Machine learning, Libor and an inside view of Jump

Machine learning hits explainability barrier

Banks hire AI industry experts in face of growing regulatory scrutiny

Jump: inside the secretive e-trading giant

Execs at Chicago prop firm wish whole world was a Clob, but as bilateral volumes rise, they’ve decided to go with the flow

More carrot, less stick in US Libor transition

Risk USA: US regulators take softer approach than UK counterparts

COMMENTARY: A Jump on the competition

Our look this week at Chicago prop giant Jump Trading, the low-profile king of the worldwide US Treasuries markets, provides more than just an insight into a major prop trading player. Jump’s story stands as an example of how scope rather than speed is now the foundation of power in data-driven industries like finance – a shift that has important implications for the future of artificial intelligence research, financial markets and much else.

AI and machine learning are becoming ubiquitous in all areas of financial services, to the point where a major problem is explaining the models to both senior management and regulators – a quest some machine learning advocates say will prove futile and should be abandoned. 

The shift from speed to scope is one big reason why explaining the models is proving hard. Modern AI systems are more than simply fast processors. They use huge volumes of data (and huge amounts of processing power) to train themselves to process conclusions from further huge volumes of data. They’re not just a human analyst speeded up – they are operating at a far broader scale than any human brain could manage on any schedule. Jump Trading’s competing internal models rely on the company’s unparalleled access to market data, derived from its dominant market position, to train and modify themselves.

All this should sound familiar. It’s very similar to the process by which dominant internet players such as Amazon, Google and Facebook have become successful – the network effect. The more searches Google conducts, the better its algorithms become (because it can observe which of the many results to each query its users click on). The more people are on Facebook, the stronger the pressure for others to join. Self-driving car developers are following the same route – the training data here is road miles driven (autonomously or by a human driver) and the best algorithm will become the most popular, allowing it to accumulate more experience and get better faster than its rivals.

Our reporting – and the regulators’ attention – is on the problems of understanding these immensely complex systems and assessing their risks, which may be beyond the abilities not only of the average senior manager but of any human brain. But the analogy with other internet giants raises another risk – that of dominant market players entrenching themselves unassailably by using their information advantage. 

STAT OF THE WEEK

Italian banks lost €9 billion (8.9%) of CET1 capital on the transition to IFRS 9 in January this year. French lenders lost €4.9 billion (1.6%), and Spanish firms wrote off €3.6 billion (2.7%).

QUOTE OF THE WEEK

“The EU digital sales tax is targeted at the big US digital businesses, including auction websites and other online marketplaces. But the definition is so broad that I’m concerned it also taxes financial market trading venues and other financial market infrastructure” – Dan Neidle, Clifford Chance