After 2008, stress tests became central to banking supervision. They ask a blunt question: if the economy takes a severe hit, does a bank still have enough capital to keep lending? Regulators now ask a version of that question about climate change. The answer is proving harder to pin down than many expected.

Since 2021, supervisors have run climate scenario exercises, among them the Bank of England, the European Central Bank and the US Federal Reserve, with others following from Canada to Australia. Their starting point was Mark Carney's 2015 warning about the tragedy of the horizon. Climate risks, he argued, arrive beyond the planning window of most financial institutions, and by the time they become a clear threat to stability, it may be too late to manage them.

The early results are now in, and they are instructive in an unexpected way. They say less about how much banks stand to lose than about how poorly the financial system can currently see the risk. For anyone pricing, lending against or regulating a warming economy, the findings matter. So do the blind spots.

Most climate stress tests build on scenarios from the Network for Greening the Financial System (NGFS), a coalition of central banks and supervisors. The scenarios fall into three broad families. In an orderly transition, carbon prices rise early and predictably. In a disorderly transition, action is delayed and then abrupt. In a hot house world, policy stalls and physical damages accumulate.

These families capture a core trade-off. Transition risk includes stranded assets, repriced carbon-intensive firms and falling collateral values, and it peaks when policy moves late and fast. Physical risk includes floods, heat, drought and storms, and it peaks when policy barely moves at all. No pathway avoids both entirely. Supervisors therefore typically run several scenarios side by side rather than selecting a single severe case.

Mechanically, the exercises translate climate variables into economic ones. Carbon prices and emissions pathways feed integrated assessment models, which produce GDP, energy prices and sector-level output. Banks then map these outputs onto borrower-level probabilities of default and loss given default. For their largest exposures, they often do this counterparty by counterparty. Physical risk modules overlay flood, storm and heat hazard maps onto mortgage books and the locations of corporate assets.

Two design choices stand out. First, horizons stretch far beyond convention. The Bank of England's 2021 exercise ran for 30 years, compared with the three to five years typical of capital stress tests. Second, balance sheets are often held static, or only limited management responses are permitted, which isolates the underlying risk before any adaptation. Each choice makes the exercise tractable. Each also shapes, and constrains, what the exercise can reveal.

Takeaway

A climate stress test is less a forecast than a structured argument about which assumptions matter most. Choosing a scenario is itself a decision about which risks you are prepared to see.

On the surface, the headline finding is reassuring. The Bank of England concluded that climate losses would not threaten the solvency of UK banks and insurers. They would instead act as a persistent drag on profitability, of roughly 10 to 15 percent of annual profits. Losses were meaningfully lower in the early action scenario than in the scenarios where action was delayed or never came, reinforcing the economic case for an orderly transition.

The ECB's 2022 test produced a similar picture with sharper caveats. Combined credit and market losses for the 41 banks that submitted projections came to around €70 billion. The ECB stressed that this figure significantly understates true exposure, because the scenarios excluded second-round effects and covered only part of banks' portfolios. It also found that more than 60 percent of banks' interest income from non-financial corporates came from greenhouse-gas-intensive industries.

Sectoral and geographic concentration is the clearest signal. Transition losses cluster in fossil fuel extraction, utilities, heavy industry, and parts of transport and agriculture. Physical losses cluster by place: coastal real estate, flood plains, and regions prone to drought or wildfire. The Federal Reserve's 2023 pilot with six large banks highlighted real estate exposure in hurricane- and flood-prone areas. It also showed how heavily the banks' loss estimates depended on assumptions about insurance coverage remaining available and affordable.

Perhaps the most consequential result was qualitative. Supervisors repeatedly found that banks lacked borrower-level emissions data, granular location data for collateral, and internal models linking climate variables to credit risk. Many relied on sector averages or third-party proxies. In effect, the exercises measured capability as much as exposure, and capability was often the weaker of the two.

Takeaway

Modest aggregate losses can coexist with concentrated vulnerability. The real question is not whether the system breaks, but which borrowers, regions and lenders absorb the strain first.

Climate stress tests inherit the limits of the models beneath them. Damage functions, the equations that convert temperature rise into economic loss, descend from work by William Nordhaus and others. Critics have long argued they produce smooth, modest damages even at high levels of warming. When the NGFS updated its damage function in 2024 using newer empirical research, estimated physical losses rose substantially. On paper, the same climate had become far more costly.

Several features of real climate risk sit largely outside the frame. Tipping points are difficult to model. So are compound events, such as a drought coinciding with an energy price shock. Second-round effects are also largely missing. Insurers retreat from a region, collateral values fall, credit tightens and local economies slow, in a chain no single model fully traces. Contagion between institutions, and interaction with wider macroeconomic shocks, are usually excluded or heavily simplified.

Static balance sheets cut both ways. They stop banks from assuming they will exit risky sectors just in time, which is useful discipline. But they also ignore that collective exit can itself be destabilising, pushing down asset prices for everyone at once. They overlook the possibility that adaptation investment could reduce losses. Over a 30-year horizon, a frozen balance sheet is less a projection than a thought experiment.

The appropriate reading, then, is comparative rather than absolute. A €70 billion figure is not a forecast. It is a lower bound, conditional on chosen assumptions. The more useful outputs are relative: which scenarios hurt most, which portfolios are most sensitive, and where data gaps are deepest. Read this way, stress tests sharpen risk management. Read as precise capital calibrations, they risk producing false comfort.

Takeaway

When a model produces a small number, ask what it was designed to leave out. In climate risk, the omitted variables are often the ones that matter most.

The first generation of climate stress tests has done what early exercises usually do. It has framed the questions more clearly than it has answered them. Aggregate losses look manageable, but those estimates rest on models that likely understate tail risks and on data that banks are still building.

The next phase is moving toward shorter horizons, sharper physical risk scenarios and closer integration with mainstream risk frameworks. Commitment among supervisors has also diverged, which may itself create uneven exposure across jurisdictions.

For practitioners, the lasting value lies in the discipline of tracing how climate transmits into credit, collateral and capital. The numbers will keep changing. The habit of looking is what endures.