Notes on The Man Who Solved The Market (Jim Simons)

Like seemingly everyone who works on the buy side, I have been reading the Zuckerman book about Jim Simons of Renaissance. The book spends a lot of time with the big personalities who have worked at Renaissance over the years. However there are some mathematical and technology hints:

Jim Simons’s academic field is geometric invariants in algebraic geometry.

Renaissance Technology worked with hidden Markov models, fit using the Baum-Welch algorithm. This algorithm has a Bayes update step, and a backward & forward process that feels like backprop. They also used high-dimensional kernel regression.

Henry Laufer worked with (vector) embeddings, very early. He also pushed for a single cross-asset and cross-asset-class model, so they could use all the cleaned “pricing” (market) data. They included assets with known bad data, but assets that nonetheless looked like existing assets in the model. This is maybe what we would call clustering, now-a-days. Everyone had access to the source code, even the administrative staff at first.

They tempted academics by working just one day per week, to see if they found trading interesting. They explicitly avoided trying to find economic sensibility for their strategies, but still followed a “scientific” method.

René Carmona imputed market data, which seems controversial.

Jim Simons invested in private companies alongside the systematic trading, especially in technology companies. This is probably because of the Nasdaq bubble.

They used a simple day-of-week seasonality model, at least for their futures trading.

They provided liquidity when “locals” de-risk, being in the “insurance business”.

They had a betting algorithm around the probability of moves in futures, not used for stocks at first. This was presented as opposed to statistical arbitrage with a factor model.

Their stock transaction cost model was the “secret weapon”. It was self-correcting, by “searching for buy-or-sell orders to nudge the portfolio back”. This increased their holding period to two days, on average. The strategy had very low capacity at first, however. In general they were not the best at trading, but at “estimating the cost of a trade”.

Around the time of the Nasdaq bubble bursting, they were trading 8,000 stocks. However this strategy was only 10% of the business. The futures strategy was still the mainstay, which was probably their chartist model.

Their use of basket options were 1) a tax optimization, but also 2) a way to cap downside, 3) isolate risk, and 4) increase leverage.

The internal Medallion fund is short-term, capped at $5b capital with all external clients eventually bought-out. The maximum capital is difficult to reconcile with Jim Simons’s compensation motivation, so it probably reflects limited capacity of the strategy, which means it trades in less liquid, smaller stocks and futures. In 2002 they were running 12x leverage in Medallion ($5b capital, $60b exposure given options).

They sought out astronomers because of their understanding of low signal-to-noise problems. Their named “Déjà Vu” strategy seems like a pairs or cointegration strategy.

Their strategy gets a 50.75% hit rate.

Why write this book now? Jim Simons is nearing the end of his career. Could also be transparency after the Mercers helped get Trump elected.

They at least use the terminology of risk factors and baskets: “RenTec decrypts them [inefficiencies]. We find them across time, across risk factors, across sectors and industries.”

The author uses strange terminology, suggesting that Zuckerman has actually not talked with many quants. For example, “trenders” for momentum style traders, and “data hunters” instead of “data scientists”.

Also the anecdote about Magerman unleashing “a computer virus that was infecting Renaissance’s computers” is clearly bullshit, and in a way that makes me doubt the author’s understanding of technology in general.

Sequential Learning Book

Things have been quiet around here since the winter because I have been focusing my modest writing and research skills on a new book for O’Reilly. We signed the contract a few days ago, so now I get to embrace a draconian authorship schedule over the next year. The book is titled Sequential Machine Learning and will focus on data mining techniques that train on petabytes of data. That is, far more training data that can fit in the memory of your entire Hadoop cluster.

Sequential machine learning algorithms do this by guaranteeing constant memory footprint and processing requirements. This ends up being an eyes-wide-open compromise in accuracy and non-linearity for serious scalability. Another perk is trivial out-of-sample model validation. The Vowpal Wabbit project by John Langford will be the book’s featured technology, and John Langford has graciously offered to help out with a foreword. Therefore the book will also serve as a detailed tutorial on using Vowpal Wabbit, for those of us who are more coder or hacker than statistician or academic.

The academic literature often uses the term “online learning” for this approach, but I find that term way too confusing given what startups like Coursera and Khan Academy are doing. (Note the terminology at the end of Hilary Mason’s excellent Bacon talk back in April.) So, resident O’Reilly geekery-evangelist Mike Loukides and I are going to do a bit of trailblazing with the more descriptive term “sequential.” Bear with us.

From the most basic principles, I will build up the foundation for statistical modeling. Several chapters assume statistical natural language processing as a typical use case, so sentiment analysis experts and Twitter miners should also have fun. My readers should know computers but need not be mathematicians. Although I have insisted that the O’Reilly toolstack support a few of my old-school LaTeX formulas…