Initially he bought and sold commodities, making his bets based on fundamentals such as supply and demand. He found the experience gut wrenching, so he turned to his network of cryptographers and mathematicians for help looking at patterns: Elwyn Berlekamp and Leonard Baum, former colleagues from IDA, and Stony Brook professors Henry Laufer and James Ax. “Maybe there were some ways to predict prices statistically,” Simons said in a 2015 interview with Numberphile. “Gradually we built models.”

At their core, such models usually fall into one of two camps, trend-following or mean-reversion. Renaissance’s system had a foot in both. Its results were mixed at first: up 8.8 percent in 1988, its first year, and down 4.1 percent in 1989. But in 1990, after focusing exclusively on shorter-term trading, Medallion chalked up a 56 percent return, net of fees. “I was confident that the models would work better,” says Berlekamp, who returned to academia in 1991 and is now a professor emeritus at the University of California at Berkeley. “I didn’t think they would be as good as they were.”

Eventually the scientists went so far as to develop an in-house programming language for their models rather than settle for a numbercentric option such as ASCII, which was popular at the time. Today, Medallion uses dozens of “strategies” that run together as one system. The code powering the fund includes several million lines, according to people familiar with the company. Various teams are responsible for specific areas of research, but in practice everybody can work on everything. There’s a meeting every Tuesday to hash out ideas.

David Fuller's view

The biggest crowds, in terms of wealth, move the markets with their money.  For this reason, the best black box or mathematical models, whatever you want to call them, will be technically rather than fundamentally driven.

Note the sentence which I emboldened above.

The biggest crowds, in terms of wealth, move the markets with their money.  For this reason, the best black box or mathematical models, whatever you want to call them, will be technically rather than fundamentally driven.

Note the sentence which I emboldened above.