IRB survey: Could AI reshape internal ratings-based models?

While IRB models offer a more sensitive and bespoke view of a bank’s credit risk, which can ultimately reduce capital requirements, they’re inherently complex to develop and maintain. They’re also subject to regulatory approval, which can take time and increase the cost of compliance compared to the standardised approach (SA).  

Meanwhile, IRB-banks face additional reforms through Basel 3.1 and the upcoming CRR III, including changes in eligibility criteria, enhanced technical requirements and extended timelines for low-default portfolios. Crucially, this includes Basel 3.1’s new output floor, set at 72.5% of RWA calculated using the SA. To establish where that floor sits, IRB-banks must effectively apply both models – putting greater pressure on resources and compliance costs moving forward.

With these challenges in mind, we wanted to understand whether banks plan to continue using the IRB approach and, if so, what support they need to streamline the associated workflows.  

What are the key findings?

IRB adoption is widespread for larger banks and its use is increasing. Of the nine firms we surveyed, just three (all with total assets under €50bn) currently use the SA, and two of those are applying to become IRB firms. Only one firm planned to predominantly use SA in the longer term. Other findings include:

Use of AI: 67% are in favour applying advanced modelling techniques to IRB models, for specific use cases

Long approval processes: For 78% of banks with approved IRB models, the process (from model development to approval) took more than 18 months

Regulatory feedback for non-approved models: 33% waited for over two years to receive regulatory feedback, and when they did, 33% weren’t clear on what was required to achieve approval

Four of our respondents are significant institutions – and these firms all apply the IRB approach for credit risk calculations, typically for more than 70% of exposures. Of the remaining five (non-significant institutions), just two currently use the IRB approach – but they use it as extensively as significant institutions.  

Looking ahead, 56% of banks expect their IRB usage to increase over the next few years. That includes 66% of firms currently using the SA approach (both already planning to migrate), and 40% of firms that are already using it, apply it to over 70% of exposures.

IRB models require a significant commitment

From our sample, the majority of firms had just one or more model approved and only two had three or more. However, getting to the point of approval is time consuming, with 78% of firms taking more than 18 months to develop the model and receive approval. For those that aren’t approved, a third waited over a year for feedback and a third waited more than two – but despite the long turnaround, 33% were still unclear on next steps to gain approval. Firms didn’t go into detail over the areas of regulatory concern during the approval process, but some mentioned model cyclicality, model documentation quality, effectiveness of model validation and divergence between PRA guidance and ECB’s expectations.

In addition to the approval timelines, IRB models can also be resource heavy with staffing estimates for development, monitoring and validation ranging from 1-75 full time staff members (with significant variation between firms). Moving forward, AI could be an integral tool to streamline workloads and increase the rate of development.

Despite the significant commitment and the resource demands, just 22% of respondents considered IRB models to be a key driver of strategic decision making, with 44% viewing them as tool for regulatory compliance.  

Banks recognise the benefits of AI

Looking ahead, firms were broadly keen to adopt advanced modelling techniques using machine learning or artificial intelligence, and this could be a critical area for future developments in modelling. 67% were in favour of using them for specific use cases – such as retail lending (46%) or high-default portfolios (31%) – while 11% of respondents were keen to see the techniques applied more broadly.  

The hesitancy around broader adoption could be due to concerns over the associated risks, with: 

• 32% citing concerns over model complexity and explainability
• 32% sharing concerns over regulatory compliance
• 24% worried about model overfitting
• 12% worried about model deployment and use by the business.

However, perceived benefits include more informed modelling decisions, enhanced stress testing and scenario analysis, and improved predicative accuracy.  

AI ambition outstrips capability

While many firms can see the benefits of advanced modelling techniques, firms aren’t rolling them out just yet. 11% of respondents had developed AI or ML models, but none had gained regulatory approval to use them. A further 22% were considering developing AI or ML models but hadn’t put them into action yet.

Despite concerns over compliance, firms were divided in terms of what the regulators’ role should be. 25% believed the industry should lead the innovation, while 31% believed that further regulatory guidance is needed prior to adoption. This was supported by 67% of firms noting the use of AI and ML as an area where they’d like to see more prescriptive regulatory guidance. 

Treatment of low default portfolios 

Under Basel 3.1, banks can no longer apply advanced IRB models to assess credit risk in low default portfolios and will need to use the foundation IRB or the standardised approach. As such, we wanted to understand what firms classified as a low default model, and how firms plan to approach the associated regulatory change:

• Firms typically classified a low default portfolio as one with 20 or less defaults a year (33%), or across the portfolio’s history (22%).
• 22% felt that partial standardisation is beneficial, but flexibility is needed, 11% felt the standardised approach was necessary to avoid risk underestimation, 22% felt the standardised approach is necessary to optimise costs.

That paints a fragmented picture, and the direction of travel may largely depend on which alternative techniques firms are currently using. This includes the Pluto Tasche method, expert judgment-based approaches and use of external data. Just 6% of firms use machine learning techniques, but this could change as firms become more confident in its application.

Moving forward 

Our survey reflects the challenges banks face around implementing IRB models, including significant operational strain, uncertainty over the regulatory approvals process and gaps in data. Despite these challenges, banks feel that the IRB approach still offers a clearer view of their credit risk. As such, banks want to keep using it but recognise that there is room for greater regulatory support and improved clarity throughout the approvals process.

Moving forward, AI and machine learning will undoubtedly play an important role in model development, validation and risk management – but we aren’t there yet. There is considerable work ahead, from the regulators and the banking sector, to establish what good looks like and robustly apply it to these crucial models. 

For more information about IRB model development, AI integration or supervisory engagement, contact Vivian Lagan.