The Cracks Between Banks and Insurers
In a recent piece we traced how physical climate risk migrates from insurers, which can reprice or withdraw coverage every year, into banks that hold the same property as collateral for thirty years. The wall between the two sectors is opaque for the obvious reasons of separate regulators and commercially guarded data. There is a less obvious reason, and it is the one that matters for what happens in the next iteration of these industries. For most of the period in which this risk has been building, no one could put a number on it. The Bank for International Settlements said as much, noting that the central obstacle for banks was the absence of generally accepted models for quantifying these exposures.
That absence was not an oversight. It reflects a structural mismatch in how the two systems represent uncertainty. Climate science produces its outputs as probability distributions, ranges of likely outcomes across multi-decade horizons that depend on emissions pathways, regional hazard models and many other factors. A bank needs something the science does not naturally provide, a point estimate it can book this quarter, a number that flows into a credit rating, a capital charge, and a regulatory return. The gap between a distribution and a bookable figure is where the transmission channels have been sitting. These are real in theory but unmeasured in practice.
The bridge between the two now exists, and it is built largely out of asset-level data and enabled by AI and machine learning. The studies we share below have features that are easy to miss when you read them one at a time. Each uses computational methods to render a relationship that was previously intractable, and together they show a financial system learning to translate climate physics into the language of capital.
The Translation Problem
Turning climate risk into financial risk requires more than knowing which hazards are probable at which locations. It requires knowing which specific assets sit in the path of those hazards, a mortgaged property at a particular Florida ZIP code, a power plant on a floodplain, a commercial loan book concentrated in hurricane-exposed counties, and what damage to those assets does at the level of the loan, the issuer, and the portfolio. Before AI-driven methods, that chain of inference was close to impossible at any useful scale. Each link demanded manual matching, single-scenario assumptions, and datasets that lived in separate institutional silos and could rarely interact meaningfully with one another.
The work that has changed is the work of building that chain. It now runs in three stages, and each stage has a body of recent research behind it.
Resolving the Exposure
The first stage overlays physical hazard data onto financial exposure data at the level of the individual asset, and the immediate payoff is that resolution reveals concentration that averages hide. A 2024 study in the European Journal of Operational Research combined survival models with a gradient-boosting algorithm across a portfolio of 69,046 Florida mortgages at ZIP-code resolution to estimate how tropical cyclone intensity and flood exposure drive default. Extending the model to 2050 flood projections under a moderate emissions pathway, it found that the average loan sees its default probability rise by about one percentage point, while the loan at the 99th percentile sees a rise of seven. This tail is where the risk lives, and the tail is invisible to any model built on portfolio averages.
The same logic is now available to commercial lending. Federal Reserve economists have built a dataset that maps large banks’ commercial and industrial loan books by the location of borrower facilities rather than corporate headquarters, then overlaid those geographies with physical hazard data. The result is a screen that conventional underwriting never runs, surfacing the counties where a bank’s lending concentration and a region’s physical hazard happen to coincide. A loan booked to a firm headquartered in a safe city can be exposed through a plant several states away, and only facility-level resolution makes that legible.
Turning Exposure into Loss
The second stage translates asset-level hazard exposure into financial loss through structural credit models, replacing narrative proxies with chains grounded in climate science. One widely cited asset-level methodology runs from the geolocation of productive facilities, through hazard exposure under standard climate pathways, to damage functions and finally to issuer-level loss, supporting a portfolio Climate Value-at-Risk built on hazard rather than assumption. Research published in Nature Climate Change found that losses amplify substantially once spatial concentration and cross-exposure are accounted for, and that they are systematically understated when physical risk is modelled at the asset-class level without geographic granularity.
For infrastructure and project finance, where data is patchier and assets are long-lived, the modeling is more advanced still. A World Resources Institute analysis of a development bank’s power-generation portfolio showed that recurrent neural networks trained on hydrology and climate-projection data can generate asset-level risk scores flexible enough to work with the uneven data that characterizes real portfolios. The output is the investment-grade risk metric an allocation decision actually requires, produced from a probabilistic hazard input that could not previously be used directly.
Making it Usable for Capital
The third stage takes the translated loss estimates and expresses them in the metrics that govern bank and insurer capital. A 2026 framework in the Journal of Risk and Financial Management combines regional hazard and transition variables with climate-adjusted asset volatilities to produce portfolio Value-at-Risk and Expected Shortfall built to feed directly into the Internal Capital Adequacy Assessment Process and Pillar 2 reporting. It is engineered to fix the specific weakness of conventional stress tests, their assumption of smooth, normally distributed losses, which cannot represent the fat-tailed and geographically uneven distributions that physical climate risk actually produces.
A 2024 review of machine learning in climate finance maps this same three-stage progression across the literature and notes that the methods have gained the most traction precisely where the work requires integrating large spatial datasets with non-linear hazard-to-damage relationships and scenario-conditioned projections. This is the translation problem, surfacing wherever the data is hardest to assemble by hand.
The Measurement Chain
| Stage | What it newly makes measurable | Representative evidence |
|---|---|---|
| Resolve the exposure | Which specific assets sit in the path of which hazards, and where lending concentration and hazard coincide | Florida mortgage default study (EJOR, 2024); Federal Reserve commercial-loan exposure mapping (2025) |
| Translate to loss | Hazard exposure converted to issuer- and portfolio-level loss through damage functions, not proxies | Climate VaR methodology (2022); Nature Climate Change (2025); WRI power-portfolio analysis (2021) |
| Make it capital-usable | Loss estimates expressed as fat-tailed, prudentially aligned capital metrics | Spatial climate VaR for ICAAP and Pillar 2 (JRFM, 2026) |
Where the Translation Matters Most
The stakes are highest where two parts of the financial system are exposed to the same physical risk but manage it on different clocks. Insurers reprice annually and can withdraw, while banks hold the loan for decades and have generally treated the insured collateral behind it as protected for the life of the loan. The UK regulator’s December 2025 supervisory statement now expects banks and insurers to evaluate collateral impairment and insurability directly, which is a regulator’s way of saying the old assumption no longer holds.
In our own research we map six channels through which risk moves between the two sectors: collateral value erosion, borrower creditworthiness deterioration, protection-gap amplification, reinsurance market stress, liquidity and funding pressure, and macro-financial feedback. Most of these were until recently conceptual, real in theory and unpopulated in practice. AI-enabled translation is what turns them into variables a model can carry. The figure below sets out the structure, with the direction of transmission marked on each channel.
Climate Risk Transmission Pathways Between Banks and Insurers
The clearest empirical proof comes from the Federal Reserve Bank of Dallas, where economists linked home-insurance policy data for 6.7 million borrowers to mortgage and credit outcomes. A one-standard-deviation rise in premiums increases the probability of mortgage delinquency by 0.6 percentage points, a 16 percent jump from the portfolio mean, and the effect runs roughly three times larger for non-jumbo loans and concentrates among borrowers with high debt-to-income and loan-to-value ratios. It then spreads, raising credit-card delinquency and worsening creditworthiness, which tells us the mechanism is liquidity rather than inattention. An instrumental-variable design establishes that the relationship is causal. That closes the circuit the map describes. The insurer’s pricing decision becomes a quantifiable input to the bank’s credit book.
Pricing in One Channel
It is worth showing what the difference in resolution does to a number. Take the channel where insurer repricing flows into borrower delinquency, and run it two ways. A conventional credit model that works from portfolio averages scores this channel at close to nothing, because it neither resolves the hazard concentration nor connects the insurer’s decision to the bank’s book. Resolve both, and the same channel becomes material in the segment where it lands.
The Insurer-Repricing Channel Shown Two Ways (illustrative)
| Modelling choice | Portfolio-average model(Baseline) | Asset-level resolution(Solution) |
|---|---|---|
| Hazard exposure | Averaged across the book, diluted toward zero | Resolved to the asset, concentrated in the exposed segment |
| Insurer repricing linked to credit model | Not linked | Linked via the measured premium-to-delinquency effect |
| Incremental delinquency in exposed segment | ~0 | +0.6 pp(source: Federal Reserve Bank of Dallas, 2025) |
| Exposed segment as share of book | n/a | ~⅙(estimation of high-risk locations plus high Debt-to-Income household vulnerability and/or high Loan-to-Value ratio proportion, per Federal Reserve Bank of Dallas, 2025) |
| Channel contribution | ≈ 0 bp | ≈ 100 bp of segment |
Assumptions: The incremental delinquency figure is anchored on the Federal Reserve study’s verified one-standard-deviation effect and expressed as additional at-risk exposure in the exposed segment, before any loss-given-default or roll-rate adjustment. The exposed segment is taken at roughly one-sixth of the book. The documented three-times-larger effect on non-jumbo loans is held out as upside, not used here. The figure is illustrative. Spread across the whole book the channel is worth approximately fifteen basis points, which is why an aggregate model scores it near zero, and that is exactly the problem.
What Measurement Changes
None of this removes the judgment the work requires. Choosing which hazard scenarios to model, which damage functions fit which assets, and which metrics matter for a given book sits upstream of any algorithm, and that is where domain expertise earns its place. What the computational layer changes is reach. It absorbs the data assembly and the non-linear modelling that used to put those decisions out of a practical range, which is the difference between a channel you can describe and one you can put a number against.
For most of the past decade, physical climate risk lived in disclosures and long-horizon scenarios, acknowledged, increasingly believable, but rarely actionable. Measurement moves it into the systems where real capital moves, in front of the people who price, underwrite, and lend, on a timeline they can still act on. That is the difference between reporting a risk and managing it, and it is where the value of this effort in sustainability now lies.
Ready for the Future, Ready Now
Tackling the challenges of today and creating the opportunities for tomorrow requires an end-to-end approach that turns challenge into opportunity. Partner with us to see how sustainable innovation augmented by AI can make meaningful business impact.
