At the close of 2025, Zillow, the largest residential property listing platform in the US, quietly withdrew its climate risk scores from property listings.
According to Consilient, the move followed complaints from estate agents and homeowners who argued that prominently displayed climate ratings were distorting valuations and undermining transactions.
The platform had embedded climate exposure metrics – including flood and wildfire risk – directly into individual listings. Prospective buyers could see the score instantly. In some cases, sales activity slowed. In one high-profile example, a Florida mansion listed at $295m – then the highest asking price in the US – underwent repeated price reductions before being withdrawn. The flood risk had not shifted. The property had not relocated. What changed was the visibility of the score and the weight buyers attached to it.
That moment marked a turning point. The climate score stopped being a neutral informational overlay and became a market actor. When model outputs are surfaced at the individual-asset level and placed directly in front of decision-makers, they do not merely inform opinion. At scale, they influence pricing, liquidity and behaviour.
Regulated financial institutions have long operated under this assumption. When models determine who receives credit, on what terms, or whether access is denied altogether, governance frameworks require far more than technical validation. Outputs must be explainable, challengeable and defensible at the level of the individual. Once economic outcomes are affected, oversight intensifies.
If a climate risk score with the potential to move asset values were deployed inside a regulated bank, the first step would be a material impact assessment. Could it influence lending, collateral valuation or pricing? If so, it would likely be categorised as high impact, triggering senior accountability and formal risk sign-off.
Independent validation would extend beyond predictive accuracy. Reviewers would examine conceptual soundness, sensitivity to assumptions and how uncertainty is handled. Climate risk modelling, in particular, involves probabilistic assessments layered over complex environmental data. Presenting a single, context-free score for a specific property risks creating false precision where uncertainty remains significant.
Crucially, where model outputs affect valuation or access to finance, institutions are expected to explain outcomes in plain terms. Drivers such as elevation, flood pathways or historical data would need to be surfaced, alongside a clear route for dispute and correction. Monitoring would continue post-deployment, tracking unintended consequences such as disproportionate impacts on certain regions or communities. Evidence of consumer detriment or market distortion would typically prompt recalibration or suspension before withdrawal became necessary.
In the US, statutes such as the Fair Credit Reporting Act and the Equal Credit Opportunity Act formalise these expectations, while the Consumer Financial Protection Bureau has reinforced that they apply even where complex models are involved. In the UK, the Financial Conduct Authority emphasises fair outcomes under its principles-based regime. Within the EU, the EU AI Act classifies many financial services applications as high risk, attaching governance and redress requirements accordingly.
These standards exist for a simple reason: once a model shapes value or access at the individual level, aggregate performance is no longer enough. The relevant question becomes whether a specific outcome can be defended for a specific person, based on the information available at the time.
Notification and complaints processes rarely suffice. They address harm after it occurs. In high-volume, model-driven environments, that lag creates exposure. Explanations may be too abstract for meaningful challenge. Corrections can be slow and inconsistently fed back into model recalibration. A system that performs well on average can still generate concentrated harm at the margins.
Financial services encountered these governance pressures early because models began determining credit access and pricing decades ago. Insurance adopted sophisticated modelling as well, but underwriting and claims decisions traditionally retained human discretion. Models informed judgement rather than replacing it, absorbing some error before it reached consumers. As automation accelerates and human intervention diminishes, similar governance tensions are emerging there too.
The lesson extends beyond finance. Real estate platforms, climate scoring systems and other market-facing technologies are increasingly operating in high-stakes territory. Once scores move markets rather than simply describe them, governance expectations shift.
Zillow’s experience illustrates that transition. The issue was not solely model quality. It was the absence of robust mechanisms for uncertainty handling, explanation, challenge and redress once the scores began shaping individual asset outcomes. Financial services offers a tested blueprint: high-impact models can operate at scale, but only if they remain individually defensible and subject to continuous oversight.
Copyright © 2026 RegTech Analyst
Copyright © 2018 RegTech Analyst
