The Gateway to Algorithmic and Automated Trading

A few minutes with… Spencer Greenberg of Rebellion Research

Published in Automated Trader Magazine Issue 29 Q2 2013

Automated Trader catches up with Spencer Greenberg of Rebellion Research to hear about their work in artificial intelligence.

Spencer Greenberg

Spencer Greenberg

AT: How is a longer-term horizon different than what other players are doing?

Spencer: It is a different game when you are in a longer term horizon. You are looking for different types of signals. In our case we look for signals that are more stable, specifically seeking out ones that are likely to be maintained over long periods. You want your algorithm to learn about things that are going to continue applying in the future.

AT: How many variables do you take into account, how large are your data sets?

Spencer: There are thousands of variables in our data set. For a given decision on a given asset or stock, generally it will be a small number because [the system] reduces it down to the relevant ones. Often it will take into account 40 or 50 factors while making a single decision. But it really varies a lot.

AT: Some people point to heuristics as advantageous in building models in financial markets because they are simpler. What do you say to that argument?

Spencer: In some ways heuristics, or rule-based predictions, have nice features. You understand exactly what the rules are because you programmed them to do a certain thing. There also are some disadvantages, however. Rules tend to be static. If there is a market change, unless you go and change your rule it is going to stay the same and may no longer be appropriate. Machine learning is more complex in most cases. But it can have significant advantages. It can learn, it can change, it can update itself automatically if done right; so there are trade-offs.

AT: One of those trade-offs is that complex systems tend to be more brittle. How do you work around that?

Spencer: I think this is less an issue of machine learning per se and more an issue of being good at designing quantitative models. First of all, when designing your model, you need to make sure that what it is relying on is something that is likely to be maintained over time. Second of all, you don't want to be making money in just one way, generally. If that one thing you are taking advantage of suddenly changes it could be disastrous, so you want to rely on lots of different signals that aren't related to each other or have multiple ways of making money. That is something that applies across all types of quantitative strategies.

AT: What kinds of strategies do your machines use?

Spencer: Our machine learning automatically learns what strategies to apply. We have a large set of factors it analyses. Its goal is to automatically learn which of these are predictive of future performance, and then for a given asset, to take everything it's learned and make predictions based on that. There are some firms that use machine learning to generate a strategy, and then they will have that fixed strategy run so the machine learning is more a part of the exploration. In that case, the machine learning is not continually being activated, only used to find an initial strategy. For us it is not like that. Every day our system learns more and updates itself based on that new information.

AT: And what kind of computer programming languages are you using?

Spencer: We use Python primarily because it allows for very rapid development of ideas.

AT: There are concerns over the issue of over-automation in financial markets and how AI might be contributing to that. Do you share those concerns?

Spencer: The kinds of strategies that have created sudden issues in the market, like the Flash Crash, are ones that are reacting to very short-term information. It can destabilise markets when you have a lot of algorithms trading similar strategies on very short time scales. There is reason to be concerned that such a process can run out of control. With much longer term trading strategies there tends to be less of an issue. But it really depends on the type of algorithm. Some create volatility and some remove it. If an algo tends to buy when other people are irrationally selling out of fear, it will make more stability because it will push things more into equilibrium, whereas an algo that mimics the fear of other people can be destabilising.

Rebellion highlights reel

R ebellion says that in the year to date its global equities strategy is beating global indexes by between 5% and 7%, while an absolute return strategy based on the economic forecasts is up 4.6%. Here are some of the calls their system has made…

April 2007: Rebellion's machines started sending out warning signals on the US real estate market.

January 2009: Issues in Europe were identified. Then in September 2009, the machines downgraded the Greek equity market almost a month before any of the major rating agencies.

End-December 2011: Rebellion gave Argentina an F-rating - displacing Greece at the bottom of markets covered by the firm - at a time when the benchmark Merval index was recovering from steep falls between July and October. The index then fell to near two-year lows between March and July 2012 before a rebound beginning in late 2012, according to Bloomberg data.

Beginning of December 2012: The firm went negative on the yen though Alexander Fleiss, Rebellion's CIO and chairman, notes that signals in March showed Japan could be turning. Argentina's consumer rebound is also coming in with further positive signals.

Since last summer: The firm has been positive on the US housing recovery and, looking forward, indications for Turkey's economy gaining steam are strong.