Skip to content
Glossary

Model validation

Model validation is the independent assessment of whether a model is conceptually sound, correctly implemented and performing as intended for its specific use.

Validation provides effective challenge to model development and use. Under PRA SS1/23 it covers an independent review of conceptual soundness and suitability, verification that the model is implemented correctly, ongoing performance monitoring, and periodic revalidation at a frequency consistent with the model’s tier. MaRisk requires regular validation that examines the accuracy, stability and consistency of model results, together with the limits arising from assumptions and data.

Boards rely on validation as the main assurance that model outputs used for credit, capital, pricing or compliance decisions can be trusted. It only provides that assurance if validators are independent of developers and owners, have the standing to challenge, and see their recommendations acted upon. FINMA observed in its guidance on AI that development and independent review were not always clearly separated.

Common pitfalls are validation reduced to replicating calculations, reliance on vendor documentation without own testing, findings that stay open for years while the model remains in use, and revalidation schedules that ignore changes in data, portfolio or use.

In practice

  1. 01
    Separates the reporting lines of validators from those of model developers and owners.
  2. 02
    Covers conceptual soundness, data quality, implementation, outcomes analysis such as backtesting and benchmarking, and limitations.
  3. 03
    Rates findings by severity and links each one to an owner, a deadline and any conditions on model use.
  4. 04
    Sets revalidation frequency by model tier and triggers an ad hoc validation after material changes.
  5. 05
    Tests vendor and AI models with the bank’s own data rather than relying on supplier claims.
In MRM-Solution

How MRM-Solution supports it

  1. 01Keep every model in one inventory with owner, tier, EU AI Act risk class and re-validation dates, linked to the committees that approve its use.
  2. 02Treat AI findings as proposals: a verdict counts only after a person has confirmed it against the exact texts.
Book a demo

Sources

Too good to be true? Then we should talk.

30 minutes. Your use case, maximum impact.

info@mrm-solution.com
Area of interest