Jim Simons, the legendary mathematician and founder of Renaissance Technologies, just passed away. He lived a fascinating life, moving from academia to building one of the world’s greatest money-making machines in Renaissance’s flagship hedge fund.
Renaissance, or RenTech, was a pioneer in algorithmic trading, using data to gain a small edge and making many trades so that in aggregate it made money. It’s been staggeringly successful, and gave rise to a wave of quant trading that is now a huge force in public markets. I enjoyed this book about the subject if' you’re interested.
Simons’ death made me reflect on my last job, where we worked on using quantitative strategies to make venture capital investments, specifically in early stage CPG brands (think Halo Top, Spindrift, Dollar Shave Club — this was the go-go years of the 2010s!)
I had read about the quant revolution in public markets and I listened to this podcast describing the strategy in private markets, which I found very compelling. CPG brands spin off a lot of data (they have sales from the very beginning, unlike many tech companies), so they seemed like a perfect place to apply a data-driven strategy.
This is how I described the job to my uncle (and noted Joe’s-substack-reader):
He responded, “Congratulations! Not sure how they do it but looks interesting!”
His confusion was warranted: there were a lot of difficulties using quant techniques in early stage venture that I hadn’t anticipated, enough that I now believe that until private markets become much more liquid, it won’t be a profitable exercise.
Data gives you a small edge
This is the root cause of the issues we faced: using quantitative methods makes your decisions better, but only marginally. For a hedge fund in the public markets like RenTech, this is still hugely profitable, because they can trade in and out of hundreds of positions per day. The analogy I’ve often heard is that successful quant trading is like betting on a biased coin, with a 50.1% chance of coming up heads. If you bet on it enough, the law of large numbers will take over and you’ll make money.
But in early stage venture, that small edge really doesn’t matter very much.
First, venture fund portfolios are relatively small, usually with fewer than 50 companies in a fund. So you can’t flip the coin thousands of times like a hedge fund can; you are limited in your chances to play.
Second, and what I drastically underestimated: you have to convince the companies to take your money! As a private investor, you have to chase down founders of companies and convince them that taking your fund’s money is better than taking someone else’s. That can mean cold-emailing them, complimenting their performance at a Barry’s Bootcamp, or treating them to drinks at Soho house. This takes a lot of time and work (venture associates are very busy) and adds a large transaction cost to each investment. This transaction costs eats up the edge you get from being data-driven.
Compare this to quant funds in the public markets, where investors can take positions without permission and without high transaction costs because the securities are traded on liquid public exchanges. The small edge that data gives you is much more powerful there.
Third, you learn the results (and get paid) far in the future. A benefit of public markets is that you can see the price of a security (and buy or sell it) on any given day. If I built a model to predict the price of Sweetgreen stock based on the daily wait times at the location near my office (sell! sell!), I could check and see how my model was performing every day, because the stock is publicly traded.
If, however, I’d built this same model 10 years ago, when Sweetgreen was still private, I could only see how my model performed once a year or so, because the price of a private company is only determined when it raises money. Sure, the data on wait times could help make me more informed about the company’s operations, but since I’d be investing privately, I’d be in it for the long haul — I couldn’t trade on it as a signal.
In the venture capital model, it takes a long time for you to make money on a position (usually years, until the company gets acquired or goes public). You’re compensated for this — the winners can make you 20x your money or much more. But because it takes so long, it’s very hard to attribute a return to an early signal. It is a lot easier to be confident in a model that predicted something yesterday than one that predicted something 3 years ago. Renaissance’s usual holding period is said to be "between hours and weeks.” This is much more conducive to using data than the years that VC requires.
Quant strategies are expensive
All of these combine to mean that the investment required to build a quant trading system (which is substantial: at this company there were PhDs with wall street experience, machine learning engineers from FAANG companies etc) does not end up paying off. The edge it gives you is not large enough to matter.
The only way I see this changing is for private markets to start to look more like public markets. This is slowly happening: late stage companies have secondary markets, and companies like Carta have come up with private exchanges. Some crypto technologies (remember ICOs, initial coin offerings?) create liquid, tradable securities. But there is a long way to go, and I’m not sure if it will ever happen at the really early stages of venture.
Seeing these challenges play out was a good lesson that thinking by analogy can be dangerous. I thought, “quant trading revolutionized public markets in the 80s and 90s — shouldn’t it happen in the private markets too?” But I didn’t understand the mechanics enough to truly evaluate whether that made sense.
I’d love to hear others’ opinions on this as I have a limited view — please comment or email me! This Acquired podcast on RenTech is also great if you want to learn more about quant investing.
I really enjoyed The Man Who Solved the Market as well, thanks for recommending it a while back!
I agree it would take a complete market transformation for Quant to work in Private Markets (that 50.1% edge you mentioned). I'm not sure it will ever come, but I'd like to see it! Could bring about a more just/fair PE environment.
Another provocative post. I would love to see a thoughtful analysis showing the kinds of problems where AI is likely to be useful.
As you note, the AI answer must be timely and clearly worth its cost or no one will pay. And the client—not the AI nerd—decides what counts as “useful.”
Useful compared to what? If existing methods are pretty good and the consequences of being wrong are small, then it’ll be hard to get much traction for an AI solution.
For example, say you would like to predict the solubility of a new small molecule drug candidate. I’m sure you could find a big data set of chemical structures and their solubility constants, and you could train an AI model to “discover“ the underlying causes of solubility. Quickly you would zero in on an AI model that successfully predicts the solubility of new molecules.
However, chemists already have molecular modeling software to do that, and more. And they also have beakers and water and if they want to, they can synthesize the new drug and drop it in the beaker, stir it, pour off the liquid, let it evaporate, weigh what’s left, and find the solubility that way! So they might pay something for your AI solution, or they may not depending on price and depending on how quickly you can deliver it and how accurate it is in relation to alternatives. Even if it’s better as you point out in your post, they still might not buy the AI if its answer isn’t timely.
A further constraint, and I would like to see proper research on this, is the size of the data that you need for AI to be able to pull out anything useful. I believe that there are problems in physical science where huge data sets are available,, but much in economics and business just doesn’t have the requisite scale. Activity data sets such as you had in your old company are very, very small in relation to those in physical science. For example, someone in my old company estimated that the number of synthesizable new drugs (using common chemical fragments and using common methods of synthesis) was about 10^26th potential molecules. Now that’s a dataset!
In contrast there are 3000 NASDAQ companies, so 20 years of closing price data is only 3000*20*250 is only 15 million data points. It speaks to how inefficient markets were that Renaissance was able to make any money at all!