E662 | Damian Cristian & Guy Conway, Rule 30: Building the First Fully Systematic VC

0:00Welcome back everyone to another episode of the UVC podcast. Today we're diving into a topic that's sparking intense debate in our community, quantitative venture capital. What happens when data, AI and probabilistic reasoning meets the art of early stage investing? And how will this reshape the craft and the careers of VCs across Europe? To unpack that, we're joined by the team at Rule 30, an AI research lab building the world's first fully systematic venture strategy. I have to say, today, we're not just going to talk about Rule 30 because it is an incredibly important topic that everyone has a perspective on.

0:36So for that reason, we'll dive into, of course, the very interesting model behind World 30 and how you work as a VC. But of course, we're also going to dive into the broader topic.

0:48Tear down this wall. It's more than just an alliance. This is a union of values. Let's start acting. From first time to seasoned investors, the EUVC Syndicate is not just about capital. Get involved, co-lead deals, share insights, bring in great people, and help shape the pipeline with us. Visit eu.vc forward slash syndicate to learn more. This show is not investment advice, and the hosts of this episode may be invested in the funds and companies featured. And let's start on the broader topic. But before we go there, to both of you, say hi to everyone. Say who you are. Hi guys, I'm Damien. I'm one of the founders of Rule 30 and my focus is mainly on the technology side and obviously on the fundraising too.

1:39My background, I'm a mathematician. I'm from Romania, originally moved to London about 15 years ago. And about nine years ago, Guy and I set up a venture builder together. That's how we met. And yeah, through that, we stumbled across the problem. raising capital for the company's intervention builder made us think that the way we see, especially at the early stage, take decision is pretty random. So we looked a bit deeper into how we can systematize this and this is how RoelTucky was born. Well, that's a frustration every founder knows. Hi, I'm Guy, the founder at Roel30. Yeah, background in engineering education and some working in management consulting and private equity.

2:24I got introduced to Damien eight years ago now. Started up a venture builder, so the work he was doing was far more exciting and inspiring than my day spent in tabs of spreadsheets. Yeah, and we've worked together for the last eight years. As Damien said, it was the experience with the venture builder that gave us the insight into venture capital. Yeah, what we could see is that everyone was talking about how startups operate on this power or distribution. Not many people were talking about how venture capital funds operate, on that same power or distribution of what we could see from the successful funds that very few of them could be successful two times in a row and almost nobody, maybe benchmark, maybe Sequoia, but almost nobody could be successful three times in a row.

3:03So the excitement around what you could do with a systematic strategy to more consistently hit those returns was too much to turn away. And we ended up pivoting the business back in end of 2020. And it's been a roll 30 ever since. Yeah. And it's, of course, the dirty little secret of venture, which most GPs do not like saying to LPs, which is that VCs very much, as you said, operate on a power law. It's not just the founders. It's unfortunately also the venture firms. There are many reasons for that that we can dive into today. We'll probably only dive into it from the decision-making perspective, from the smart use of data perspective.

3:40I think there are other systematic strivers of it as well. But let's not dive into that now. Let me open this conversation up with the very first question of how you think we best define quantitative venture capital. And then from that, you can go into talking about rule 30 versus traditional VCs. If you think of venture capital, there are a few elements that can be used as levers for firms to basically do better investments and manage DPI. but if you look across that so you definitely have the sourcing part you have the decision part so what what you invest in and then you have the portfolio construction part in the big areas and obviously later down the stage is the support that you give to founders on our side and what we we how we are looking at it a true quantitative venture capital needs to have solutions for every decision across the first part up until selection so starting with the portfolio strategy that has to be packed so the portfolio strategy has to be packed to your advantages and your operational capacity and what you bring to the table the decision making per se is the most important because even though there's an obsession in venture capital with access like with discovery the reality is it plays quite quite differently than people actually assume because if you step back in theory and the easiest example to give is for a later stage one for a series a fund so in theory no series a fund has any form of discovery problem there's tens of thousands of seed stage companies that are publicly available data that they could go for the problem comes at the point of being able to actually analyze the data and know what to focus on, right?

5:29And that part is very, very important. How you analyze, how you take a very big list of companies to a very small one that you can actually focus on. And once you are able to focus on how you take decisions on what you invest in and then further down how you balance your portfolio. So it's the entire decision stack across those parts. Here you said that the CRSA funds should not have a discovery problem. Would you also say that that is true on the earlier stage or is quantitative VC only relevant at this point at Series A onwards? No, it's for everyone else. It's more a VC type of problem. And I think that the focus on what people spend their time on actually and spend their intelligence on.

6:13If you look at Preced, you have quite a host of tools that appeared in the last couple of years, in the last few years on the market, like Harmoni, for example, that basically collect, have suites of scrapers that collect founder signals. Now, these founder signals are basically generating new lists, almost think about the crunch based to point O, basically. This new list, and this can be quite exhaustive. So on our side, we get in the hundreds of these new, like our scraper bring hundreds of new signals every week. But then the problem comes, if you have a small team, what do you focus on out of those hundreds?

6:50And the only things people resort to are pedigree heuristics, what the founder basically was doing before or who knows them in their ecosystem. But that gets people to actually not focus and not be able to analyze the entire ecosystem. So our power and the power of QAN VC is actually not only being able to pull the data in, which is fairly trivial on the scraping side, but then it's that ability from pulling of hundreds of signals to triage down to a list that has a distribution of winners to losers that's a lot bigger than the basically random VC space. The important point to make here, the important differentiation is between, maybe we're going to come to this later, but between quant VC and data-driven VC.

7:41We see a lot of funds talking about data-driven VC, especially in the last three or four years there's been this movement of data-driven VCs. It's in the way we look at it, it's almost hygiene, right? What they're doing in terms of using the available data points to inform better decisions. The challenge that we have with data driven VC is that it's not a decision engine. Okay. So it's, it's helping to see more startups or to make better decisions. But if it's not a decision engine and then you rely on the same IC processes where the IC decides what they want to do, they'll use the data where it agrees with the culture of the push in the room and they won't use the data where it, where it doesn't.

8:19And so, yeah, for us, it's trying to distinguish that POMVC is that we rely on our algorithms to make our decisions. We're not just using data points that we can find available. And let's talk about the elephant in the room, what you've been describing here, that you actually use the data to make the decisions, that you have a decision engine that's driven by AI. Let's unpack that, because I think we all understand. We all know the harmonic platforms today, I think, and the likes of that. And I think that where many would challenge you and someone doing what you're doing is, can you really go further than just making sure that you see the right things?

9:00Can you really go to the decision point? So, yeah, I think that's a very, very common question, as you can know. And then people make this assumption, this automatic assumption that there is not enough data to predict at pre-seed stage. Well, I think the correct phrase there or the correct truth is there is not enough data that can be humanly computed that leads to that. And to give you an example, let's say you're investing in deep tech companies and you have two companies that come and present you with quite nascent technology to solve a big problem. And each of them, very good researchers from good universities and look at the problem differently.

9:43As a human, yes, you could do some due diligence in terms of asking research peers for an opinion for that solution. But the piece of data that exists that the proper quant model can use is like taking all the white papers in the world, building attribution model behind all the researchers, figuring out how the researchers point to this problem. And from that, get quite a rich feature set that you can feed into a machine learning. There's just one. It's a lot more others in terms of how you analyze founders in compared to cohorts and baselines from where they started in time series ways. So we can get quite geeky here, depending how much you want to dive into it.

10:23But there's a huge host of data that is not easily computable by humans that algorithms can take into account. It's just very, very hard to build the data layer and to create it. But we've been doing that for the last five years. So we know for sure there is more than enough data. I think we should be super geeky because we think of our three constituents as audience. We have the VCs to whom this might be incredibly disruptive. We have the founders who are about to be evaluated by computers instead of people. And then we have LPs that are trying to figure out, am I in a debt asset class? OK, so let's then unpack exactly what the decision means.

11:06If you look at the other funds and if you look at the power law, and everyone in VC kind of talks about the power law, but I don't think many fully understand what it means. But what it drives, if you look in the end, you have a universe of companies that are raising capital. That's the universe that you can go for. Any form of selection process, regardless if it's human driven or if it's machine driven, regardless if it's based on childhood trauma and a deep understanding of markets or a mix of it, is in essence a filter on the distribution that you're looking at. So you're saying the easiest example to say, saying I only invest in repeat founders conceptually.

11:43That's the easiest filter to conceptualize. Once you take all the founders, all the companies that you could potentially go for and analyze, you apply this filter with repeat founders. All you're left is with a sub-distribution and you hope there's a better DPI there. Once you have that sub-distribution in the actual VC decision-making, even for funds like a normal fund will have roughly 2.5 % of their investments that will drive roughly 80 % of their debt returns. And that's the action of the power law. Even in brilliant funds, even if you look at funds like Benchmark, for example, that stat will be roughly 4.5 % driving 70 % of the returns.

12:24So, thinking on these success metrics, you're basically understanding that everyone is filtering a distribution and then kind of picking at random from that sub-distribution. That's literally what happens for every form of system, regardless if quant or not. Now, on our side, and the way the system was trained, it was trained by tens of thousands of companies from the past in terms of objective data. So it has no opinion, it has no weights, it's no statistical model in that sense. It's a model that took every yearly cohort from 2010 up until as close as possible we can get to today. Then you have other algorithms and math functions that aim to figure out which one was the top 10 % of the distribution in terms of the delta evaluation from the first round.

13:19Those ones, those 10 % are basically trained. The model is trained on understanding what makes those 10 % winners. The other ones are defined as losers. So it's not about picking unicorns at all. It's about maximizing the DPI at the distribution level in terms of the way you can pick. So we tested it. Since we ended up having a product that we can test, which was roughly a year and a half ago, we run it against many, many portfolios. and we back-tested it against the portfolios of many investors. And absolutely all the data shows so far that we're beating that capacity to filter a better distribution by many times that a maximum human can achieve.

14:10What it also unlocks, the capacity to actually look at scaled distribution because that was what I was intuiting at the beginning, right? So if you have a fund that has three analysts, even if the data exists, even if you bring to any form of tools, any form of data, if you bring 100 companies per month in your pipeline, you almost don't have time to fully analyze them. On our side, a Quant VC can actually bring 10 ,000 companies if they are and analyze them deeply. So you're already starting with a much richer capacity to analyze bigger distribution so that can go downstream in terms of how good you triage from the total available companies.

14:56Can I ask you about, because obviously the data set that you are able to triage on and filter on the back of is everything. And it's both in terms of the training set that you've used and it's in terms of the available data for the founders that you're looking at. Many would say that picking a founder, you need to engage with that individual founder to uncover important traits, dynamics in the team, so on and so forth, which means that you can't do it on the back of a deck and you can't do it purely on the back of public data. So what type of process is needed to run with startups and founders to be able to get to a point where your triaging model works?

15:43Do you have to meet them first and get that first data point from how they act in a meeting, how they interact in a due diligence process, all that stuff? Or can you do it all on the back of what's available publicly? Look, not really. This is, again, a function of the way the human mind and the way the PC works. When you don't have any form of data as a human to go for because your brain is not set up to compute the data that you can get from a deck, then you have to rely on intuition, right? Or like the intuition, how you perceive the founders and all these questions. Now, in essence, I think that's a brilliant, brilliant data source.

16:20Like it's like if you can decode, it's a brilliant data source. But there's also a huge assumption on every VC in the world that there are very good, well-trained psychologists that can actually compute that data, right? And the reality doesn't show it. in the end, regardless if they have this process, and everyone says that, if you look at the fund performance or what they're driving, it doesn't correlate. So in the end, it goes back to something very, very simple. You can analyze in your way, there's many ways of doing venture and many ways of taking any form of decision. We love the story of the early quant hedge funds.

16:59We may have sort of spoke about this before, but people used to analyze companies using their P &L data, their balance sheet data, some of the early, the quant hedge funds, right? They figured out that if they were looking at Walmart for a stock, for example, right? They, they could take satellite images of the car parks of Walmart. They could overlay some weather data to understand if those car parks were full or not in relation to the weather. And then that was a great predictor of what the sales were going to, were going to be for, for the stores. So you go from these trailing edge indicators, these very forward looking indicators.

17:31And you know, it's the same as we look at it, that historically people have always told us in venture that it's a boutique industry, it's an artisan trade, you have to meet a founder, you have to shake their hand, you have to look them in the eye. And that's how you get to good decisions. And you see in the stats, right? Everyone talks about the, well, you can see the performance of funds isn't there. So it's, we're not saying that the traditional ways of doing venture are wrong. There's some funds that do it very well. We're just saying that the advancement in technology and the advancement in data availability over the past 10 years has created this opportunity for new strategies to emerge.

18:05And from what we see from the LP conversations we have, they're crying out for new ideas to tackle venture because people know it's a rich asset class, right? People see the returns that you can make if you can get it right. What is the corollary to taking satellite photos of Walmart car parks in venture then? i it's it's very hard to answer because simply especially at pre-seed you can't have you can't have linear models or decision tree based model based on static variables so it's quite hard to answer a question it's a combination of many things but to to give you an example here yeah what i was saying personality is not a thing like any other thing nothing sits in a vacuum so the personality will permeate to to every decision that the founders make to the way they communicate to the way they attack their business so for example one one one feature which holds quite a lot of that personality the personality proxy is basically studying the evolution of a founder in terms of their like the progression through life both from a brought from a professional sense and a network type of sense in a time series way and if you study that and take each founder and compare them to similar cohorts of founders within the same geography they started in with the same age started from the same baseline and studying those slopes of those evolutions across the graph and the experiences they've done you're gonna see that some founders have outliers trajectory like for example reach certain statuses or reach certain highs followed by blows or more heights within age frames that other didn't get that's a feature that It takes a lot of the personality.

19:53And this goes to show that it's what you're looking at here are scrapers or data sources that are far beyond. Did this person change their LinkedIn status to stealth mode? Look, exactly as I said, that's the part of actually getting data in. That's trivial. How many people think when they talk about CoinVC, they think it's a scraper of LinkedIn sourcing. Well, that's basically not. It's almost a hygiene. Then the decision is the problem. Two years building up datasets before we ran models. And interestingly, what we found when we started to dive into this problem was the source datasets, the pitch books, the crunch base, the usual suspects, right?

20:39They're not enough on their own to drive the predictability that we wanted. The datasets required huge amounts of cleaning, contextualization, work, new pipelines. And what we found, one of the kind of the discoveries in the early stages was that there was, it was painful work, right? It was arduous, painful work that took many years. And we realized that no academic is going to solve this problem because no one wants to spend two years of a three-year, four-year PhD just cleaning data sources. When did academics last solve the problem? And then no VC is going to solve this problem because, again, the work is so painstaking that they're not incentivized to do it.

21:17And it's not work you can just hand over to a junior. It's not work you can outsource. It's not work you can pay something to do. It required the deep subject matter knowledge that we had picked up in the years running our venture building. Can you share some of the, while you say that it is hygienic to get the right data and so on, I think it could be interesting if you could just share some of the most surprising or shocking or important data sets that you're using, that you're pulling, where people would be like, I had not expected that. Yeah. Look, again, it's not in the data. The raw data comes from 14, 15 different sources.

21:57Like, obviously, you have data from the startup providers, which actually helps you do the label. You have web search data. We have a scraper infrastructure that goes across 6 million websites to find signals, to bid market competitors, and so on. That part, basically, it's almost in its raw form. It's fairly useless by itself, like the transformation. And to build a data layer, we run 14 other models in between. So when the data source is coming, that's the first data pipeline reconciliation, because many sources will contradict itself. So you need to kind of establish your ground truth of what you believe.

22:33Once that is done, the hardest part is across the enrichment of the data layer and structuring the features in a way that make the relevance to the model. And most of the features, especially because it's pretty easy, they have to be quite composite type of features. But the one that I explained to you before is one that's very important. So for mapping out someone's trajectory, so founder's trajectory of work, you need to figure out a lot of things. Like, obviously, the easy stuff to figure out for each experience that they had in the past, for example, obviously, you need to figure out what they were doing.

23:08So if they were, I don't know, doing data science or podcasting or how they're going about it. But when the problem comes in, it's even like conceptualizing your quality of a company that impacts, like the company that they work for, we kind of intuitively know that it impacts what they did. So it's a big difference if you work at a short-mashop down the corner or if you work at OpenAI, right? But then there is more complication because not everyone is the same. Then you need also to end up having functions that actually can compare, I don't know, Morgan Stanley with a CDSA startup from Poland, randomly, for example.

23:47For that, we use graph databases, which get into a lot of complications on how you set your graph, what type of algorithms you use to actually understand the weights in the graph. So you have all these multi-variables that you have to stitch together, which makes it painstakingly hard because you can have your assumptions in the beginning, but the way you can massage your variables is almost infinite. So it's a lot of trial and error and more intuition building. And you end up, or what we ended up is with a feature set that is very counterintuitive from a human perspective, simply because the algo's thing different and the available data that you can get is much richer than the classic human understanding on how big the tan is and how nice the founder smiles, basically.

24:37You have to dig a lot deeper. Yeah. Now let me go to another question, which is a bit about rule 30 then. And it's that if someone goes to your website, They'll see that you're using the word AI research lab about yourselves, not welcome to the next gen VC something, something. Tell me about that decision. What does that mean? What does that imply? Look, our ambition is not to build a fund and have a nice life in that sense. We're not chasing the VC life. What we think is that the VC space is the hedge fund moment of the 80s and 90s. and basically the ambition is to build a renaissance capital for the private market we started with vc once because we knew the ecosystem but also because it's the by far the hardest technical problem that you can solve so building models that precede is according to gpt because i don't have my opinion here between 50x and undertext harder than building them at series b but by conquering vc well conquering precede in vc building upstage model becomes fairly trivial because you would have had gone through that initial data structuring and layering and so on and it feels like the best beach hack to start from what's also access to these companies is a lot different than in later stages at series beats humanly obvious a lot more competition the idea is build unlocking what we unlocked in pre-seed allows us to go upstream and basically build models and strategies across multiple assets in the private market So the future of Rule 30 is as an asset manager, it's not as a system that powers the decision making of A16 set and the like.

26:28I mean, we don't sell alpha on this. It's like we had a lot of push and we had to bargain still to survive all these years because obviously no VC actually wanted to give us some proper financing, even though when you do the testing with them, they mostly get completely flabbergasted and some of them disappear for a few weeks and then they come back and want access to the algo. But we always resisted, always resisted. We're not a SaaS business. We're sure we built Alpha. And it's all about using the technology to mine the Alpha. It might be that we're going to use it in strategic partnerships.

27:09We will potentially power also LPs with our technologies. But in no way, shape or form, we're going to power other VCs. Maybe just talk about the defensible edge of, Because that's one of the things that we often talk about when we talk about this datafication, quantification of ventures. Will we have an actual edge? Will any firms have that? Or does it all just become a set of very large asset managers that deploy on the bag of models? Yeah, I think the way that we see venture evolving in the next decade is there will be the emergence of firms and funds like ours who build decision engines, who find signal or patterns that can push that decision faster.

27:56And we'll find ways of putting money into companies at a quicker and quicker stage, in an early stage. which is in the way that we also see it is that hedge funds will realize what we've realized. And they have the skills and the experience to enter the space. And in the past, we've seen them trying to enter funds like Tiger wanted to do venture. It didn't pan out exactly as they thought. But people know that the returns that can be made in the asset class are trying to figure out ways to do it. And I think what we'll find on the other side is the solo GP, the emerging GP, the groups who have the really deep-rooted networks that don't rely on the data platforms or the signal providers to find deals and are able to quietly go about finding those companies because they come through very trusted sources within their networks.

28:46And what you'll end up finding is that the deals that will happen will be the combination of the QAM funds will come in with the money quickly and the more network-driven VCs will partner up with them and deploy capital that way. I just remember that I wanted to ask you because you said before that in your model, and maybe I heard it incorrectly, but you said that would you really optimize for the delta between the initial round that you would come into and the follow on round? Did I hear that correctly or is that wrong? now basically when you create a label to teach me and this is a very very hard problem like if i if i would say look look at the past and decide what you would have wanted to invest in right and that's basically when you build this binary classifier which our decision engine is is basically saying what is the success class what should i learn from that is good and what should i learn from that is bad that's a very tricky problem in venture because you know because this huge timelines so to actually look at the single decision and be 100 sure that you want your algorithm to pick that theoretically would have to wait 10 12 years and a good example is builder ai almost conceptually would you train your preced algorithm to pick builder ai or not in in that sense right and because you have these long waiting periods you have to find solutions to create labels to train your model on recent companies.

30:14What we saw after a few years from the first funding round, for almost five years from the first funding round, at the portfolio level, so not at the singular decision level, there is a very high correlation between the outcomes that happen at the general portfolio level between those five-year period and 12-year period, right? And then you use that five-year period to build evidence, basically to build evidence that, look, from this vintage over five years, top 10%, right? To allow yourself to build a label closer to hope. So basically, in essence, you're trying to shorten the time period in which you can create a label data set.

30:56But you still have to keep it correlated with the final outcome at the portfolio level. Yeah, of course. Okay. I just wanted to double click on that. No, it's not about follow-ons. You kind of need to get, you need to predict at the level of the outcome almost becomes asymptotic with, yeah. There's a saying in venture, which is that the best predictor of future success of a startup is who leads the follow-on route. So basically, if you have one of the very successful firms, that is actually the best predictor of what will happen to the portfolio company in the future. Is that true based on what you found?