What is model validation?
Model validation is the structured confirmation that a model is suitable for the purpose it will serve. It goes beyond model evaluation, which measures performance. Validation also examines whether the training data represents the population, whether the features are available at prediction time, whether the method suits the problem, where the model is known to be weak, and whether the documentation would let someone else reproduce the result. In regulated industries it is performed by people who did not build the model.
The term has a narrower technical meaning as well. Validation data is the slice of data held back from training to tune hyperparameters, and cross-validation is the technique of rotating that slice. Those are tools within validation. The broader discipline is what a bank examiner or an auditor means by the word.
Why model validation matters
Banks have been required to validate models for years under model risk management rules, and the same expectation is spreading through insurance, healthcare, and any company subject to the EU AI Act. The practical reason is older than the regulations. A validated model has a written statement of what it is for, what data it was tested on, and under which conditions it should be retired. A model without one is an unknown liability on the day it starts producing strange results.
Model validation in practice
A regional bank adopts a model that scores small-business loan applications. The validation team, separate from the data science team, re-derives the performance figures from the raw data, checks the applicant population against the bank’s actual footprint, tests the model on applicants from the last two downturns, and challenges the inclusion of two features that act as proxies for a protected class. The report records three limitations and a required annual re-validation. The model goes live with those limitations attached to it in the model registry.
Kirk Tech Solutions typically runs a lighter version of this review for companies outside regulated industries, since the questions are the same even when no examiner will ask them.
What to watch for
- Separate the people who validate from the people who built the model, even in a small company.
- Validate the data pipeline and the feature definitions, because most failures start there and not in the algorithm.
- Record limitations and the conditions that trigger re-validation, and store them with the model.
- Ask a vendor for its validation report, and treat its absence as a finding.