💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

원문 렌더가 준비되기 전까지 텍스트 가이드로 표시합니다.

In September 2024, Hurricane Helene tore through the US Southeast and triggered around 50B USD of insured losses; the following month Milton hit Florida and added another 30B USD. In January 2024 the Noto Peninsula earthquake (M7.6) burned through an entire quarter of Japan's earthquake-insurance pool. Korea, too, has seen typhoon and flood reinsurance rates rise more than 35% cumulatively since Hinnamnor in 2022. By 2026 reinsurance and cat-risk modeling are no longer a back-office function — they are core infrastructure deciding how much underwriting capital the industry has. This piece walks through the big four European reinsurers (Munich Re, Swiss Re, Hannover Re, SCOR), the Bermuda and US camp (RenaissanceRe, Berkshire Hathaway Re), the four-way model-vendor fight (Verisk AIR Worldwide, RMS / Moody's, CoreLogic, KCC and Aon's Impact Forecasting), and the Korean and Japanese players (Korean Re, Hanwha Re, Japan REIN, Tokio Marine retrocession).

What Is Reinsurance — Insurance for Insurers

Reinsurance is a contract where one insurer (the cedent) transfers part of its underwritten risk to another insurance company (the reinsurer). The cedent cedes part of its premium and, when large losses occur, recovers from the reinsurer. This lets small primary insurers underwrite enormous single risks (apartment complexes, refineries, large vessels) and stay capital-stable across natural-catastrophe accumulations. Capital regimes such as Solvency II and K-ICS reflect the diversification benefit of reinsurance directly in capital cost. Reinsurance is not just a cost-saving tool — it is a capital-efficiency lever.

Reinsurance vs Primary Insurance — Structural Differences

Primary insurance is a one-to-one contract between an insurer and a policyholder, while reinsurance is a B2B contract between insurers. Primary policy wording is standardised and heavily regulated; reinsurance terms are freely negotiated by both parties. In primary insurance the policyholder makes the claim; in reinsurance the cedent aggregates losses above an attachment and bills the reinsurer quarterly or annually. Reinsurance also almost always involves a broker (Aon, Guy Carpenter, Howden Re) because market intelligence, capacity pricing and program structuring are too complex for a cedent to handle alone. Over 70% of global reinsurance flows through brokers.

Treaty Reinsurance — Proportional vs Non-proportional

A treaty automatically cedes a defined block of business (for example all new Korean property policies). A proportional treaty splits premium and losses by a fixed ratio. A quota share cedes the same percentage of every policy (say 30%); a surplus cedes only the part above a retention. A non-proportional treaty does not share premiums proportionally; the reinsurer only pays losses above a defined attachment. The classic non-proportional structure is excess of loss (XOL). A layer written as `1,000M USD xs 500M USD` means the reinsurer takes losses from 500M USD up to 1.5B USD. Per-risk XOL applies per individual risk, cat XOL per single catastrophe event, and aggregate XOL per accumulated annual loss.

Facultative Reinsurance — Negotiating Mega Risks One at a Time

If treaty is portfolio-level automatic cession, facultative reinsurance is per-risk negotiation. It is used for mega risks where one location exceeds the cedent's appetite — semiconductor fabs, LNG terminals, space-launch vehicles, large wind farms. The cedent's broker takes the offer to the global market and Munich Re, Swiss Re, Lloyd's syndicates and others independently write a "line" (a percentage of the risk they take). It is slower and more expensive than treaty, but every detail of one large risk can be negotiated.

Cat Bonds and ILS — Tapping Capital Markets

Traditional reinsurance does not have enough capital to cover the 1-in-250-year tail. That is where cat bonds and the broader ILS (Insurance-Linked Securities) world come in. The cedent sets up an SPV (special purpose vehicle), the SPV issues a bond to capital-market investors, the proceeds are held as collateral, and if a defined trigger fires (for example a Saffir-Simpson 4 Florida hurricane) principal is reduced. Global cat bond issuance hit a record 17.5B USD in 2024, and per artemis.bm Q1 2026 outstanding crossed 51B USD. Pension funds and hedge funds are the main investors, drawn by low correlation between natural catastrophes and broader capital markets. ILS spans cat bonds, sidecars, collateralised reinsurance and industry loss warranties.

What Is Cat Modeling — The Four Modules of a Probabilistic Model

A cat model is not simple statistics but a probabilistic simulation built from four modules: hazard, exposure, vulnerability and financial. The hazard module generates tens of thousands of synthetic events (hurricane tracks, earthquake scenarios). The exposure module holds insured buildings, contents and business interruption coded with coordinates, construction, occupancy and number of stories. The vulnerability module expresses the loss ratio of each asset at a given hazard intensity via vulnerability functions. The financial module applies policy terms (deductibles, limits, coinsurance) to produce gross and net insured losses. The four modules are looped tens or hundreds of thousands of times via Monte Carlo to build a loss distribution, summarised as an EP (Exceedance Probability) curve and PML (Probable Maximum Loss).

Hurricane — Saffir-Simpson Scale and the NHC Best Track

Atlantic and Pacific hurricanes are classified on the Saffir-Simpson 1–5 scale based on 1-minute sustained wind. SS1 is 119–153 km/h, SS5 exceeds 252 km/h. Cat models build their stochastic catalogs on NHC HURDAT2 (1851 to present) and NOAA NHC best track data. One cycle of a catalog typically contains 100,000 to 250,000 years of synthetic seasons. Landfall location, angle, forward speed and decay rate are decisive for loss. Hurricane Helene in 2024 made landfall at Florida's Big Bend at SS4 (225 km/h) and pushed deep inland into North Carolina. Models with weaker inland-flood physics under-predicted PML by more than 30%.

Earthquake — PGA and Return Period

Earthquake models use PGA (Peak Ground Acceleration, in %g or cm/s^2) and PGV (Peak Ground Velocity) as hazard intensity, not wind. USGS NSHM (National Seismic Hazard Map), Japan's J-SHIS and Korea's KMA seismic-hazard maps are the foundational inputs. Catalog construction blends a fault rupture model (which fault breaks at which magnitude how often) and a GMPE (Ground Motion Prediction Equation) that translates that into PGA at site, corrected for distance and soil. A 1-in-475-year PGA is the basis of Korean building structural codes; 1-in-2475-year underpins nuclear-safety assessment. The 2024 Noto Peninsula quake recorded 2,800 gal of PGA at some sites — close to 1-in-1000-year intensity.

Flood — Return Periods and the 1-in-100-Year Trap

Flood models flow from rainfall to runoff to river stage to inundation footprint to depth-damage to insured loss. A 1-in-100-year flood means the annual exceedance probability is 1%, not that the event happens once every 100 years. The probability of seeing at least one such flood across a 30-year mortgage is about 26%. FEMA NFIP in the US, the Environment Agency Flood Map in the UK, hazard maps from Japan's ministries and the Ministry of Environment's flood maps in Korea provide base layers. As rainfall intensity rises with climate change, what used to be 1-in-100-year is increasingly 1-in-50-year or less. Climate-conditioned catalogs that fold this in are becoming standard in the 2024–2026 generation of RMS, Verisk and Fathom flood models.

Monte Carlo Simulation — Computing PML Directly

PML is not an average — it is a tail quantile of the loss distribution. The 1-in-250-year PML is the 99.6th percentile. Assuming a lognormal or GPD tail, we can compute it directly with Monte Carlo:

rng = np.random.default_rng(42)

Number of events per year: Poisson(lambda=3)

Loss per event: lognormal(mu=18, sigma=1.4) — mean around 200M USD, fat tail

N_SIMULATIONS = 200_000

annual_loss = np.zeros(N_SIMULATIONS)

events = rng.poisson(lam=3.0, size=N_SIMULATIONS)

for i in range(N_SIMULATIONS):

if events[i] > 0:

losses = rng.lognormal(mean=18.0, sigma=1.4, size=events[i])

annual_loss[i] = losses.sum()

aal = annual_loss.mean() # Average Annual Loss

pml_100 = np.percentile(annual_loss, 99) # 1-in-100-year

pml_250 = np.percentile(annual_loss, 99.6) # 1-in-250-year

print(f"AAL : {aal/1e6:,.1f} M USD")

print(f"1-in-100 PML : {pml_100/1e6:,.1f} M USD")

print(f"1-in-250 PML : {pml_250/1e6:,.1f} M USD")

Replace the loss draws with a real event catalog and add vulnerability curves and you have the spine of a full cat model. Commercial models run hundreds of thousands of events with 100K to 1M simulation passes.

AIR Touchstone API — Calling a Model Vendor

Verisk AIR Worldwide's Touchstone is one of the de facto standards in cat modeling. You can drive it via a REST API:

API = "https://api.air-worldwide.com/v3"

TOKEN = os.environ["AIR_TOKEN"]

1) Upload exposure in CEDE format

files = {"file": open("exposure_kr_property.cede", "rb")}

exposure_id = requests.post(

f"{API}/exposures",

headers={"Authorization": f"Bearer {TOKEN}"},

files=files,

).json()["exposure_id"]

2) Trigger an analysis

payload = {

"exposure_id": exposure_id,

"model": "AIR Typhoon Model for Korea v3.0",

"perspective": "GU", # Ground Up

"event_set": "100K-year stochastic",

"currency": "KRW",

}

analysis = requests.post(

f"{API}/analyses",

headers={"Authorization": f"Bearer {TOKEN}"},

json=payload,

).json()

3) Poll for results

results = requests.get(

f"{API}/analyses/{analysis['id']}/results",

headers={"Authorization": f"Bearer {TOKEN}"},

).json()

print(f"AAL : {results['aal_krw']:,.0f} KRW")

print(f"100-yr : {results['pml_100']:,.0f} KRW")

print(f"250-yr : {results['pml_250']:,.0f} KRW")

Moody's RMS RiskLink and the Intelligent Risk Platform expose similar REST surfaces, and KCC RiskInsight, CoreLogic RQE and Aon Impact Forecasting's ELEMENTS each have their own APIs.

Cat Bond Payoff Structure — Parametric vs Indemnity

Cat bond triggers fall into four broad families. Indemnity triggers reference the cedent's actual loss and are most accurate but slow to settle. Parametric triggers reference an objective measurement (wind speed, seismic intensity) and pay immediately when a threshold is crossed. Industry-loss triggers reference PCS or Verisk PCS Catastrophe Bulletin industry losses. Modeled-loss triggers re-run a cat model on the actual event. Parametric is simple but carries basis risk — the gap between actual loss and the trigger value. Here is a simplified parametric trigger:

def cat_bond_payoff(event_wind_kph, attachment, exhaustion, principal):

"""

Parametric cat bond payoff for hurricane wind.

attachment : trigger wind in kph (e.g. 220)

exhaustion : full payout point (e.g. 260)

principal : bond principal (e.g. 100M USD)

Returns : principal remaining for investors (USD)

"""

if event_wind_kph < attachment:

return principal # No trigger, full principal returned

if event_wind_kph >= exhaustion:

return 0 # Total principal loss

pct = (event_wind_kph - attachment) / (exhaustion - attachment)

return principal * (1 - pct)

Example: 500M USD cat bond, attachment 220 kph, exhaustion 260 kph

remaining = cat_bond_payoff(245, 220, 260, 500_000_000)

print(f"Principal remaining for investors: {remaining/1e6:,.1f} M USD")

→ Attachment 220, observed wind 245, exhaustion 260 → ~62% remaining, 37.5% loss

Climate Change Impact — RCP, SSP and NGFS Scenarios

Traditional cat models assumed stationarity of the 1971–2020 climate record. Climate change breaks that assumption. The 2026 standard is to reweight future catalogs using RCP (Representative Concentration Pathway) or SSP (Shared Socioeconomic Pathway) scenarios. RCP 4.5 is a midline stabilisation path; RCP 8.5 is a high-emissions path. The NGFS (Network for Greening the Financial System) publishes six scenarios (Net Zero 2050, Below 2 deg C, Delayed Transition, Current Policies, Fragmented World, NDCs). Insurers map their cat models onto NGFS scenarios and disclose PML evolution in 2030, 2050 and 2080. Munich Re's 2024 annual report projected average global hurricane PML to rise by 25–40% by 2050 under SSP5-8.5.

1-in-100/250-Year PML — How It Maps to Capital

Solvency II calibrates the SCR (Solvency Capital Requirement) to a 1-in-200-year confidence level. US NAIC RBC and Korea's K-ICS use different levels but the same underlying logic. Reinsurers must hold capital in excess of the 1-in-250-year or 1-in-500-year PML of their entire portfolio. To do this they recalculate cat-model PML regularly and offload the tail via retrocession (a reinsurer's own reinsurance). The retro market is a narrow universe of fewer than 50 specialist traders, most of them Bermuda-based, plus ILS funds. Average retro rates rose by more than 40% in 2023 and again in 2024–2025.

Korean Re and Hanwha Re — East Asia Typhoon and Flood Modeling

Korean Re is Korea's only stand-alone reinsurer, founded in 1963. Its 2024–2025 GoLog digital-transformation project consolidates underwriting, claims and analytics on a cloud-native stack. It runs an in-house typhoon and flood model built on KMA typhoon best-track, Ministry of Environment flood maps and KIGAM earthquake catalogs, while also licensing AIR Worldwide Typhoon Model for Korea v3 and RMS Asia-Pacific Typhoon HD Model. Hanwha Re began as the reinsurance arm of Hanwha General Insurance and underwrites Hanwha-group exposures plus external cedents. Korean reinsurance combined ratio softened to around 102% in 2025, with typhoon, flood and construction LOC exposures the main drivers.

Japan REIN, Tokio Marine Retrocession and the JER Pool

Japan's reinsurance market is unique because of how earthquake risk is handled. JER (Japan Earthquake Reinsurance) writes the primary reinsurance layer, then retrocedes back to private non-life insurers (Tokio Marine Nichido, Sompo Japan, Mitsui Sumitomo) and the government in a four-layer structure. The government's aggregate guarantee is 11.7 trillion yen as of 2024, structured in layers: below 50B yen the government covers 100%, between 50B and 193B yen 50%, and from 193B yen up to 11.7 trillion the government covers 99.5%. A new specialist reinsurer trading as Japan REIN has entered the market, while Tokio Marine runs its own retrocession program that places roughly 30% of its 1-in-200-year earthquake PML into the global ILS market.

Top 10 Global Reinsurers — 2026 GWP

|------|-----------|-----|------------------|---------------|

GWP is gross written premium and can swing by accounting choices and timing.

Top Cat Model Vendors Compared

|--------|------------------|--------------------|-----------|

US vs JP vs KR — Cumulative Insurance Exposure

|------|-----|-------|-------|

All figures vary by company and reporting year by roughly +/- 20%.

Pricing — Risk Premium vs Risk Load

Reinsurance is not priced by simply adding up expected losses (pure premium or AAL). A risk load is added to compensate for tail risk and capital usage. A typical pricing formula is:

Reinsurance Premium = AAL

+ Standard Deviation Load (kappa * stdev(loss))

+ Capital Cost (CoC * required capital)

+ Expense Load (broker, analytics, operations)

+ Margin

Industry convention is kappa around 0.1–0.5 and CoC around 8–12%. For a 1-in-250-year layer the standard deviation is huge, so premium often runs 5–15x the AAL. That is not gouging — it is the cost of locking up capital to support an extreme tail.

Retrocession Market — The Last Backstop

Retrocession is a reinsurer's own reinsurance, the place where 1-in-1000-year tails go. It is a very narrow market with roughly 30–50 specialist traders at its core, dominated by Bermuda (RenaissanceRe, Validus, Hiscox Re), the Lloyd's market and ILS funds (Fermat, RenaissanceRe DaVinci, Stone Ridge). After Harvey, Irma and Maria in 2017 retro rates jumped 30%; after Ian (2022) and Helene and Milton (2024) they accumulated a further 40%+ increase. A retro capacity crunch is an early signal that primary reinsurers will have to cut their own underwriting.

ILW and Sidecars — Simple Triggers and Quick Capacity

An Industry Loss Warranty (ILW) uses industry-wide loss, not the cedent's actual loss, as its trigger. For example: "if US hurricane PCS industry loss exceeds 30B USD, pay 100M USD." Wording disputes are minimal and basis risk is borne by the cedent. Hedge funds and ILS funds are heavy underwriters; the ILW market sits at roughly 4–6B USD in 2025. A sidecar is an SPV co-founded by a reinsurer and outside investors to share a defined book (often cat XOL) for a defined time (typically 1–3 years). The reinsurer supplies underwriting and market access; investors supply capital. RenaissanceRe Upsilon, Hannover Re K-Cessions and Everest's Mt. Logan are the canonical names.

Data Quality — Why Exposure Beats Models

Cat-model accuracy depends more on exposure data quality than on hazard or vulnerability assumptions, more often than people expect. Whether a building's location is the zip-code centroid or a real GPS coordinate, whether construction (concrete, steel, wood) is coded, and whether the number of stories and occupancy is captured can swing PML by more than +/- 30%. In the US, LexisNexis ASD and CoreLogic Property Quote are common enrichment sources; in Japan, ZmapTOWN II and LandPro; in Korea, the road-name address KAIS API plus KEPCO location data. Korean Re's GoLog program treats an enriched exposure database as one of its core deliverables.

Model Uncertainty and Ensemble — Don't Trust One Vendor

For the same portfolio, the four major cat-model vendors (AIR, RMS, KCC, CoreLogic) can produce 1-in-250-year PMLs that differ by more than 50%. Differences come from hazard assumptions (landfall frequency, GMPE choice), vulnerability curves, and demand-surge and LAE assumptions. By 2026 the standard is therefore ensemble: take a mean or median of multiple models for capital and challenge the outlier when one disagrees sharply. Munich Re (NATHAN) and Swiss Re (CatNet) run in-house models alongside the commercial vendors.

Claims and Settlement Cycle — 18 to 36 Months After an Event

Catastrophe loss settlement takes much longer than most outsiders assume. For a large hurricane, the first industry loss estimate is published within a week, but ULAE (Unallocated Loss Adjustment Expense) and IBNR (Incurred But Not Reported) only stabilise after 18 to 36 months. Hurricane Harvey's final loss number was not within a few percent until 2019. The cedent sends quarterly cession statements to the reinsurer and large events are often closed via commutation — a lump-sum settlement that buys out remaining future liabilities.

ESG and Climate Disclosure — TCFD and IFRS S2

Reinsurers have been subject to IFRS S2 climate disclosures since 2024. They must publish quantitative scenario analysis (at least two of the six NGFS scenarios), transition risk and physical risk exposure. Munich Re and Swiss Re publish standalone annual climate-risk reports, and Hannover Re has issued a 2050 net-zero underwriting roadmap. Korean Re aligns climate scenarios with K-ICS reporting under Korea's domestic ESG framework. Japan runs TCFD alongside IFRS S2 with a 2027 mandatory deadline in view.

The Next Five Years — What to Prepare For

Across 2026–2031, reinsurance and cat-risk modeling change along three axes. First, climate conditioning becomes standard: RCP/SSP reweighting and climate-conditioned catalogs are the default in every cat model. Second, data fusion accelerates: satellite imagery (Maxar, Planet), IoT sensors and municipal GIS feed exposure and post-event adjustment in near real time. Third, the boundary between insurance and capital markets blurs further: cat bonds, ILS and tokenised parametrics are projected to reach about 25% of traditional reinsurance capacity by 2030. Korea and Japan are starting policy discussions on how to redefine the split between public pools (JER, Korea's wind and flood subsidy) and private reinsurance.

References

- artemis.bm — Catastrophe bond and ILS market intelligence, "Q1 2026 cat bond issuance and outstanding": [https://www.artemis.bm/](https://www.artemis.bm/)

- Swiss Re Institute, sigma report "Natural catastrophes in 2024": [https://www.swissre.com/institute/research/sigma-research.html](https://www.swissre.com/institute/research/sigma-research.html)

- Munich Re, NatCatSERVICE statistics and annual review 2024: [https://www.munichre.com/en/risks/natural-disasters.html](https://www.munichre.com/en/risks/natural-disasters.html)

- Hannover Re, "Cat budget and 2026 outlook" investor day materials: [https://www.hannover-re.com/](https://www.hannover-re.com/)

- SCOR, "Global Risk Center" climate scenarios: [https://www.scor.com/en](https://www.scor.com/en)

- Verisk AIR Worldwide, Touchstone documentation portal: [https://www.air-worldwide.com/](https://www.air-worldwide.com/)

- Moody's RMS, Intelligent Risk Platform technical brief: [https://www.rms.com/](https://www.rms.com/)

- CoreLogic RQE, hazard model overview: [https://www.corelogic.com/intelligence/insurance/](https://www.corelogic.com/intelligence/insurance/)

- Karen Clark and Company (KCC), RiskInsight platform: [https://www.karenclarkandco.com/](https://www.karenclarkandco.com/)

- Aon Impact Forecasting, ELEMENTS modeling platform: [https://www.aon.com/impact-forecasting/](https://www.aon.com/impact-forecasting/)

- NGFS, "Climate scenarios for central banks and supervisors" (2024 phase V): [https://www.ngfs.net/en/ngfs-climate-scenarios](https://www.ngfs.net/en/ngfs-climate-scenarios)

- Korean Re, GoLog digital transformation announcement 2024: [https://www.koreanre.co.kr/](https://www.koreanre.co.kr/)

- Japan Earthquake Reinsurance Co., Ltd. (JER), structure of government reinsurance: [https://www.nihonjishin.co.jp/](https://www.nihonjishin.co.jp/)

- USGS National Seismic Hazard Maps: [https://www.usgs.gov/programs/earthquake-hazards/national-seismic-hazard-maps](https://www.usgs.gov/programs/earthquake-hazards/national-seismic-hazard-maps)

- FEMA NFIP and BureauNet data: [https://www.fema.gov/flood-insurance](https://www.fema.gov/flood-insurance)

- NOAA NHC HURDAT2 hurricane database: [https://www.nhc.noaa.gov/data/](https://www.nhc.noaa.gov/data/)

- IAIS, ICS 2.0 (Insurance Capital Standard) implementation status: [https://www.iaisweb.org/](https://www.iaisweb.org/)