Stress testing illusions exposed as AI reshapes model risk

The credibility of stress testing frameworks in modern financial institutions is under growing scrutiny, as extreme events expose fundamental weaknesses in how models are designed, governed, and interpreted.

According to Tanveer Bhatti, Founder and Senior Adviser on AI Governance, Model Risk and Regulatory Readiness at Fintech Investor, the challenge is not simply technical, but philosophical.

“I saw this managing model risk in a G SIB during 2008,” Bhatti said. “The models kept outputting numbers right through the crisis. Validity collapses at the moment you need it most.”

Bhatti, who was formerly Group Head of Model Risk at Revolut, said the experience of both the global financial crisis and the COVID shock revealed a deeper flaw in how institutions rely on quantitative frameworks.

Models built on repeatable patterns are being used to assess events that are inherently unpredictable.

“You're asking models built on repeatable mass phenomena to price non repeatable events,” he said. “That's a category error disguised as a quantitative one.”

Instead of relying on point forecasts, Bhatti argues that institutions must design frameworks that are capable of failing safely.

This includes setting explicit uncertainty bounds, identifying key sensitivities, and implementing hard circuit breakers when outputs fall outside credible ranges.

The challenge becomes even more acute when considering interconnected risks. During periods of stress, financial systems behave less like isolated components and more like tightly coupled networks, where disruptions cascade across functions.

“If the stress framework doesn't model that chain, you're testing nodes and calling it a network test,” Bhatti said, pointing to the way margin calls, funding pressures, and client behavior interact dynamically.

Capturing these cascading effects requires a more pragmatic approach. Bhatti emphasized the importance of triangulation, using multiple modeling techniques to approximate complex system behavior rather than relying on a single methodology.

“Analytical rigor is useless if it can't be run and explained and governed,” he said.