Risk.Net | By: Louie Woodall | August 17, 2017:
US regulators are raising concerns about the use of machine learning techniques to assess contagion risks in bank model networks.
Last year, certain entities supervised by the US Federal Reserve were asked to analyse their aggregate model risk – essentially the interactions and dependencies between various risk and pricing models. Banks responded by experimenting with advanced computational techniques to understand model interconnectedness, including machine learning, network theory and probabilistic graphical models.
However, regulators are cautioning banks that these approaches lack transparency and could obscure the true extent of their vulnerabilities.
“Regulators want transparency and you cannot hide behind a complex tool,” says Nikolai Kukharkin, a senior risk manager and the former global head of model risk management and control at UBS. “There is always concern if you don’t understand something and lose the intuition behind it. If you apply machine learning to assess model risk, this algorithm is itself a model and you have to prove that it works.”
In conversations with banks, the Fed has said it does not want them to develop sophisticated approaches that generate singular measures of their aggregate model risk, as these can be misleading.
To read full article – please click here.
