Climate cat model validation:
SS1/23 applied to vendor models

Validating climate-conditioned catastrophe models under PRA SS1/23 — five principles applied to cat models, vendor model assessment, and the evidence supervisors expect to see.

Why this is harder than non-climate cat validation

Validating a non-climate vendor cat model is well-trodden — backtests against historical loss experience, vendor documentation review, sensitivity tests, governance sign-off. Climate-conditioned versions of the same models break that pattern in three specific ways.

First, by design they aren't backtested against history. A 2050-conditioned hurricane model isn't validated against 1980–2024 hurricane losses; that's the whole point. Conventional backtest evidence doesn't apply.

Second, vendor methodologies vary substantially. Verisk, RMS, KCC, Moody's RMS each handle climate-conditioning differently — different scenario references, different parameter perturbation logic, different uncertainty bands. The firm's internal model may use two of them; the same RDS scenario produces different numbers.

Third, the climate-adjustment layer is poorly documented. Vendor reference materials describe the base hazard module in detail; the climate-conditioning is often "informed by IPCC AR6" with thin transparency on the scientific assumptions actually feeding the perturbation.

Five SS1/23 principles applied to climate cat models

1. Model identification — every climate-conditioned cat model variant in production is named, including the specific scenario reference (e.g., "Verisk Climate Conditioning v3.2, RCP 4.5, 2050 horizon"). Variants used for different decisions are tracked separately.
2. Model risk governance — climate-conditioned cat models are tiered consistently with non-climate models, owned by a named senior modeller, with a documented use register.
3. Model development — for vendor models, this becomes vendor model assessment. The firm cannot directly observe the development; what it can do is challenge the vendor on methodology transparency, request scientific assumption disclosures, and benchmark across vendors.
4. Model validation — non-historical validation: scenario consistency checks (does the climate adjustment match a stated published scenario?), inter-vendor benchmark tests, expert challenge against academic literature, sensitivity to parameter uncertainty.
5. Model use — explicit use record showing which decisions the climate-conditioned model informed, who approved that use, and what the override logic is when the model output is challenged.

What supervisors actually want to see

• Documented vendor challenge. Email correspondence, methodology questions, written vendor responses on the climate adjustment specifics.
• Inter-vendor or in-house benchmark. At least one independent comparison against vendor output.
• Scenario consistency check. The vendor's climate-conditioning scenario reference (RCP, SSP, NGFS variant) is explicitly named and matches the firm's stated scenario assumption.
• Uncertainty in the use. The capital number that emerges from the vendor model is presented with an uncertainty band, not a point estimate.
• Validation refresh cycle. Annual re-validation logged, with a trigger-based mid-cycle refresh for major vendor methodology releases.

Where the Abgalis Engine fits

The Abgalis Engine doesn't replace vendor cat models — it sits next to them. The Engine's contribution is the inter-vendor benchmark layer, the channel structure that propagates cat-model output into the credit, liquidity and operational domains, and the uncertainty quantification that turns a vendor point estimate into an SS1/23-defensible band.