Model risk management is evolving: regulation, volatility, machine learning, AI

Due to the impact of the Covid-19 pandemic, the world is now dealing with geopolitical uncertainty, heightened counterparty risk concerns and rising interest rates, all of which present new challenges for model risk managers. increase. Thomas Oliver, Head of Model Validation at Quantifi, explores how model risk is managed (MRM) the situation is changing in response to these challenges

What model risks are regulators most concerned about in 2023?

Due to the major failures of entities such as Archegos, FTXSilicon Valley Bank, Credit Suisse, Englandof prudential regulators (PRA) requires banks to ensure that they have completed adequate assessments of liquidity, credit risk and counterparty risk. This includes fraud due diligence and a satisfactory credit risk methodology to assess entities with high concentrations of operational exposure, such as crypto assets, sector-specific concentrations or asset-liability mismatches. .long-term concerns raised by we Federal Reserve and PRA Relevant to model estimates of climate change risk in long-term commitments such as infrastructure and mortgage lending.

As a general principle, MRM Established for several years through the Federal Reserve Board SR11-07 Based on guidance and a targeted review of the European Union’s internal models, regulators are now assessing banks’ internal model risks (model inventories, review life cycle, use governance). However, regulators are sympathetic to the additional risk management complexities facing companies during the pandemic and in the aftermath of the Ukraine invasion, with many extending the full Basel timescale. Ⅲ Compliance – e.g. until January 2025 England and EU.

Lessons learned from the pandemic

The sudden and severe economic shock from the pandemic has created a model risk challenge due to the magnitude of behavioral and market changes and the pace at which these changes occur. Decreased oil consumption during lockdown has helped the West Texan Intermediate (WTI) Futures that trade below zero.

Credit scoring models also struggled to value individuals and small businesses suffering from an income collapse but receiving unprecedented government support. EnglandFor example, despite a -9.7% reduction in GDP In 2020, fewer companies will go bankrupt in 2020 and 2021 than in the years before the pandemic.

As historical dynamics rarely reflect recent conditions, it is increasingly important to adjust model calibration procedures and ranges of use in a timely manner. Agencies need to be able to quickly monitor rapid performance changes and quickly adjust model processes and usage patterns. This requires good communication between model developers, validators, and other stakeholders, and an established backup business logic to handle the model degradation period.

How does machine learning affect MRM?

Machine learning has the potential to transform multiple areas. MRMOnce machine learning models are used in production, there are additional requirements to prove model robustness and explain model decisions. Complex machine learning models may require the use of decomposition analysis using techniques such as locally interpretable model-agnostic explanations, Shapley additive explanations, or dedicated interpretability model implementations. Sufficient operational and implementation procedures (enterprise machine learning operations) additional steps to ensure complex models have reproducible results, timely recalibration, and accurate alignment between methodology and production deployment. Evidence may also be required. More sophisticated models in areas such as natural language processing may also rely on external models (transfer learning) such as: GPT-4, adaptation to specific tasks rather than training from in-house start. In some cases, this means that modelers need to understand the risk of bias or imperfections, even if the original training data or calibration code is not visible.

On the validation side, machine learning uses multi-model automated machine learning checks to ensure that proposed models utilize all potentially relevant data and identify probabilities. , etc., promises to strengthen support for validators to systematically benchmark their models and performance. Enhanced.

If we Regulators limit the role of internal models for credit risk measurement. MRM?

Regulators have multiple concerns about reliance on internal models. Risks from complex models can be much more difficult to understand and mitigate than the known limitations of cruder but more transparent methodologies. For internal models to be sufficiently conservative, there must be internal capacity within the bank to evaluate the model and external regulatory capacity to ensure that internal control processes are robust. The desire of banks’ revenue departments to lend more or capture a larger market position creates incentive risks that hide favorable biases in model complexity.

from MRM Moving to a simpler standard model simplifies regulatory compliance. However, regulators recognize that forcing all banks to use a common modeling framework discourages innovation and reduces diversity of opinion on valuation and risk management. This can lead to market distortions that are now priced by the same regulatory capital considerations. Enforcement of generic and simplistic methodological prescriptions can exacerbate the impact of systematic model errors when all participants rather than a single bank utilize a common flawed model. we As such, regulators are weighing the benefits of transparency against the risks of market distortion. If we conclude that the difficulty of assessing firm-specific internal credit model conservatism is inherently too high, standardization is the best approach to ensure adequate conservatism across all regulated entities. could be.

how to recruit FRTB influence MRM Because of market risk?

Fundamental review of the trading book (FRTB) is part of Basel Ⅲ It has a particular focus on estimating the market risk of banks’ trading portfolios. Basel III has already led to significant increases in bank buffers between 2011 and 2021.

FRTB Allow internal model processing of positions only when there is a liquid market, to obtain realistic estimates of historical volatility and correlation for such positions. A position’s risk representation should capture sufficient elements of actual market gains and losses and, through backtesting, provide an accurate forecast of historical losses. Additionally, the testing horizon is extended to all available historical data, and less liquid positions should use larger change horizons. This significantly increases the amount of data that needs to be retained and increases the comprehensiveness of commodity treatment in the market risk methodology.

With these more standardized standards, FRTB Move the sufficiency of representation debate, historically determined by each bank and each regulator, to a more standardized format. When there is a change in market behavior, such as when various risk factors become dominant, or through position hedging, the bank builds material exposure to poorly modeled or omitted risk factors. Model risk still occurs if it means . Model validation can therefore focus more on model robustness to changes in market regimes, quality of underlying valuation models, and hedging behavior against intended model use.