Climate and nature risk are already credit risk. What happens when the past stops being a guide?

UK banks already carry material climate and nature risk on their balance sheets, largely through residential mortgages, commercial real estate, and other income-producing property. The value, liquidity, and cash-flow arising from these assets depend on physical conditions that are changing in ways most credit models were never designed to handle. We often hear about emergent risks, but these particular climate and nature risks are not emerging in any meaningful sense because are already embedded in portfolios; they are already interacting with borrower behaviour, and already shaping loss outcomes (even if they are rarely described in those terms).

The real problem is not a lack of awareness, nor a failure of disclosure. It is that the core machinery used to measure credit and collateral risk still leans heavily on historical relationships that quietly assume the future will behave broadly like the past. That assumption is becoming harder to defend.

This matters because credit risk, in its most operational form, is not an abstract concept. Banks estimate the probability of default (PD), which captures the likelihood that a borrower fails to meet contractual obligations, and the loss given default (LGD), which reflects how much is ultimately lost once recoveries on collateral are taken into account. Both quantities are grounded in historical data, typically spanning multiple economic cycles, and both assume that the relationships linking hazards, collateral value, borrower stress, and recovery outcomes remain broadly stable over time. For many macroeconomic risks this approximation works well enough. For climate- and nature-driven hazards, I’m not so sure it works anymore.

Physical hazards are shifting faster than historical data can capture

Physical hazards such as flooding, heat stress, drought-related subsidence, and water scarcity are not simply intensifying at the margin. They are becoming more frequent, more spatially concentrated, and more correlated with one another in ways that historical loss data struggle to represent. When hazard regimes shift, the mapping from exposure to damage, from damage to borrower distress, and from distress to realised loss shifts with them. Models calibrated on the past do not fail because they are unsophisticated, but because the statistical ground beneath them is moving.

Banks often counter this critique by pointing to through-the-cycle modelling, which is designed to smooth short-term volatility and avoid pro-cyclical swings in capital requirements. That design choice is defensible when dealing with business cycles, where deviations are expected to revert. It is much less convincing when the underlying risk process is changing directionally. Through-the-cycle models do not simply average noise. They also dampen signals that look unusual relative to the past, even when those signals reflect structural change. Climate and nature risk sit squarely in that category.

Valuation and insurance assumptions quietly embed climate fragility

The same assumption of stability is embedded, often more quietly, in property valuation. Valuations anchor credit decisions, loan-to-value ratios, and recovery assumptions, yet they remain largely tied to recent transactions, comparable assets, and observed market prices. Climate and nature variables rarely enter explicitly. Instead, there is an implicit belief that markets have already priced in all material risks. That belief only holds if buyers, sellers, insurers, and lenders share a clear-sighted view of future physical conditions. In many UK property markets, particularly those exposed to flood or subsidence, that condition is simply not met. When prices are backward-looking, any credit model that inherits those prices is backward-looking too, regardless of how complex it appears.

Insurance is often treated as the mechanism that closes this gap, allowing banks to assume that losses will be capped even if hazards intensify. This introduces yet another assumption of stationarity, this time about insurance markets themselves. Insurance does not respond smoothly to rising physical risk. Coverage is withdrawn, exclusions appear, and premiums spike in ways that can materially alter loss outcomes just when stress is highest. Credit models that treat insurance availability as stable, or binary, are assuming away precisely the dynamics that matter most under stress.

Stress testing looks forward - but the loss engine stays anchored

Climate stress testing has helped elevate these issues to board level, but it has not resolved the underlying modelling tension. Scenarios may be forward-looking, but the translation from scenario to loss is often performed using exposure, vulnerability, and correlation structures that remain largely fixed. Asset-level heterogeneity is averaged away. Concentration risk is muted. The scenario moves forward in time, while the loss engine remains anchored to the present. The result is a form of analysis that looks rigorous but rarely challenges pricing, provisioning, or capital decisions in a material way.

At this point, the discussion often stalls for a more pragmatic reason. Enterprise risk teams are accountable for numbers that must be validated, audited, and defended to supervisors. They are rightly cautious about introducing new quantitative elements into regulated models before methods, data, and governance are ready.

Climate and nature risk frequently enter organisations through sustainability-led initiatives that surface real issues but cannot yet support defensible model changes. Everyone agrees the risk exists, but there is no one wants to own outputs they cannot explain (which is very reasonable). Moreover, many well-intentioned approaches create more model risk than they remove. They assume that taking climate and nature risk seriously requires pushing new variables directly into probability of default or loss given default models. For most banks, that is neither realistic nor desirable in the short term – “You should insert added complexity into your historically high-performance models because we are confident they will no longer perform” is not an easy sell. It risks creating liabilities that sit awkwardly with existing governance frameworks and supervisory expectations.

In any case, I argue for a much simpler and easier-to justify solution.

A practical step: use climate and nature signals to challenge assumptions

A more workable approach is to use climate and nature information to challenge assumptions, rather than to replace models. At refinq, this has meant providing asset-level signals that highlight where physical climate risk assumptions may already be optimistic because the natural systems that buffer hazards are degrading. These signals are deliberately limited in scope. They do not attempt to model losses or predict outcomes. They are explicitly directional, transparent in construction, and positioned outside regulated models.

In practice, this kind of information is used to segment portfolios, prioritise review, and structure internal challenge. It helps risk teams identify where climate stress test results deserve closer scrutiny, where provisioning overlays rely on fragile physical assumptions, and where transition narratives quietly depend on environmental stability that may no longer hold. Crucially, it allows these conversations to take place within existing governance structures. Outputs are used to question, not to override. They strengthen internal challenge without creating a parallel shadow model.

For Chief Risk Officers, this framing matters. The issue is not whether climate and nature risk are real, but where existing models are implicitly assuming they are not. When hazard frequency and severity shift, probability of default rises as borrowers face higher operating, repair, and insurance costs, while loss given default increases as collateral values fall or become less liquid. At the same time, defaults become more correlated because hazards are spatially clustered, and recoveries take longer due to repeat damage and insurance friction. These effects compound, and they do so in ways that historical averages systematically understate.

At that point, climate and nature risk are no longer long-term concerns. They are present-day pricing errors. Treating them as such does not require pretending the science is finished. What it does require is a recognition of where assumptions embedded in current models are most likely to break, and the creation of governance space for that recognition to influence judgement (before losses or supervisors force the issue).

We are entering a different phase. Long-dated property assets financed with short-dated assumptions create blind spots. Climate and nature risk expose those blind spots because they break the idea that the past is a reliable guide. For senior credit risk teams and CROs, the most productive question is no longer whether climate risk matters. It is where, precisely, existing models are assuming that it does not.

References

Integrating climate risk into credit risk modelling

Impact of climate risk on banks and ECL (analysis of climate risk impact on credit models)

Bridging climate and credit risk (global survey of banking practices)

Incorporating physical climate risks into banks’ credit risk models (noting absence of consensus models)

Climate risks in banking (supervisory expectations and risk detection)