AI Climate Risk Pricing: What the Models Now Show

AI climate risk pricing has moved from a conference slogan to a set of working models with named authors and published numbers. Central banks use machine learning to project sovereign credit ratings to 2050. Academics train tree ensembles on catastrophe bond coupons. Data vendors score a billion rooftops. What the field lacks is a single view of what these models say, asset class by asset class, and a plain statement of where they disagree with the market.

This piece supplies both. It sets out what the models now show for sovereign ratings, sovereign bonds, mortgages and insurance, and catastrophe bonds, and then the finding that should worry any investor holding long-dated debt: the models call climate risk material, while the bond market still prices physical risk only faintly.

What AI climate risk pricing means, and why it lagged

Todd Cort of the Yale School of Management calls this category unpriced risk: a risk that is apparent, not a black swan, yet lags in pricing because the data is difficult, quantifying it is costly, or the problem is simply daunting. His argument is that AI's capacity to parse and organise very large, messy datasets is what now lets sustainability teams state these risks in the quantitative language a chief financial officer accepts.

Two kinds of risk are in play. Physical risk is the damage from heat, flood, storm and drought. Transition risk is the repricing that follows policy, technology and demand shifting toward lower emissions. Machine learning earns its place because the links between climate variables and financial outcomes are non-linear, and because, as KPMG notes, these models capture complex relationships that traditional models struggle to incorporate. The same note is blunt about maturity: adoption beyond insurance is held back by data availability, model accuracy and regulatory uncertainty, and the models are harder to govern for explainability and fairness.

Sovereign credit ratings: the random forest result

The most cited piece of AI climate risk pricing is a sovereign ratings study from Cambridge. It matters because ratings are decision-ready. An ESG score leaves the investor to decide what it means; a rating already drives capital requirements, collateral rules and borrowing costs.

108 countries, three warming scenarios

Klusak, Agarwala, Burke, Kraemer and Mohaddes built what they describe as the world's first climate-adjusted sovereign credit rating for 108 countries. Under a high-emissions path (RCP 8.5), 63 sovereigns suffer climate-induced downgrades by 2030, averaging 1.02 notches, rising to 80 sovereigns and an average of 2.48 notches by 2100. The cost is concrete: an extra US$137 to 205 billion a year in sovereign interest payments under RCP 8.5, against US$22 to 33 billion under RCP 2.6. Stringent policy consistent with the Paris Agreement nearly eliminates the effect.

The method is worth stating because it is replicable. The authors trained a random forest on S&P sovereign ratings from 2015 to 2020, then fed it macroeconomic data adjusted for each warming scenario. The ratings in question gatekeep more than US$66 trillion in sovereign debt. Under RCP 8.5, Chile, China, Slovakia, Malaysia, Mexico, India, Peru and Canada each fall by more than five notches by 2100. Rich and cold countries are not exempt.

The Bank of England adopts the method

A research result becomes a pricing tool when a regulator uses it on its own balance sheet. The Bank of England's 2024 Quarterly Bulletin projects sovereign credit ratings to 2050 with a random forest of seven features, citing the same non-linearity argument. The Bank applies scenario analysis to the three asset classes it is most exposed to: sovereign bonds, corporate bonds and residential mortgages.

Sovereign bonds: from ratings to yield curves

A rating is an input. Investors need a price. The next step in AI climate risk pricing is to translate projected ratings and debt ratios into yields and portfolio values.

Shocked yield curves and portfolio losses

The Bank of England builds climate-shocked yield curves by adding the change in sovereign credit risk to a scenario-specific risk-free curve, then reprices bonds by duration and convexity. The exercise assumes a "climate Minsky moment", in which investors wake up to the risk at once. The result is not small: a longer-dated portfolio loses more than 9% of its value in every scenario, and more than 20% under Divergent Net Zero. Credit risk shocks are substantially larger at longer tenors.

Chronic physical damage at regional resolution

The ratings approach inherits the coarseness of national data. A sovereign climate rating built by EDHEC Climate Institute and Scientific Climate Ratings goes below the country level: 191 countries, more than 95% of world output, 3,400 regions, and the equivalent of 180 billion data points processed from NASA downscaled climate data. Damages are mapped onto the seven NGFS scenarios plus two more extreme ones and expressed as an A to G rating, with G-rated sovereigns facing structural losses by mid-century. For the United States the model shows close to a 5% GDP loss by 2035 and 10.4% by 2050.

The effect reaches the supposedly safe end of the curve too. A stochastic integrated assessment model by Chikhani and Renne finds that climate risks lower the 10-year real rate by around 30 basis points in its base calibration, driven by weaker growth and precautionary saving. The authors caution that climate risk premiums are particularly sensitive to damage assumptions, a caveat that applies to every model in this article.

Mortgages, insurance and municipal debt

For households, climate risk is priced first through insurance. The Bank of England's mortgage method couples catastrophe model data from JBA with ordinary financial modelling. In the United Kingdom, the government-backed FloodRe scheme caps premiums for many flood-prone homes until 2039. Once the cap rolls off, the Bank estimates that the fall in house prices in the top 10% of flood-exposed areas is on average 6.5 times larger than in a median-risk area, because buyers will price in higher premiums and the chance that a home becomes unmortgageable.

In the United States the repricing is already visible. ICE reports that home insurance premiums are up 52% on average since 2019 in high-risk areas such as New Orleans, and by as much as 90% in some places, while elsewhere cover cannot be found. The same data set covers more than one billion rooftops and scores municipal issuers' flood, wildfire and hurricane exposure from zero to five, because municipal bonds depend on a tax base that can move away. Research cited there suggests adaptation needs could double annual municipal bond issuance by the mid-2030s.

Catastrophe bonds: learning the price directly

Catastrophe bonds are where machine learning prices climate risk most directly, because the payoff is tied to a physical event. A study of ML models for predicting CAT bond coupons adds climate indices, including the Oceanic Niño Index, the North Atlantic Oscillation and sea surface temperatures, to the standard features. Including climate variables improves predictive accuracy across every model tested, with extremely randomized trees achieving the lowest error. Large-scale climate variability, in other words, is measurably in the price.

A second line of work uses neural networks as the pricing engine itself. Sester and Xu train networks on simulated CAT bond prices and report near-instant valuations, several orders of magnitude faster than Monte Carlo methods, with accuracy maintained. Speed matters here for risk management: a fast pricer can be interrogated for sensitivities that a slow one cannot.

The gap the models expose: markets still underprice physical risk

Put the modelling together and it points one way. Climate change downgrades sovereigns, raises borrowing costs, cuts long-dated portfolio values and reprices homes. The market evidence is more equivocal.

A Riksbank working paper covering 52 developed and developing countries over two decades finds that transition risk does raise sovereign yields, particularly in developing and high-emission countries after the Paris Agreement. Physical risks such as temperature anomalies, however, generally are not priced in. High debt amplifies yield increases during acute events, and natural disasters move yields mainly through ordinary macroeconomic channels, debt, growth and inflation, rather than through a distinct climate premium.

That is the gap, and it cuts two ways. One reading is that the models overstate damage, which the Chikhani and Renne caveat about damage assumptions would support. The other is the Bank of England's Minsky scenario: the risk is real, the market has not yet recognised it, and the repricing will arrive late and all at once. EDHEC's stated purpose is exactly to surface structural vulnerabilities "before these are priced into the spread". No one in the current literature can yet say which reading is right. Closing that question needs research, market data and the institutions that hold sovereign debt working from the same evidence, which is a coordination problem as much as a modelling one. For the parallel story on the harm-reduction side, see our analysis of AI in climate change adaptation, and for the wider shift in the profession, how AI is changing finance.

Where this is being worked on in November

The EX Future Summit runs 18 to 20 November 2026 in three rooms at once, online, Las Palmas and Bali, and lists Willi Semmler, Senior Researcher at IIASA and professor at The New School for Social Research, among its speakers. Its programme includes a Finance track and a Global Financial Raise track on R&D funding pathways across venture capital, private equity, national grants and Horizon Europe. Online attendance is free for verified researchers and students. For what to expect from the finance programme specifically, see our note on the AI in finance summit.

FAQ

Which machine learning models are used to price climate risk?

Random forests are the standard for projecting sovereign credit ratings, used both by the Cambridge study and by the Bank of England. For catastrophe bond coupons, extremely randomized trees, gradient boosting and random forests have been compared, and neural networks are used as fast pricing engines trained on simulated prices.

Is climate risk already priced into sovereign bond yields?

Partly. The Riksbank evidence across 52 countries finds transition risk priced into yields since the Paris Agreement, especially for high emitters, but physical risk such as temperature anomalies generally not.

What is a climate-adjusted sovereign credit rating?

It is the rating a model predicts for a country once its macroeconomic inputs are adjusted for a given warming scenario. The Cambridge team built the first one for 108 countries by training on real S&P ratings.

Can climate risk models be trusted for pricing today?

With caution. Results are highly sensitive to damage assumptions, data coverage is uneven, and machine learning models raise explainability and fairness questions that supervisors expect firms to govern.