It's earlyish on a Tuesday morning in a room off the 'buttery' (read dining hall) at Oxford's illustrious Christchurch College. Paul Bilokon, a former quant at Citi, Nomura and Deutsche Bank-turned data science and machine learning lecturer at London's Imperial College, is bustling around in full academic regalia. Matthew Dixon, a former academic in the computer science department at Stanford University and quant at Lehman Brothers and Barclays Capital, is about to deliver a class. The room is packed.
The 30 or so attendees are drawn from a sturdy cross-section of finance society. Sitting elbow to elbow are media-shy bank staff, hedge fund business development professionals, consultants, senior ETF traders, portfolio managers and heads of fixed income technology.
They've come from Asia. They've come from Europe and they've come from the Americas. And they've each paid upwards of £2k ($3.6k) to hear Bilokon and Dixon expound upon the application of machine learning in financial services.
"I'm here to deepen my understanding of neural networks," says Joachim Tigler, head of exchange traded fund (ETF) trading at ADG Trading and a former Deutsche Bank managing director and head of the fixed income derivatives group at the German bank. Machine learning can help clean and structure data, optimize execution and pick stocks, Tigler adds, knowledgeably.
The people in the room are just the tip of the iceberg. Bilokon, Dixon and Saeed Amen (now head of Cuemacro, a consulting and research firm focused on systematic trading) are founders of the Thalesians, a self-described 'consultancy and think tank' comprised of, around 2,500 'dedicated professionals working in quant finance, economics, mathematics, physics and computer science.' This particular machine learning seminar is just one of their events: there have been 15 others this year alone. Most but not all were related to finance: others covered tumour formation, quantum computing, and energy bills.
Machine learning is a proudly academic area. Bilokon himself is fully credentialled: he has a masters in mathematical finance from Oxford University (distinction) and a masters (mathematics and computer science) and PhD (mathematics and computing) from Imperial College. But the Thalesians aren't just about rarefied academia: a poster at seminar advertises their tagline - 'It’s easy for philosophers to be rich if they choose.'
Whether Bilokon et al are "rich" is open to question, but some in the machine learning community certainly are. "Renaissance has been doing this for a long time," says Dixon at one point during his presentation on the application of neural networks in finance and the frontiers of machine learning. The renaissance he's referring to is Renaissance Technologies, the quantitative hedge fund whose founder, Jim Simons, is thought to be worth $18.5bn.
Dixon uses part of his session to clarify how machine learning is different to other forms of quantitative finance. For example, he points out that accuracy should not be used as a measure of performance in supervised learning models (models where you know the right answer and teach the machine how to achieve it). If a machine learning model is rewarded for accuracy, it will focus on the 98% of times that it's right and ignore the 2% of times that it's wrong. "You will get supervised learning that's very good at predicting zeroes because if has seen a lot of them, but which won't know what to do with the 1s [which it's seen fewer of]," says Dixon. "In finance, the tail events are quite important. It's easy to brush them under the carpet because they are inconvenient, but you need to face them head-on," he adds.
Dixon also expounds upon unsupervised learning, where models aren't told what the correct answer is, and on reinforcement learning. The latter requires dynamic programming so that the machine is penalized for making a wrong decision and rewarded for making a good one. Q learning is a subset of reinforcement learning where you look at the probability distribution of responses to various actions. This kind of machine learning is distinct from statistics, explains Dixon: "Reinforcement learning comes from optimization and decision science."
For this reason, the best machine learning practitioners in finance may not be statisticians at all. - Dixon says reinforcement learning is best suited to financial markets. "In supervised learning, you don’t account for the fact that your decision changes the state of the world. You are observing the data and making decisions – nothing about your decision feeds back into the market. This isn’t the case in the markets, so by by definition reinforcement learning seems to most appropriate."
Tigler, who knows Bilokon from Deutsche Bank and has been to various Thalesians events, thinks the seminar was was very well worth it. Now that artificial intelligence is the new hot thing, he says a lot of events are "superficial", but that the Thalesians are distinct: "There are people here with significant experience and influence who understand the mathematical foundations of machine learning and its practical applications in finance."
And if you can't attend a $2.6k seminar to get with machine learning? Dixon recommends a book: Marcos López de Prado, the founder of Guggenheim Partners' Quantitative Investment Strategies (QIS) business, has just published a new book 'Advances in Machine Learning.' "He's against heavy mathematics and is very practical," says Dixon. "The book has a lot of wisdom on the practicalities."
Have a confidential story, tip, or comment you’d like to share? Contact: firstname.lastname@example.org
Bear with us if you leave a comment at the bottom of this article: all our comments are moderated by human beings. Sometimes these humans might be asleep, or away from their desks, so it may take a while for your comment to appear. Eventually it will – unless it’s offensive or libelous (in which case it won’t.)