go and spend a million dollars a year on data you can use numerize data and it's free and it's
52:36
clean it's ready and so you know you only need one dollar to stake on nibirai but you need 10 million dollars to set up a proper Prime brokerage account and it's like this isn't that good I do believe in markets and I do believe we want to have the US always have the most efficient markets and have bright people's intelligence applied to them uh 100 simpatico on that I I love this in so many ways and and that is one of the
53:08
ways that I really love this the the ability to empower very very smart people should not have to have that 10 billion or 10 million uh where they can actually apply their intelligence immediately I just I I am so
53:25
uh it's supportive of that uh as well as the outcomes how do you deal are are these are are they formally uh uh bound to you through a contract are they uh what what are what are your data scientists uh within the organization yeah they they
53:46
they agreed to our service on the websites um and those are bikes uh specific in certain areas but also very free like it's I think it says explicitly yeah if you do submit predictions to us uh they become our predictions but uh the model you use to train you know that's yours you you are on the IP and retain the IP for that um and so yeah there isn't uh there isn't too much where the whole point is to bring the friction down uh for all for
54:21
contributions uh to the to the system and I would say that some some startups I'll say Robin Hood I don't like Robin Hood but uh I'll just I'll just talk about Ramen that it's like you can't say you're democratizing access to the market and actually you're not you're actually making it's very easy for ordinary people to lose money um because you know they don't have an information Edge and you know you're selling them options
54:54
with like 100 annualized volatility which will be zeros uh in the next drawdown which is inevitable um so I think numerized a little bit more like less from oxidized access to the market but in a way that's like honest about how hard it is to participate in the Marcus like you have to have a really good data set and you have to be really good at data science uh to help our Quantum fund so but but I but your contributions are actually
55:26
being used to power or or hedge fund have any of your data scientists uh uh tried their own uh startup uh basically as you mentioned they keep the IP of their model um have any of them said thank you this has been great but I'm I'm getting
55:46
funded by seanice Ventures and and that's where we're gonna go yeah the well the important thing is once they've trained a model for sure they've learned about data science and it loaded about how to do solve our problem
56:00
but their model obviously won't work on other data it's designed for our data they don't even know what the model's predicting what the features are but still it's possible they could become very skilled and then say you know what
56:12
um I'm done with liberal I'm gonna go work at a big hedge fund um uh but for most of our users and the types of uses we have it's typically kind of like they're working at meta uh or apple or it's machine learning people
56:27
they don't want jobs on Wall Street it's the last thing they want they want to be able to have the flexibility and freedom to come to numerize whatever they want uh and so I think it's not really a risk but I do think people are learning things and then right that they're applying elsewhere and I think that's great yeah and that was a big part of I mentioned the idea for gray Swan that was that was the model uh it was we we
56:56
don't want you to work for us enjoy your job wherever it happens to be this is gonna be something that you can do or not do whenever you want to do it you can you can use uh the resources for other things you can get other insights
57:12
uh and you know my I I have some old school lawyers who like literally almost have aneurysms when they were looking at at what what I was doing um but um in terms of IP for example you obviously own all of the IP here right
57:33
well we don't audit any of the iPad used to develop the model um if you trained a neural network with a special Transformer architecture that you've read in a new paper that's all all that code is yours you never send it to us right you send to us as the predictions um so in some ways it's more like you're a little IP but we're licensing some of the predictions yeah uh the other thing that I love about this is in many ways if you
58:02
abstract it even further um it it becomes this idea of you know sort of smart crowds uh which I'm very enamored of and that that also changes the very nature of traditional capitalism and traditional markets uh do you see any extensions of this type of process that you've developed for specifically for a market neutral Equity long short uh if I was going to ask you to speculate uh well where would you see um this um you know what are the nearest
58:42
neighbors or this type of process could uh could take off I don't it's very I think it's gonna be would be very important to do it in another industry uh versus Finance like if you had people people have said to me look if your approach is the best way to do machine learning because you have so many different models and they're all ensembled together why wouldn't someone want to put like Health Care data or something and model
59:10
model that data and get get the best results but I think many problems in machine learning are basically like sold like if you want to make a model that predicts uh that create generates images uh their way is to do that and they're like extremely good and uh if you want to make the model that does like face detection or something to its images it's typically used like a convolutional neural network and then that ends up being like state
59:42
of the art uh with just one engineer working on it um so for most problems they actually are kind of sore because they're working in a space where the data is stationary like that's the difference with the stock market it literally is changing all the time because of Market participants and new data sets and new information coming into it so that's why it's especially good for this um and I don't think you should you
1:00:11
could benefit that much from doing crowdsourcing on other data um so uh you know uh if I ask you to tell me you know you already have uh quite well articulated the grand plan uh but um other than Regulators screwing things up uh is there any anything within the process that you could see for example blowing up um yeah I mean there's always always it is possible you know for Quant funds to have bugs right I mean I don't think uh
1:00:57
two Sigma marketing department or other funds would say you met all that Justice guys we found a bug today in our trading system um but it obviously happening it's not like immune from all the problems of software and bugs so you know there
1:01:14
there once was a time uh where we released a handful of features that uh were we shouldn't have released they were they were dangerous features um now it's possible that some of our users might have figured out that they were
1:01:33
dangerous or strange in some way they looked really good and they stopped working and they could have figured that out and maybe taken them out of their models but our kind of promise to our users is that we're doing that type of
1:01:42
test and they can just go ahead and model it and we ended up finding this out and we're like whoa how did this this get in and we found a kind of a bug um and we wrote a forum post about it telling all our users these features are
1:01:57
kind of dangerous they won't be in the future data releases and uh and they might have heard performance uh for a little while not a lot but a tiny bit I mean these are three features at two thousand it's not like catastrophic um but it is a it's such a high Precision uh game Corn finance that you know it's sometimes when an engineer joins and they're used to sort of hacking on a website we're like okay no this is a much this
1:02:28
is a very different game do you realize when you push that code it immediately influenced 50 million dollars of Trades and this is like a science and sub more and more are the people we have are experienced uh in this environment you have one engineer who talks about how at his old job in the UK as asset manager uh there's a I think there's a special term in law for it or something where he can basically go to jail if he makes a
1:02:54
mistake it's not about like it's okay to make mistakes it's an environment that's like this is a very uh this is a danger paid job you're an engineer but this is a danger a job and and you can't make mistakes so it makes for a very different environment to end up becoming a very slow hedge fund if you start caring at that level um but these these types of things do happen so I think that's always one thing uh to worry about and then um
1:03:25
the market you know although we we're modeling the market and the risk in the market we don't know what it will do right every day we don't know uh and what you want is your volatility estimates to be the same in the future as they were as your estimates are saying and but if this even slightly different uh suddenly you're like well the bad test we could have uh a drawdown of this magnitude but in live without having to draw down a
1:03:53
little bit bigger why did that happen and you you might not even be able to diagnose that um but yeah the the fact of running so much uh so much concentration uh you know even if we're not back to neutral we're still not immediate if another covered what coveted 19 would have happened that's actually a really tough time in our back tests and with many Quant funds you know we don't go down nearly as much as the market but uh we do go down and
1:04:23
everything we look at we can't we don't know how to hedge out that risk so if that happens again that'll simply just happen to us again uh because not every risk is hedgeable what's been your biggest drawdown with the live money um we are having it now how exciting uh you heard it first yeah um yeah what yeah one one reason is we uh We've it's actually looks competitive with uh with the covid-19 drawdown um one of the reason is the fact that we
1:05:04
increased volatility on purpose right we we wanted to take on um more risk now that we're comfortable with the with the system so we've sort of in the past we've been running on sometimes as low as seven percent
1:05:17
volatility extremely boring fund grew it to 11 ball and then when we release new Optimizer settings which just came out uh after the drawdown I should say uh they they they they're gonna take our uh forward Vault to you something more like
1:05:34
15 um so in that in that type of volatility you can expect one or two months to be quite large in in drawdown and and what specifically was the percentage drawdown that you're having now I'll tell you because it's going to be live on all websites uh in a few days but basically we went from yeah we're this month of May down about 10 percent uh wow you you got me ready for a much bigger number than ten percent it's like us secretly
1:06:10
no I know it's 10 sounds pretty chill if you've ever invested in uh venture or crypto or it doesn't bother you at all no that's like uh not even an hour but but for us it's uh they've never had a month that big so uh interesting uh hopefully it's as rare as as the past but we we never know well Richard this has been absolutely delightful I loved the fact um that you are uh doing what you're doing I'm a little jealous I will admit
1:06:46
because that was going to be my Prime vertical in O'Shaughnessy Adventures um you had a good idea yeah well I had a good idea but you are implementing a good idea and there's a big difference between having a good idea and implementing one uh because uh a good idea without any action around it is just a Daydream as opposed to an actionable item uh like you've achieved uh this is absolutely fascinating I will watch uh your progress with great
1:07:19
interest um at the end of our podcast uh we we do this with all of our guests and and so I'm gonna do it with you um and that is uh uh we say that we're gonna make you the emperor of the world for a day you can't kill anyone if you
1:07:36
can't put anyone in a re-education camp but what you can do is we're gonna add you a magic microphone and you can speak two things into it and what's going to happen is the entire population of the Earth is going to wake
1:07:52
up whenever their next day is and they're going to think I just had two of the most brilliant ideas and I'm going to act on them what two things are you going to incept in the world's population this is super cheesy and lamb the first one uh just like um life is long like like knowing knowing that life is long and the repercussions of that like be so be help be healthy take care of yourself type of thing um because that's one that's very easy
1:08:31
to get for people who are hard workers um and and also there's the life is short meme which I think is dangerous for an investor um and then there's uh AI is not going to kill us all on that one I absolutely 100 agree uh I I I uh am usually quite patient um but watching some of these uh the hysterics um it's very difficult for me it's wild uh yeah because it's it just it's all the worst parts of human OS bubbling to
1:09:23
the top uh yeah have you heard of the cognitive distortion to testropizing yeah yes I am and uh I'm I'm watching many many people who hey the one that really kills me or the few that really kill me and I won't name names but there are
1:09:41
some people who I had tremendous amount of intellectual respect for um that I have seen that intellectual respect for those people get chipped away rather brutally I know I know but that but as you say life is long so so
1:10:01
uh we will we will watch with interest tell us where we can find where our listeners and viewers can find but you uh you can find me on Twitter richardgrade c-r-a-ib and uh and at me Mirai and uh you can take a look at numero.ai terrific well Richard I wish you the absolute best I'm as I said a little jealous I think uh it's a fantastic idea and I wish you the greatest of success and thank you for coming on infinite Loops thank you