Machine learning, stress tests and Sonia

JP Morgan data scientist on mining and machine learning

Asset management arm looks to trawl internal data for investment edge

European banks face ‘bottleneck’ to complete EBA stress test

New accounting rules and supervisor demands squeeze teams prepping for 2018’s exercise

EIB shrugs off term RFR worries with Sonia bond plan

Issuer to use daily compounded, backward-looking rate with time lag for sterling benchmark

COMMENTARY: The limits of the machine

It’s not going too far to suggest that, in 10 or 15 years’ time, the job of a risk manager will be largely that of a machine overseer. Many risk managers may feel this is already the case. Our coverage this week highlights some of the advantages and the dangers of the change.

At JP Morgan Asset Management and Citadel, the development of big data technology is producing tempting results – the companies’ existing stocks of data on credit history and investment performance can now be exploited as never before. The implications are interesting; if the edge for asset managers of the future comes from sophisticated processing of existing information – what the intelligence community would call open-source intelligence – then the incentives for market abuse through insider trading may be reduced. What would be the point in stealing private information when, 90% of the time, you can reach the same conclusion from sophisticated processing of public information? On the other hand, if monumental amounts of big-data processing power are the key to success in the markets, the barrier to entry to the asset management business will be high. (Not that it isn’t already.)

But there’s a bigger caveat, especially around suggestions that machine learning systems should take over responsibilities for designing strategies. As Citadel’s chief risk officer Joanna Welsh and her counterparts warn, it could become all too easy for risk managers to get suckered into uncritical dependence on their machines. In aviation, this phenomenon is known as “keeping your head inside the cockpit” – and many air disasters have occurred as a result, when sophisticated machinery erodes the crew’s responsibility to maintain situational awareness, pushing them into a state of cognitive tunnelling. The more sophisticated machine learning systems get, the more difficult – and the more important – it becomes to ensure there are humans watching over them. And this will be especially true during times of crisis, when the systems go outside the limits of their training sets; unfortunately that’s also when it will be especially difficult, as pressure on the human overseers mounts.

STAT OF THE WEEK

Securitisations made up just 0.7% of European banks’ total risk exposures in 2017, down from 3.2% in 2008, a collapse that testifies to the ongoing stigma securitisation techniques face across the European Union in the wake of the financial crisis

QUOTE OF THE WEEK

Cryptocurrencies live in their own digital, nationless realm and can largely function in isolation from existing institutional environments or other infrastructure… As a result they can be regulated only indirectly – Bank for International Settlements