Quant funds are using machine-learning technology such as image recognition in their race to pick datasets for new investment strategies ahead of rivals.
Funds are expected to spend $7 billion a year by 2020 on so-called alternative data, such as satellite imagery and social media information, according to consultancy Opimas. But selecting datasets that contain the seeds of workable investment programmes can take months.
If funds lose time in processing new information, they risk rival funds discovering it too, said Elliot Noma, managing director of Garrett Asset Management, a commodity trading adviser based in New York.
“If someone else also figures out a feed is good, it becomes a race,” Noma said. “Not only [is the question] how quickly can I get the data, but how quickly can I process it and make the trade.”
Increasingly, funds aim to “fail quickly”, Noma said. If machine-learning techniques find nothing in the data, the fund can move on to analysing new datasets.
Noma was speaking at Risk.net’s Quant Summit in New York on July 11.
Four-fifths of investors are looking to buy alternative data, research firm Greenwich Associates says in a recent report. However, quants at JP Morgan have warned of the “blind alleys” awaiting managers if they waste time on datasets that do not contain alpha – either because they generate signals that have too little investment capacity or decay quickly, or because the data is simply too expensive.
As many as eight out of nine efforts to build strategies based on alternative data fail, Wesley Chan, director of stock selection research at Acadian Asset Management, told Risk.net in April.
Machine learning has long been used by firms such as Facebook to speed up the process of analysing and categorising images. For example, if an object has a nose, eyes and a mouth, an algorithm can be trained to recognise it as a face.
The same technology – known as feature extraction – might be applied to satellite images to identify features that can be used in quant investment strategies such as numbers of vehicles in shopping mall car parks or customers entering retail outlets.
When there is no previous model, or the quality of the data is suspect and other people are looking at the data and competing, machine learning is a major advantage in terms of speedElliot Noma, Garrett Asset Management
Garrett Asset Management is in the early stages of exploring the applications of machine learning to unstructured data, but estimates it could cut the feature extraction process from years and months to days or even hours.
“Alternative data has certain characteristics, but when there is no previous model, or the quality of the data is suspect and other people are looking at the data and competing, machine learning is a major advantage in terms of speed,” Noma said.
Elsewhere at the summit, Dilip Madan, a professor of mathematical finance at Robert H Smith School of Business at the University of Maryland, said automating data-intensive activities is one of the better applications of machine learning.
That contrasts with the struggle to apply machine learning in predictive models. “Even if a model is successful and learns something, five minutes later, the market changes and you have to learn again,” he said.
Others questioned the value of alternative data for quant funds as well as the potential of machine learning as a processing tool. “I am doubtful machines can be taught to replicate the creativeness of human analysis,” said an equity market neutral portfolio manager at a New York-based quantitative hedge fund.
Satellite imagery could be useful for macro traders, he said, but is “a waste of time” for equities traders and funds with large portfolios. Satellite images tend to focus on a small number of datapoints, so cannot be modelled robustly, he said.
The week on Risk.net, October 6-12, 2017Receive this by email