Venture Capital and Moneyball

Or why most venture firms still aren’t optimizing their operations.

Over the last couple of weeks I had a bunch of conversations with friends in the venture space about the purely operational aspects of the business. How they translate reams of available data of wildly varying quality and structure into venture returns via collecting, aggregating, storing, and transforming the data into investment decisions.

The overall picture that emerged is that very little of this happens systematically. Venture capital is still relationship-centric, even at the top end where in-house data science teams try to instill some quantitative discipline.

Granted, venture capital is an exceedingly tricky industry with long cycles, high fail rates, highly unstructured (and hard to access) data, and a far-from-obvious connection between the input data and any kind of success metric. Under these circumstances, it’s not surprising that “process optimization” still tends to play second fiddle to “relationship optimization”.

But there’s a number of reasons to expect that this will change, and change might come very quickly.

AI-based assistants (of the type that currently receives the bulk of venture money) will automate many if not most of the repetitive front office and back office tasks. As a result, venture funds will not only become smaller (measured in partners) but also leaner (measured in support staff per partner). Solo VCs will become a common feature of the landscape.

Speed will increasingly become the determining success factor.

Admittedly, some venture firms already tried to speed up their funding decisions during the feeding frenzy of 2021–22. But that did not end well.

If your processes are still mostly manual, speed will inevitably translate into cutting corners on quality, which is why any kind of push towards increasing velocity, from discovery to due diligence, contracting, and governance, will have to start from a coherent process model.

This where Moneyball comes in

“Moneyball", the shorthand for data-driven athlete sourcing, started out with the Oakland Athletics at the turn of the millennium, and has since taken over not only baseball, but pretty much all of pro sports. From there it moved on to politics via Nate Silver’s 538 website (Nate started out as a baseball quant geek) and in general had a hand in the emergence of “data-driven anything", aka data science.

The underlying motivation for the cash-strapped A’s was to detect underrated skills sets that evaded the traditional “tools" oriented scouting operations and that would allow the team to sign players that weren’t on anyone else’s radar.

The parallel should become obvious now: scouting for athletes, like scouting for startups, isn’t just about locating the targets that seem to have the highest likelihood of succeeding — it’s also about finding the targets few others have on their radar. A bit more formally, it’s about the divergence of beliefs. (I’ve written more about it here.)

In other words, it’s about recognizing opportunities no one else sees, to see them earlier, and to avoid the fallbacks of herding.

The reason Moneyball started out in baseball has a lot to do with the fact that it’s the most “individualistic" of the team sports, with reams and reams of historical data, broken down to individual plays and players.

From there it moved on the other American sports and ultimately found its way to association football (aka soccer) despite that fact that soccer is much harder to break down and the data is much less structured.

This is a problem that besets venture capital even more. In addition, much of the available data on startups is qualitative, and in the case where quantitative data is available, it’s much more speculative and idiosyncratic.

That’s also a reason why in venture capital the competition between “trads" and “quants", believers in experience-based quick judgment vs believers in measurable input, still leans much more in favor of giving human judgement the final say.

That’s not an unresolvable conflict. The division of labor clearly goes in the direction of using quant methods early on to sift through large amounts of candidates and gradually lean more on the human judgement element as the number of options has dwindled down to a manageable number.

The crucial question is then where to make the transition and how to handle situations where the two camps diverge.

The loop around the “human in the loop"

Moneyball has been around for 25 years, and data-driven decision making has conquered most fields, especially those where lots of money is at stake and uncertainty is high.

That venture capital, which played a big role in ushering this era in, still leans heavily on the human element might be surprising, but we have to consider that it is still mostly a relationship business. Handling both startups and limited partners requires a fair amount of skills that go beyond the quantitative.

Successful venture partners would then mostly describe their acumen as “data supported" rather than “data driven", and that’s unlikely to change anytime soon.

The loop in “human in the loop” will still center on the human intuition, but the loop around it is going to change very fast. And that’s not only because Generative AI, Large Language Models in particular, primarily work in qualitative (verbal, visual, acoustic) domains, but also crucially that it can consume all kinds of unstructured data and transform it into output in any desired form, from databases to radars or blog posts.

Quality is still sketchy, and maybe even more troublingly, LLMs tend to herd: they produce similar outputs when diverging perspectives are desired. One is reminded of the famous flash crash of 1987, when newly deployed automated trading algorithm all jumped at the same time, because they were all thinking the same.

Let’s be optimistic that output quality continues to improve at a similar rapid clip as the last two years. Then the two dimensions along which venture firms will compete is uniqueness of the output produced, and integration of the novel toolbox into the decision making process.

Uniqueness of the output depends to some extent on unique input, but not only. It also depends on being able to extract the best signals from what is always and inevitably noisy input — and the ability to act on it.

So to maximize the gain from deploying GenAI tools, they should be build around a decision model that minimizes loss of information and procedural friction. Given the rather patchy shape the processes at VCs are in, this push of new tools also creates the right opportunity to trigger a fundamental rethink about process optimization.

This goes beyond mere picking tools to rethinking sourcing, decision, and execution mechanisms along the whole funnel, which in turn will drive tool choice — both off the shelf and purpose-built.

Moneyball and the attention economy

I’ve likened venture funding to a rigged lottery before, in the sense that the participants would like as much as possible to “rig” the odds of finding a winning ticket, but no matter how much rigging happens, the odds of hitting the jackpot remain recalcitrantly low.

The corollary is that even the superstars in the field, the Sequoias, the Kleiner Perkinses, the Y Combinators are still operating far below their hypothetical performance envelope — most of their funding allocations still go nowhere, which also means that the potential upside of an optimized sourcing operation is still massive.

When I write about the attention economy, I don’t usually mean “social media and content creation”, but something very specific. In an environment where there is far more data available than we can consume, in other words a world of constant cognitive overload, we need to build systems that put the scarce but critical resource — human cognitive effort — at the center of the process to be optimized.

Expressed differently, we need to steer human attention to where it is most usefully deployed, and this is what Moneyball ultimately did, expanding by orders of magnitude the information that can be ingested to support human decision making.

The information-theoretic expression for this is “surprise” (or surprisal), a measure for how far an observation deviates from the expected.

Venture capital translates into a constant scanning for positive surprises, but the sourcing part of it — the discovery of technological opportunities and the players that translate them into economic opportunities — is only the front end of a multi-stage process.

The commonality of that multi-stage process, the “funnel”, is that at each stage, a decision has to be made on which additional resources will be allocated to which most promising target.

The type of resource differs by stage, from exploration to matching to contracting and agreement on terms to an ongoing advisory relationship to finding the right exit, different skills are required and different people will be engaged.

The junctures inbetween are the decision points, and they are currently the least optimized parts of the operation. This is where the friction tends to come in and where the proverbial ball gets dropped.

Funnels as selection processes work by iteratively weeding out the least promising options. But in a business where “least promising” can be a positive signal, a single-elimination funnel can produce costly false negatives.

Like sports, the venture world is full of comeback stories, athletes that were dismissed by their youth teams but ended up playing in the World Cup Final or the Super Bowl. Similarly, all successful venture firms know that they missed out on a few later unicorns. The good ones can laugh about it.

This world, where a very large universe of potentials, out of which a vanishingly small fraction will determine the success or failure of a scouting operation, will be completely upended because in short order we will have automated systems that can emulate rather than complement human effort, at scale and at speed.

Venture capital is the main driver of this revolution, but in the near future, it will also be its target.