Institutional Investor's Alpha Magazine - March 2008 - (Page 28) Quant Investing that more than half of his Ph.D. and postdoctoral students have gone into finance. “A lot of my colleagues in computer science and engineering departments have had the same experience,” he adds. Like an increasing number of quants, including Renaissance Technologies’ Simons, Moody is applying his expertise well beyond the equity markets. The $114 million JEM Commodity Relative Value Program, which he launched in May 2006, invests solely in commodity futures. JEM stands out from other commodity hedge funds because it uses statistical arbitrage and has a diversified portfolio. The fund trades in five sectors — energy, grains, livestock, metals and soft commodities — and balances long and short positions. JEM returned 13.9 percent last year, according to the Barclay Hedge d at aba se. From i nc ept ion through February, it has a compounded average annual return of 22.27 percent. Moody, who has used machine-learning algorithms to forecast commodity price movements, says the biggest challenge with financial data is that it’s so noisy. The more f lexible and — TANYA STYBLO BEDER, complex a quantitative model is, CHAIRWOMAN, SBCC GROUP the greater its chances of finding spurious patterns with no forecasting value. But Moody and other machine-learning researchers have found ways to build flexible models that work with a limited number of parameters. As a result, they minimize the overfitting that can come from using too many parameters while still capturing subtle relationships in the data. “The techniques that have been developed in machine learning are especially good at maximizing predictive power,” he says. One common style of machine learning is reinforcement, whereby an algorithm makes sequential decisions — while playing a board game like backgammon or checkers, for example. In finance, Moody says, reinforcement learning can be applied to portfolio rebalancing, where calls about what stock to drop or add depend on the success or failure of past choices. Another application is making trading decisions. Moody says he and his colleagues have created reinforcementlearning techniques that maximize the risk-adjusted returns of a trading system while taking into account transaction costs. “We’ll come up with a much different structure for a model than we would if we were to simply try to predict whether the market’s going to go up or down and then do trading and risk management as an afterthought,” he says. Anyone can see this and other work on trading strategies in academic papers that Moody has co-authored with colleagues and students. He admits that he and others in the machine-learning community regret sharing some of their discoveries. “I think a few of us wish that we hadn’t published them,” he says. Because markets are dynamic, machine-learning and other statistical models need constant updating. The key is feeding them the right mix of older and newer data. “You’re trying to find a sweet spot between using data that is stale, and therefore misleading, versus having enough recent data so that you can estimate a model reliably,” Moody explains. Incorporating unprecedented events like last August into machine-learning models may be difficult, but Moody says it is possible to simulate how a model will behave under some extreme circumstances. One approach is to take, say, eight years of daily market data and do computer-based Monte Carlo simulations — which use a random sampling of numbers to create potential outcomes — to generate a hypothetical 800-year trading history. If this is done properly, the synthesized data will reflect previously observed major disruptions. Investors can use such an approach to estimate the frequency and magnitude of extreme market events, as well as the associated trading risks. Moody notes that most simulation methods can’t capture disruptions or changes in market behavior that have no historical precedent. Still, they can be used to create a model that generates trading signals — and takes fewer risks. “By using these kinds of techniques, you’re informing the model of the possibility of extreme events,” Moody explains. “And that will give rise to a more conservative strategy.” Moody says the advantage of a systematic method like machine learning is that it removes the emotion from trading decisions. But at the same time, it’s capable of taking market psychology into account. “A lot of market behavior is actually somewhat predictable based upon human emotion,” Moody notes. “An appropriate statistical model should be able to capture that.” DIMITRI SOGOLOFF’S QUANT START-UP, Horton Point, has its roots in an act of generosity. In 2004, Sogoloff and Yuri Kuperin co-founded an academic program at St. Petersburg State University, in his native Russia, that trains science Ph.D.s to work in financial services. (Kuperin is head of the econophysics program there.) Sogoloff began wondering about the relationship between finance and pure science — and whether there was a systematic way to capture it without using just statistics. According to Sogoloff, 46, the main flaw of most quant strategies is their heavy reliance on historical data. Even though no strategy works all the time, he explains, this backward-looking approach makes statistical arbitrage especially vulnerable to unforeseen market moves. “If there is an event that is not captured in the prices today, the statistical arbitrage box will completely fall apart,” he says. “It’s a fantastic tool, but that’s all it is. It’s a tool in what needs to be a larger toolbox.” At Horton Point, Sogoloff and co-founder Vladimir “If you can figure out where the money’s going and what it’s coming out of, you should be able to make some pretty decent dough.” 28 • INSTITUTIONAL INVESTOR’S ALPHA • MARCH 2008
For optimal viewing of this digital publication, please enable JavaScript and then refresh the page. If you would like to try to load the digital publication without using Flash Player detection, please click here.