Steven Cohen’s experiment at diving into quantitative investing may have hit a speed bump on his way to launching a new hedge fund with the widely hailed Quantopian platform. The Wall Street Journal reported that the experiment at unleashing a horde of mathematicians, engineers, and computer programmers onto the market panoply has been met with anemic initial returns. What could they do to improve performance? The lackluster performance was followed by Jonathan Larkin, the chief investment officer who put together the algorithmic strategy, leaving the firm.
Recruiting those without an understanding of how markets operate on a beta level leads to quantitative failure
It wasn't long ago that Steve Cohen was decrying the lack of talent in the hedge fund industry. Now the issue may be coming back to bite Cohen in poor initial Quantopian platform performance. What could go wrong?
Perhaps the most significant challenge those without a fundamental understanding how markets work face when developing algorithms can be seen in one sentence from the Journal article:
"Recruiting and training mechanical engineers, nuclear scientists and other amateurs around the globe to be quantitative traders may be more difficult than some had expected."
What many of these 160,000 users writing more than 600,000 trading algorithms on the Quantopian platform have in common is a strong technical knowledge. The Boston-based company had earlier announced a bold vision: to create a platform of tools and applications that would allow computer programmers to program and test their algorithms in a cost-free environment.
They built tools that are widely considered a significant advancement, creating shortcuts and access to data that were previously roadblocks to widespread quantitative adoption.
Through this process, the firm selected what it considered the best of the algorithms based on a backtest, placing them into a diversified blended portfolio.
So what went could have gone wrong? While there are a lot of potential issues, ValueWalk considers one core issue of how algorithms are organized and correlated.
ValueWalk questions about the Quantopian platform that go unanswered:
ValueWalk sent Quantopian questions in an effort to determine where the problem might have occurred. After initially agreeing to answer questions, Quantopian declined to respond after receiving the questions.
Question 1) How are the algorithms organized/categorized?
This is the first fundamental mistake that computer programmers and mathematicians can easily make is in organizing and categorizing the algorithms. At times this is down in conjunction with a fundamental market understanding that is provided as the top overlay to markets. The key in developing blended portfolios, oddly enough, is not often first found in the math of the individual algorithms, but rather in the blended strategy a developer uses -- or more common, the lack of such a clearly defined portfolio management plan.
While a few short months of performance is not sufficient to categorize the validity of the overall approach, often times multi-strategy portfolio managers who do not have a clear and understandable process for strategy diversification are also those that do not operate properly during a variety of primary beta market environments.
The strategy to diversify should include the categorization and organization of algorithms, it should not at an introductory level exceed one page and should be understood by traditional portfolio managers. The algorithms should be correlated with a market environment and each formula should have a justification as to why the statistical anomaly works and when or under what circumstances that formula might find an expiration date. If a fundamental portfolio manager cannot understand the logic behind the formula this should be given negative attribution in a scoring method.
2) How many different algorithms are used in the portfolio?
In the organization of the algorithms, allocation towards macro market environment events that have an economic justification should receive allocation percentages. Too often, a high number of algorithms leads to double covering the same beta, or repeating market event. This creates a correlation imbalance and thus during differing market environments the algorithms don’t generate positive returns. Too often this can be seen in an algorithm overweight to the market environment of price persistence, which is primarily used in trend following.
3) Is there a concise allocation/exposure strategy? The best algo exposure strategies I've seen typically are no longer than one page. Do certain algorithms receive a higher weight than other algorithms? Is the exposure timed based on beta market environment analysis or any other macro weighting system?
The key to all multi-strategy programs is the overall allocation strategy. Oddly, evaluating the math of each individual strategy might not be as important as first filtering this understanding through the beta found in the predominate market environments.
4) What is the reason given for the underperformance? Did the backtesting not live up to live results?
Backfill bias is a subtle killer. If strategy testing is done, how many primary and secondary market environments were stress tested outside of the actual performance history? Was individual validation of a strategy done during a crisis?
It is too early to call it quits on the Quantopian vision or success. Four months performance history is statistically insignificant for a fund with an average time horizon that includes mid-term and long-term focused algorithms. Regardless of the outcome, the Quantopian platform has achieved significant success at building a user base and democratizing quantitative investment methods.