Mr. Papageorgiou, who works on a news-based signal (which analyzes news coverage of companies) for the firm's regular quant engine and oversees frontier markets investments, said the aim of the machine-learning model is to better predict where to find the best equity returns.
“As return prediction becomes more accurate, we would expect returns to improve,” he said. Acadian executives say the new model will go live within the next year; they haven't decided whether to set up separate machine-learning-based strategies or to have the new model work alongside existing models.
Unlike typical quant engines based on linear regression analysis, Mr. Papageorgiou's system uses machine-learning pattern-recognition techniques. It uses the same inputs as linear regression analysis, such as price-to-earnings or price-to-book ratios, but puts them together in a very different way.
“Machine-learning systems throw out assumed relationships between (stock) fundamentals and returns ... This would be a different type of forecasting model” from what Acadian or other quant managers have used to date, he said.
For example, the model would learn to recognize return patterns such as the value premium on its own, but understand it in a much more complex way than a person would.
“The only thing that's pre-specified in this approach is the method by which it learns,” Mr. Papageorgiou said.
Because it's always learning, the model would be able to adjust to unusual environments — such as a financial crisis — better than its predecessors. “We did see that the model and system seemed to be able to recognize what was going on in the markets” through the global financial crisis, Mr. Papageorgiou said. For example, it scaled back on quality stocks during the 2009 rally, when quality didn't matter.
“We would believe — and the data would suggest — that with this type of system in place, the forecasting model would not necessarily have been subject to the significant quant underperformance we saw” in the crisis, he said.
However, Mr. Papageorgiou warns the machine-learning model is “not a panacea,” and that its forecasting capabilities might be complementary to those of linear regression models.
Mr. Papageorgiou said there's no formula for generating innovative ideas. “Personally, I start by imagining what I'd like to do without focusing on restrictions or on the practicality,” he said. “This typically entails pushing the limits of what's considered normal within a particular field.”
He suspects “relatively few of our competitors are actively working” on machine-learning models, but added “this will ... over the long term, gain additional interest from the investing community.”