Skip to main content
    JOJonas Osman
    All articles
    Insurance
    Modelling
    Climate

    The Future of Catastrophe-Risk Modelling: From Black Boxes to Transparent Frameworks

    Why the next generation of catastrophe models will be judged on transparency and explainability, not just numbers — and how secondary perils are reshaping the modelling landscape.

    By Jonas Osman Abdelghafour · · 12 min read
    The Future of Catastrophe-Risk Modelling: From Black Boxes to Transparent Frameworks — technical illustration by Jonas Osman Abdelghafour, insurance and modelling and climate risk modelling
    The Future of Catastrophe-Risk Modelling: From Black Boxes to Transparent Frameworks — Insurance · Modelling · Climate

    The year 2025 should have been a quiet one for catastrophe modellers. For the first time in a decade, no hurricane made landfall on the US coast. And yet, according to Swiss Re Institute's sigma 1/2026 report, global insured losses from natural catastrophes still reached USD 107 billion — the sixth consecutive year above the USD 100 billion mark. A record 92% of those losses came from so-called secondary perils: the January wildfires in Los Angeles alone produced around USD 40 billion of insured losses, the largest wildfire loss event on sigma records, while severe convective storms added USD 51 billion (Swiss Re Institute, 2026).

    Pause on that for a moment. The perils that traditional catastrophe models were built around — hurricanes and earthquakes — contributed a small fraction of the year's losses. The perils that dominated are precisely the ones the industry has historically modelled least well. If your view of risk is anchored in a vendor model calibrated primarily to peak perils, 2025 is a polite warning that the risk landscape and the modelling landscape have drifted apart.

    The business problem: models under pressure from three directions

    Catastrophe models have served the insurance industry well for three decades. They transformed underwriting after Hurricane Andrew, made probabilistic capital allocation possible, and underpin today's reinsurance and insurance-linked securities markets. But three pressures are now converging on the traditional model paradigm.

    First, the risk itself is non-stationary. Swiss Re's analysis of the 1970–2025 loss record attributes more than 80% of the long-term growth in weather-related insured losses to exposure growth — more valuable property built in harm's way, at higher reconstruction costs — but also finds that for some perils and regions, hazard and vulnerability appear to be evolving faster than exposure alone would imply (Swiss Re Institute, 2026). A model calibrated to the historical record embeds yesterday's climate and yesterday's building stock.

    Second, secondary perils have become primary loss drivers. Wildfire insured losses are estimated by Swiss Re to be growing at roughly 12% per year — the fastest-growing peril — yet wildfire, hail and flood models remain younger, more data-constrained and less validated than hurricane and earthquake models.

    Third, validation expectations have hardened. Under Solvency II, insurers using catastrophe output within internal models must demonstrate understanding of the models they rely on. "The vendor calibrated it" is not an acceptable answer to a supervisor, a validator, or increasingly a board. EIOPA has explicitly promoted the use of open-source modelling and data as part of addressing climate-related protection gaps, including releasing its CLIMADA-based application to make open catastrophe modelling more accessible (EIOPA, 2023a; EIOPA, 2023b).

    The business problem, then, is not that cat models are wrong. It is that opaque models are becoming difficult to govern, difficult to adjust for a changing climate, and difficult to defend.

    Overlaid physical hazard intensity layers for flood, wind and heat exposure, illustrating "The business problem: models under pressure from three directions" in The Future of Catastrophe-Risk Modelling: From Black Boxes to Transparent Frameworks
    Figure 1. Overlaid physical hazard intensity layers for flood, wind and heat exposure, in the context of the business problem: models under pressure from three directions.

    What a catastrophe model actually is

    Before discussing where the architecture is heading, it is worth stating plainly what sits inside any catastrophe model, vendor or open-source. Four modules connect in a chain:

    1. Hazard — a large catalogue of simulated events (storm tracks, fire footprints, flood extents), each with physical intensities at each location and an annual rate of occurrence.
    2. Exposure — the portfolio of insured assets: locations, values, construction and occupancy characteristics.
    3. Vulnerability — damage functions translating hazard intensity at a site into a damage ratio for a given building type.
    4. Financial — policy terms (deductibles, limits, reinsurance structures) that convert ground-up damage into insured loss.

    The core output is the exceedance-probability curve. For each simulated year, the model produces an annual loss; ranking these losses gives the probability that annual losses exceed any threshold. Two summary quantities matter most. The expected annual loss is:

    EAL = Σᵢ λᵢ · E[Lᵢ]

    where the sum runs over all events i in the catalogue, λᵢ is the annual occurrence rate of event i (how many times per year, on average, that event happens), and E[Lᵢ] is the expected insured loss if it does occur. In plain terms: multiply how often each event happens by how much it costs, and add everything up. The return-period loss — for example the 1-in-200-year annual loss that drives Solvency II capital — is read directly off the exceedance curve as the loss level exceeded with probability 0.5% in any year.

    Every catastrophe model in the market is a version of this chain. The differences — and the risk to the user — sit in the assumptions inside each module: how the event catalogue was built, which climate baseline it reflects, how damage functions were fitted, and how uncertainty is propagated. When those assumptions are invisible, the user carries model risk they cannot measure.

    The transparency movement is no longer fringe

    A structural shift is under way. The Oasis Loss Modelling Framework — a not-for-profit, fully open-source platform for building and running catastrophe models — now hosts an ecosystem of more than 18 model suppliers covering over 90 models, alongside open data standards for exposure, hazard and vulnerability designed to improve interoperability across the market (Oasis LMF, 2026). On the regulatory side, EIOPA's promotion of open-source tools, and the joint ECB–EIOPA work on European natural-catastrophe risk management, both point in the same direction: supervisors want risk assessment capabilities that can be examined, compared and stress-tested, not merely licensed (EIOPA & ECB, 2024).

    This does not mean vendor models are obsolete — far from it. Vendor models embody decades of peril science and claims calibration that open frameworks cannot simply replicate. The realistic future is a blended architecture: vendor views where they are strongest, open and in-house components where transparency, adjustability or peril coverage demand it, and a disciplined process for reconciling the two.

    Prior and posterior densities illustrating Bayesian parameter updating, illustrating "The transparency movement is no longer fringe" in The Future of Catastrophe-Risk Modelling: From Black Boxes to Transparent Frameworks
    Figure 2. Prior and posterior densities illustrating Bayesian parameter updating, in the context of the transparency movement is no longer fringe.

    A practical workflow: building an adjustable view of risk

    Consider a mid-sized European property insurer whose vendor model underweights wildfire — a reasonable concern after 2025. A pragmatic modelling workflow, of the kind Quantica Risk is developing in its own frameworks, looks like this:

    1. Decompose the vendor output by peril and region, and identify where recent loss experience or scientific evidence diverges from modelled expectations.
    2. Construct an adjustment layer: a frequency–severity model of the under-represented peril (for example, a Poisson frequency assumption with a heavy-tailed severity distribution fitted to industry loss data), run on the insurer's own exposure through an open platform.
    3. Blend the vendor and in-house views with explicit, documented weights — and record the rationale as a formal model-change decision.
    4. Validate by backtesting the blended view against realised losses and benchmarking against industry aggregates such as the sigma loss record.
    5. Govern: every assumption in the adjustment layer is written down, owned, and reviewable — which is precisely what a black box cannot offer.

    The mathematics in step 2 is deliberately classical; the innovation is architectural. Transparency is not a property of any single equation. It is a property of the workflow.

    Implications for risk leaders

    For chief risk officers and chief actuaries, three conclusions follow. First, treat your catastrophe model as a portfolio of assumptions to be governed, not a product to be consumed; ask which module — hazard, exposure, vulnerability or financial — dominates uncertainty in your book, because that is where validation effort belongs. Second, invest in the capability to adjust: the gap between modelled and emerging risk in secondary perils is now material enough that a pure licence-and-run operating model leaves earnings exposed and supervisors unsatisfied. Third, watch the open-source ecosystem seriously. Open platforms and data standards are becoming the lingua franca through which regulators, reinsurers and cedants will compare views of risk; institutions fluent in that language will negotiate, price and report from a position of strength.

    End-to-end data and calibration pipeline from source data to reported risk measures, illustrating "Implications for risk leaders" in The Future of Catastrophe-Risk Modelling: From Black Boxes to Transparent Frameworks
    Figure 3. End-to-end data and calibration pipeline from source data to reported risk measures, in the context of implications for risk leaders.

    Conclusion

    The 2025 loss year — record wildfire losses, no US hurricane landfall, and still USD 107 billion of insured losses — is a snapshot of a risk landscape that has outgrown the assumption set of the classical vendor-model paradigm. The next generation of catastrophe modelling will not be defined by a single breakthrough technique. It will be defined by architecture: modular, inspectable, adjustable frameworks in which every assumption has an owner and every change leaves a trail. The firms that build that capability now will not merely satisfy their validators. They will understand their own risk better than their competitors understand theirs.

    About the author. Jonas Osman is a risk-modelling and financial-risk professional and the founder of Quantica Risk, an AI-driven modelling company focused on insurance, banking, climate risk, actuarial analytics, and model validation.

    Quantica Risk develops transparent, data-driven modelling frameworks for financial institutions and risk-sensitive businesses. To discuss catastrophe risk, climate analytics, or model validation, visit the Quantica Risk website. [website — to be inserted when verified]

    • Next in this pillar: Frequency–Severity Modelling of Insured Catastrophe Losses (Article A2) — the classical machinery behind the adjustment layer described above.
    • Related: From Hazard to Loss: A Financial Framework for Physical Climate Risk (Article B1) — how the same four-module chain generalises to climate risk beyond insurance.
    Run-off triangle structure used for actuarial reserving and development factor estimation, illustrating "Series links" in The Future of Catastrophe-Risk Modelling: From Black Boxes to Transparent Frameworks
    Figure 4. Run-off triangle structure used for actuarial reserving and development factor estimation, in the context of series links.

    References

    Related articles

    Work with Jonas

    For partnerships, modelling reviews, advisory roles, institutional risk projects or technical collaboration.

    Contact Jonas