LONDON (Reuters) - Financial firms must avoid using trading strategies developed by artificial intelligence that seek to profit from stoking market instability, a member of the Bank of England's Financial Policy Committee said on Tuesday.
"Neural networks could learn the value of actively amplifying an external shock. The FPC's old enemy, the forces of amplification, could arise in this new form," Jonathan Hall, an external member of the FPC and former investment banker, said.
Hall said it was viable for investment firms to develop what he called 'deep trading agents' - AI-powered strategies which operate semi-autonomously from human traders and which they only partly understand as they can change on the fly.
Ongoing academic research has shown a risk that such agents could collude with each other in a way that was illegal but hard for humans to detect or seek to fuel market instability. Alternatively they could be ill-equipped for turmoil.
Before putting AI models to work, financial traders needed to extensively test them against each other and make them comply with both the spirit and the letter of regulation, Hall said in a speech at the University of Exeter in southwest England.
"If trading algorithms engage in non-compliant, harmful behaviour then the trading manager will be held responsible."
Hall said the concerns represented his personal views, rather than those of the BoE as a whole, and were largely hypothetical for now.
However, he said there were parallels with trading strategies popular when he started his career in the 1990s, which led to the collapse of the Long-Term Capital Management hedge fund in 1998.
"Taken together these create performance and regulatory risks for trading firms, and explain the current caution about using neural networks for trading," he said.
(Reporting by David Milliken; Editing by William Schomberg)
By David Milliken
