I recently finished reading “The Man Who Solved the Market: How Jim Simons Launched the Quant Revolution,” by Gregory Zuckerman. It’s the most detailed public look at how Renaissance Technologies became the most successful hedge fund in history. (Yes, they blow Bridgewater and everyone else out of the water.)
Renaissance’s flagship Medallion fund, which is run mostly for fund employees, is famed for the best track record on Wall Street, returning more than 66 percent annualized before fees and 39 percent after fees over a 30-year span from 1988 to 2018.Wikipedia
As a publicity-averse firm, that doesn’t mean much, though: the “details” of how they did it are fairly nonspecific. And, as quants, their tools are often pretty different from core fundamentals.
So, aside from it being an enjoyable, quick read that features the backstory of Robert Mercer, is there anything value investors and other long-term investors can learn from the book?
How Renaissance Works
Like other quantitative trading shops, they employ hundreds of the smartest math/astrophysics/etc. Ph.D.s they can find. They’ve also been doing it longer than almost everyone else. Long Term Capital Management was right there with them at the beginning, but obviously, we know how that turned out.
I imagine that a lot of people picked up this book to see if they could learn anything at all that might help them with their own trading. After reading it, I doubt it. But, for those who know something about machine learning and try to use predictive models in their process, here are some nuggets I gleaned:
- They mine historical data for any kind of predictive signal. They don’t care if they can explain it.
- They were one of the first to compile really high-quality proprietary data. Assuming they never stopped, they are probably using features/predictors/datasets that would blow your mind today.
- They usually hold positions for days to a couple of weeks. This allows them to train models with relatively large amounts of observations, as there are many 1-14 day non-overlapping periods in the data.
- They don’t appear to be “flash boys”/high-frequency traders that are trying to front-run trades, or at least there’s no mention of it.
- They originally used Hiden Markov Models as they considered this a good model of the market. It’s not clear that this worked or is still in use today.
- They’re also quoted as using “kernel methods,” so a Support Vector Machine or something similar, in the early days. Maybe they still have this in their model, maybe not.
- One of the most interesting things that was emphasized is that Renaissance uses a single model for all their trades, be it equities, commodities, whatever. I find this incredible, to the point that it seems not credible. How could you have the same input features/predictors for every possible trade? Anyway, this is either a pretty impressive accomplishment or is a bit misleading in terms of how things are implemented.
- They had to develop an equities model that worked in order to deploy sufficient capital, and this was a real struggle for a while. They initially used a lot of market-neutral pairs trading strategies (i.e., buy Coke sell Pepsi if their relative pricing gets out of whack in a certain direction), but I got the impression this wasn’t a big part of what they do anymore. However, their approach is still described as “market-neutral.”
Advice for People Using Machine Learning at Home
All of this is to say that even though I have access to some decent data, and I have some machine learning tools in my toolkit, I can now very confidently say that attempting to use ML/AI to find useful predictors on short timeframes would be foolish. They have much smarter people and much better data.
They employ astrophysics Ph.D.s that specialize in finding the tiniest signal from reams of data!
Actually, the most insane thing mentioned in the book is that they are so good and quick at identifying anomalies, they can trade to obfuscate the anomaly or make it look like something entirely different as they are profiting from the underlying anomaly.
So, literally, even if you have very good market data, you will probably not even see the anomalies/signals in the data over the last 20 years, as Renaissance was likely actively trading it with the intent of disguising what it was doing.
Lessons for Value Investors
I would never want to trade against Renaissance on a short term basis. If you’re trying to profit this way, good luck.
But for long-term investors, there is plenty of good news.
Long Live Long-Term Investing
The biggest is that there is simply not enough data to build really sophisticated models like Renaissance’s if your timeframe is 1 year or longer.
There are only about ~100 non-overlapping observations in anyone’s data (i.e., 1 year intervals that don’t overlap), and that’s just not enough for model training, validation, and testing. If you plan to hold for 3+ years, there’s even less data.
So, take heart: with that kind of timeframe, you are still competing in a reasonably fair race, where hard work analyzing fundamentals and evaluating intrinsic value usually pays off.
Focus on Your Process to Reduce Behavioral Biases
One other lesson for value investors from Renaissance: you can blow up your portfolio with bad behavioral investing, so dial in your investing process. It seems even the best freak out when their portfolios decline.
There were multiple examples throughout the book of the fear, uncertainty, and doubt that creeps in when losses mount — even for people that built their fortunes on their belief in computer algorithms’ superiority.
Jim Simons called up the manager of his foundation’s fund in December 2018 when the market was tanking, asking if they should be selling. His manager said they should wait until things settle before making any moves. The next day, the market rebounded. Had he sold, he would have missed out on a lot of the rebound.
Conversely, early in the book, one of Renaissance’s early employees was so overconfident he wouldn’t listen to new data. He blew up his portfolio and almost tanked Renaissance.
Investors of all breeds need to pay attention to their behavioral biases and be rigorous about their process to help mitigate them.