ModelOps for Structural Analytics
Over the past few years, MLOps has become an essential discipline for organisations deploying AI/ML models at scale. Inspired by DevOps, it focuses on the infrastructure, operating model, and toolkits needed to manage models as enterprise assets. While this has helped industrialise AI/ML pipelines, most banks and financial institutions still rely heavily on deterministic, rules-based models in critical areas like credit scoring, regulatory capital, accounting valuation, and scenario analysis — a domain we refer to as ‘structural analytics’.
Built Differently, Used Differently
In structural analytics, models do not just fall out of a training process. They have to comply with regulation, established practice, and business intuition, while also achieving an appropriate level of statistical performance. Models tend to be deterministic: when run with the same inputs, they should always produce the same outputs. Finally, these models tend to be subject to the highest levels of internal and external scrutiny
In this context, many core MLOps principles — such as versioning and deployment discipline — are still relevant. But ModelOps for structural analytics needs to address three additional priorities:
Repeatability: Systemic versioning needs to apply across entire workflows — not just model code, but data, configuration, and business logic — to ensure that model outputs are fully traceable and reproducible.
Documentation: To meet existing governance frameworks — whether MRM, financial control, or data lineage — production systems should be help to generate documentation describing how models are configured, run, and changed over time, rather than relying on post-fact manual documentation processes. This is the counterpoint to ‘explainability’ in AI/ML models.
Experimentation: Users need the ability to perform analytics without impacting the production deployment. This isn’t the same free-form flexibility that is so valued in data science — it’s about being able to flexibly experiment within the structure of a deployed model.
Conclusion: ModelOps for Structural Analytics
The highly governed models that are used in structural analytics may not be ML/AI — but they are critical business assets and need at least the same level of operational discipline. A tailored ModelOps approach for structural analytics should combine the rigour of MRM, the technical discipline of MLOps, and the realities of how these models are built and used. Without it banks will remain stuck with outdated tools and processes that comply on paper, but lack efficiency, agility, and real control.
