OSFI E-23 and Automated Valuation Models: Considerations for Canadian Lenders

For Canadian mortgage lenders, an automated valuation model (AVM) is more than a number on a screen. The value estimate may inform collateral reviews, portfolio monitoring or other property-related processes. The relevance of the estimate depends on why it is being used, what information supports it, and how its limitations are understood.

OSFI’s final Guideline E-23 – Model Risk Management (2027), published September 11, 2025 and effective May 1, 2027, offers a framework for considering those questions. This article explores selected implications for AVM users; it does not attempt to turn the guideline into a property-valuation product checklist.

What changed in the final E-23 guideline?

E-23 describes OSFI’s expectations for model-risk management across federally regulated financial institutions (FRFIs). It addresses models throughout their lifecycle, including rationale, data, development, independent review, deployment and monitoring. Its application is risk-proportionate: an institution evaluates inherent model risk, rather than treating every model or use case as identical.

The framework is not limited to models built internally. It also covers relevant models and data obtained from external sources, such as third-party vendors. OSFI’s accompanying publication letter discusses documentation, monitoring and the challenge of reviewing proprietary third-party models.

AVMs are not singled out as a compulsory tool, nor does E-23 specify an approved vendor or a standard list of real-estate input fields. Whether and how an AVM fits into an institution’s model-risk framework depends on its use and risk.

Start with the AVM’s intended use

The same property-value estimate can play different roles. An institution might use one as a preliminary screening signal, as a supplemental review input, or in a process that substantially influences a credit decision. Those uses may create materially different consequences when a model is wrong.

A practical evaluation therefore starts with questions about purpose: What decision does the model inform? Which property types and regions are in scope? Is the model being used with other evidence and professional judgment? What happens when an estimate is unavailable, uncertain or inconsistent with other evidence?

E-23 places importance on a clear model rationale, inherent risk assessment and governance proportional to that risk. It is not appropriate to infer that every AVM output needs exactly the same review process.

Property inputs can be as important as the algorithm

A valuation estimate may depend on property characteristics, transaction evidence, location, temporal market indicators and transformations of those inputs. If a key property field is missing, incorrectly joined, outdated or estimated, that can affect how the result should be interpreted.

Consider a hypothetical example: a record lists a dwelling’s living area as 1,800 square feet, but the figure is actually inferred from imagery or an ambiguous public record. The same source may provide a building footprint, which measures a different concept. Treating the two as equivalent could introduce an avoidable modelling error.

OSFI’s Principle 3.2 addresses the suitability of model-development data, including accuracy, representativeness, traceability and timeliness. For property-related models, useful institutional questions include:

  • Definition: Is a reported area a measured floor area, a footprint, a finished living area or an estimate?
  • Source and transformation: What is known about origin, collection date, standardization and derived fields?
  • Coverage: Does the information represent the relevant geography and property types, including harder-to-model segments?
  • Uncertainty: How are missing, conflicting or modelled characteristics handled and communicated?

Those questions describe review considerations, not a claim that E-23 requires each field, metric or metadata element in every AVM product.

Independent review is distinct from a vendor’s performance claim

Model developers may report performance, but an institution’s model review needs to be objective and proportionate to model risk. E-23 Principle 3.4 addresses independent assessment of conceptual soundness, model performance and fitness for purpose.

For AVM evaluations, institutions may consider whether the evidence covers representative properties and time periods; whether it distinguishes training, testing and independent review; whether errors vary by property type, region, price band or market condition; and how unusual properties or missing data are handled.

A single aggregate accuracy percentage rarely answers all these questions. Nor does a confidence indicator, standing alone, demonstrate that independent validation has occurred. The meaning of each metric should be defined and its relevance assessed against the intended use.

External AVMs remain subject to institutional judgment

Using a third-party AVM does not itself resolve model-risk obligations. OSFI expects institutions to manage externally sourced models and related data within their own frameworks and refers to Guideline B-10 for third-party risk management. The proportionality of review and monitoring remains linked to model risk.

It may be reasonable for vendors to protect proprietary intellectual property. That does not remove the institution’s need to understand and assess the model sufficiently for its intended use. In practice, institutions and providers may need to discuss available evidence, limitations and access arrangements in a controlled setting rather than placing proprietary documentation on a public website.

Monitoring matters after implementation

An AVM that performed adequately on one historical sample may face changing conditions: different transaction volumes, new data sources, revised model versions or shifts in local property markets. OSFI’s monitoring principle addresses the detection of performance deterioration, significant changes and other events requiring escalation or review.

An institution’s actual monitoring approach should reflect the model’s role, available outcome data, risk rating and governance policies. The guideline does not establish a universal AVM update cycle or one threshold suitable for every use case.

Where Gnowise is relevant

Gnowise offers Canadian residential automated valuation models, property attributes and bulk data, and related analytics available through WISE1 API, subject to coverage and product scope. Organizations considering these services can review data availability, modelled outputs and relevant limitations for their intended workflows.

These products should be evaluated as inputs to an institution’s own decisions, not as an E-23 certification or substitute for independent model governance. For a concise overview of the regulatory context, see our OSFI E-23 model-risk management resource.

References

Editorial note: General informational commentary, not legal or regulatory advice, a representation of product compliance, or an endorsement by OSFI. Last reviewed November 2026.