The Gateway to Algorithmic and Automated Trading

#GDTRM15: Machine learning is the new C++

First Published 27th May 2015

Derivative quants are finding that finance is shifting from statistical modelling to big data analytics.

Global Derivatives Trading and Risk Management Conference

Global Derivatives Trading and Risk Management Conference

Industry leader discussion: We have the data, now we just need the right questions

Amsterdam - Over a decade ago, high frequency trading was extremely obscure, with a few shops in Chicago specialising in the strategy. Now, every financial firm has high frequency algorithms.

The same is going to happen with big data, and it will transform the way practitioners operate, said Marco Lopez de Prado, senior managing director at Guggenheim Partners and research fellow at Lawrence Berkeley National Laboratory, speaking on a panel at the Global Derivatives Trading and Risk Management conference last week.

Darrell Duffie, professor at Stanford University, said that machine learning is better equipped to deal with the complex web of human interactions in the markets than traditional statistical methods, because it can adapt.

But that comes with a major caveat, all to do with concerns over correlation, causation and overfitting.

"One must be very careful about using test methods to confirm what you think is a law when in fact the underlying environment is changing," Duffie noted.

When it comes to seeking alpha, the big problem is how to gain executable insight from knowledge gleaned off of huge sets of unstructured data, said Michael Hintze, CEO and senior investment officer of asset management firm CQS.

"The knowledge-insight point is the critical point for us in terms of creating alpha and that requires context, and that is where this big data might make a massive difference. It helps us contextualise and it also helps us ask the questions…to think about decisions taken," he said.

From left to right: Tobias Preis, Associate Professor, Warwick Business School; Michael Hintze, CQS; Marcos Lopez de Prado, Guggenheim Partners; Darrell Duffie, Stanford University

The conversation about the use of machine learning in finance is become more prevalent. Over the years that Steve Wilcockson, industry manager for Financial Services at MathWorks, has been attending, he's seen a "slow, incremental evolution". It's yet to become mainstream however.

"It will be interesting in the next five or six years - will it be predominant? Or will the current regime still dominate?" he said.

Stats to streams

Quants come from a school of traditional model-centric approaches, which is all about not having sufficient data and finding ways to extrapolate.

In today's world, data is streaming from all directions - historical data on demand, images, and text for example.

"It's a little bit of the mixing of the modern world with the post-modern world, or moving from one regime to another," said Wilcockson. "How do we move away from a small data model-centred world to a big data-driven world?"

The most important difference between machine learning and statistics is a conceptual one, said Igor Halperin, executive director of Quantitative Research at JP Morgan.

"Statistics tries to explain the world by specifying a model and then estimating its parameters, while machine learning, which mostly amounts to non-parametric ways of modelling, doesn't try to explain the world it only tries to mimic it," Halperin said. "The emphasis is on predictive power of the algorithm rather than on interpretability."

Moreover, some machine learning methods, such as neural networks or support vector machines, are not probabilistic.

Reinforcing learning

The landscape of machine learning tasks can be broken into three classes: supervised, unsupervised, and reinforcement learning.

Supervised learning includes tasks such as classification, regression and speech recognition, for example, and is used in banking and risk management for credit card fraud detection, anti-money laundering and algorithmic trading.

Unsupervised learning is more applicable for modelling time series, or for various models that deal with stock segmentation, or regime change detection.

Reinforcement learning, said Halperin, might be the most exciting topic in machine learning.

"Of all those approaches, it is the one that is closest to artificial intelligence itself, and closest to the way we think, probably the best proxy," he said. "People use it to (get) robots to operate, and I think some of the tasks done in finance and risk management can be formulated as reinforcement tasks."

That includes the portfolio manager level, he added.

It certainly seems to be the way the wind is blowing for the quants of the future. Bruno Dupire, head of quantitative research at Bloomberg, said that "all the resumes we get now, people mention machine learning and big data the same way they used to put C++ in the past".