Skip to content

Content Library

Back to Articles

Article September 30, 2025

Regulatory scrutiny casts stress testing as banks' strategic compass

Regulatory scrutiny casts stress testing as banks' strategic compass

Ushnish Banerjee, EMEA QAG Vice President, Morgan Stanley

Financial institutions are rethinking stress testing as a strategic tool, using AI, smarter controls, and boundary awareness to unlock real value.

  • Stress testing evolving from compliance task to strategic tool for capital planning risk appetite and resilience
  • Bridges gap between IFRS9 CECL expected loss models and conservative Pillar1 with Pillar2 covering all risks
  • Inefficiencies in scenario design data integration validation and reporting cause hidden conservatism
  • Regulatory scrutiny and crises drive new modeling approaches as old calibrations show fragility
  • AI ML aid in data quality anomaly detection and reverse stress testing but need human oversight
  • Defining monitoring and communicating model operating boundaries prevents misuse and misinterpretation
  • Excess conservatism weakens strategic value but integration with business planning is improving
  • Strong governance advanced analytics and awareness of model limits make stress testing a shield and compass

Stress testing is undergoing a fundamental shift from a compliance ritual to a strategic enabler for capital planning, risk appetite and business resilience, according to the head of model risk quality assurance at a large international bank. 

“Treat it as a shield and a compass, not a checkbox,” the executive told delegates at this year’s Financial Crime USA conference, hosted by CeFPro.

He framed the change against a backdrop of economic shocks, regulatory tightening and rapid technological disruption, saying that a mindset change needed to be at the centre of evolution. 

Situated between highly accurate, unbiased expected loss tools such as IFRS 9 and CECL, and the conservatively framed capital models under Pillar 1, stress testing now bridges two philosophies. 

Unlike Pillar 1, which allows some electivity between standardised and internal approaches, Pillar 2 stress testing must address the full breadth of economic risks without exception. 

“That breadth drags in complexity across finance, risk, business and technology,” the speaker noted.

This complexity has exposed inefficiencies across the end-to-end process. The challenge is not only model sophistication but the framework around it - scenario design, data integration, model inputs, output validation and reporting. 

Many firms, the speaker argued, still lack a scalable, repeatable process that allocates resources optimally and avoids hidden conservatism embedded in assumptions. 

He warned that such inefficiencies can blunt commercial value, producing over-cautious results that fail to support strategic capital planning.

Technological change is forcing the pace. Expanding model inventories - rising by 10 to 25 percent annually - combined with scrutiny under regimes such as the PRA’s SS3/18 and SS1/23, have pushed banks to explore new techniques. 

Recent crises, from the pandemic to rate shocks and geopolitical turbulence, exposed the fragility of models calibrated in unusually stable periods. 

“Calibration built on calm seas will not steer you through a storm,” this senior manager said.

Artificial intelligence and machine learning are adding value in targeted ways rather than replacing core models. 

Direct use in headline stress models remains limited given explainability, supervisory compatibility and tight delivery timelines. However, AI is proving effective in supporting tasks such as data quality assurance, variable selection, anomaly detection and reverse stress testing. 

“Use AI to clean the mirror, not to change your face,” the executive said, emphasizing model validation, regulator engagement and accountability.

Automation of anomaly detection is a leading opportunity. Traditional z-score checks can be rigid, while isolation forests or density-based clustering give adaptive, data-driven ways to flag potential errors in inputs and outputs. 

Events you may be interested in

View all of our upcoming events

Sign Up to Our Newsletter

Sign up

Connect Plus

Unlock more with Connect Plus membership

This content is available to Connect Plus members.