Quantopian

» see quantopian.com as it lived, on the wayback machine

~ the obituary ~

John Fawcett and Jean Bredeche started Quantopian in Boston in 2011 with an argument that sounded reasonable and turned out to be expensive. Quantitative trading was locked inside a few dozen firms because the inputs were locked inside them: clean historical price data, and a backtesting engine that did not lie to you about what the strategy would have done. Give both away free to anyone who could write Python, run a contest with real capital behind it, and the talent would find you instead of the other way around.

It worked as a platform long before it worked as a fund. You opened a browser, got a notebook, wrote a function that took a universe of American equities and returned a set of weights, then watched it trade backward through a decade of history in a few minutes. There was no install and no data license to sign. For a generation of self-taught programmers with a finance itch this was the only door that opened, and 85,000 people were through it by July 2016.

The money side took longer. Quantopian raised $48.8 million from investors including Spark Capital and Khosla Ventures, and in July 2016 Steven Cohen's Point72 committed up to $250 million to be managed through algorithms the platform surfaced. Members whose algorithms got an allocation earned royalties on the capital above the first million dollars. By August 2018 the community was past 210,000 people.

Then the returns stopped. The house strategy was low-risk and market-neutral, and by early 2020 it had gone two years without producing anything worth charging for. Quantopian asked the community for different kinds of models, which is a difficult request to make of volunteers who joined for a contest with fixed rules.

~ the case file ~

~ the algorithm contest ~

The contest was the whole engine of the place. You submitted an algorithm, it ran on live data alongside everyone else's, and a leaderboard ranked them on out-of-sample performance over months rather than on a backtest you could have overfit. Winning meant an allocation, and an allocation meant royalty income from a hedge fund without ever having worked at one. People built careers out of it. The forums are full of members saying the platform was where they learned to program properly, which is not the usual thing to say about a trading site.

It was a closed loop around one market, though. The data was American equities and the capital behind the allocations was an American hedge fund's, with a payout path running through brokerage plumbing that assumed you lived there. From Ankara the platform was readable and the contest was watchable, and that was about as far in as the geography let you go.

~ the fund that did not work ~

Crowdsourcing alpha has a structural problem that nine years did not solve. The strategies a public contest rewards are the strategies a public contest can measure, which pushes every entrant toward the same market-neutral shape. Thousands of people optimizing independently against one scoreboard end up crowded into the same trades. Point72's commitment was reported in the hundreds of millions. The returns that came back out of it were never reported at all, which is its own kind of answer.

~ the five days ~

The notice went up on 9 November 2020. The community offering would be retired on 14 November, and anything not saved locally by then was gone: notebooks, algorithms, backtest histories, nine years of forum threads. "0-day notice! Are you kidding me > where is all the code ???" reads one of the replies, and it is representative of the thread.

Later that month Fawcett and Bredeche joined Robinhood along with most of Quantopian's headcount, which is what the whole exercise had actually produced. Zipline, the open-source backtesting engine Quantopian built and released, is still on GitHub and still runs. The old address now redirects to a community subdomain that answers with a Cloudflare page asking whether you are human.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »