BRENNEN

Industries / Finance and business services / Insurance / Reinsurance

Pricing

Setting the price of a reinsurance contract from past losses, the risks it covers and the state of the market.

6

Milestones achieved, from 1983 to 2026.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1983

    The spiral nobody saw

    In the London excess of loss market of the 1980s, reinsurers bought cover from each other so often that the same loss passed round the market again and again. James Stanard and Michael Wacek wrote that the market began to notice this "London market spiral" after Hurricane Alicia in 1983. Claims from the Piper Alpha oil platform disaster of 1988 "are said to have amounted to some 10 times the insured losses", Andrew Bain of the University of Glasgow wrote, and he quotes a finding that even underwriters in this business "did not fully understand the effect of the spiral".

    LessonPassing risk round a closed group of reinsurers multiplies the claims and hides how much each one really holds.

    casact.orggla.ac.uk

    Credit

    Stanard and Wacek, for the Casualty Actuarial Society paper; Andrew Bain, for the analysis

  2. 1990

    A public price signal

    Guy Carpenter has kept an index of US property catastrophe rate on line since 1990. Rate on line is the premium divided by the limit of cover: paying $10 million for $100 million of cover is a 10% rate on line. The broker measures the change year on year on the same programs, so the index does not depend on any model's view of the risk.

    artemis.bmguycarp.com

    Credit

    Guy Carpenter

  3. 1997

    Curves for pricing by size

    Stefan Bernegger published a family of two-parameter curves, the MBBEFD class, in the ASTIN Bulletin and related them to exposure curves Swiss Re used in practice. Exposure curves let a reinsurer price a layer from the sizes of the properties insured when there is too little loss history. A free R package, mbbefd, now gives the Swiss Re curves and notes that one matches a curve Lloyd's used for industrial risks.

    casact.orgideas.repec.orgsearch.r-project.org

    Credit

    Stefan Bernegger

  4. 2006

    Hurricanes repriced

    After Katrina, Rita and Wilma in 2005, insured losses passed an estimated $50 billion, about half of it borne by reinsurers, Guy Carpenter reported. It said insurers, reinsurers, modelers, rating agencies and regulators "recognized that the existing viewpoint grossly underestimated both the frequency and severity of North Atlantic hurricanes", and reinsurers pressed for large rises at the January 2006 renewals. No later year has moved Guy Carpenter's US index more.

    LessonWhen the industry's shared view of a risk changes, prices can be reset in a single renewal.

    insurancejournal.comguycarp.com

    Credit

    Guy Carpenter, for the account

  5. 2023

    The sixth year of rises

    Guy Carpenter's US index rose 31.3% at the January 2023 renewals, its sixth straight year of rises, which together almost doubled it. The rise was smaller than it might have been because insurers kept more of their own losses: retentions, the amount an insurer pays before the reinsurance starts, went up across the board.

    guycarp.com

    Credit

    Guy Carpenter, for the index

  6. 2026

    Prices fall again

    After the June and July 2026 renewals Guy Carpenter's US index was down 16% for the year, its largest yearly fall since 2014 and 22% below its 2024 peak, Artemis reported. The broker put this down to low losses, plentiful reinsurance capital and growing appetite for risk.

    artemis.bm

    Credit

    Guy Carpenter; Artemis, for the report

Watch for

  1. Pricing beyond catastrophes

    Minesh Shah of Hiscox Re says new platforms are replacing spreadsheets for pricing reinsurance of risks other than natural catastrophes, so reinsurers can analyze whole portfolios rather than one treaty at a time, and that this "non-cat pricing has been quickly catching up". He adds that cyber data "remain relatively sparse at the cedant level", that is, for each insurer buying cover.

    Achieved whena reinsurer or an outside party publishes measured results, such as how prices set on such platforms compared with later losses.

    bermudareinsurancemagazine.com

    Claimed by

    Hiscox Re, March 2026

Open

  1. 01

    From submission to price

    A public test of whether AI agents can price a reinsurance layer as a junior analyst would. Each task gives a synthetic submission (loss history, exposure summary, proposed terms); the agent must apply burning cost, the average of past losses that would have hit the layer, and exposure rating, and return a price with its working. Every answer is computed in advance by open actuarial code.

    The brief
    Credit · unclaimed
  2. 02

    Expected loss, checked by modeler

    A public table comparing, for each catastrophe modeling firm, the expected losses it gave for the cat bonds it modeled with what those bonds actually lost. Expected loss is the average yearly loss a model predicts, and it largely sets the price investors accept.

    The brief
    Credit · unclaimed

Open 01 · the brief

From submission to price

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of whether AI agents can price a reinsurance layer as a junior analyst would. Each task gives a synthetic submission (loss history, exposure summary, proposed terms); the agent must apply burning cost, the average of past losses that would have hit the layer, and exposure rating, and return a price with its working. Every answer is computed in advance by open actuarial code.

Why it is still open

Published tests cover only parts of the job. Sia Partners tested AI models on finding the total liability under reinsurance contracts, and a 2025 preprint scores how well models ground their answers on six reinsurance task families. Neither, as far as I found, publishes full pricing tasks with checked answers.

sia-partners.comarxiv.org

Built with what exists

GEMAct, a Python package for (re)insurance costing described in the Annals of Actuarial Science; reinspy from the Casualty Actuarial Society; the mbbefd R package for exposure curves; synthetic submissions generated by code; any current AI model, or Mistral Large 4 (Note: "A trillion open parameters").

gem-analytics.github.iogithub.com

  1. First step

    Write 20 single-layer tasks and compute their answers with GEMAct.

  2. Done looks like

    200 tasks covering quota share, per-risk and catastrophe excess of loss, each with an answer anyone can reproduce from code; scores for at least three AI models, counting a price as right if it is within 2% of the computed one.

Open 02 · the brief

Expected loss, checked by modeler

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public table comparing, for each catastrophe modeling firm, the expected losses it gave for the cat bonds it modeled with what those bonds actually lost. Expected loss is the average yearly loss a model predicts, and it largely sets the price investors accept.

Why it is still open

Lane Financial compared expected with actual losses for 404 cat bond tranches up to mid-2012 but did not "distinguish between the different risk modellers", Artemis reported. Plenum Investments, a fund manager, says expected losses can differ "from one model view of risk to another".

artemis.bmartemis.bm

Built with what exists

the Artemis deal directory, which names each bond's modeling firm and often its expected loss, used with Artemis's permission; Artemis's list of triggered cat bonds; Python.

artemis.bm

  1. First step

    Ask Artemis for permission to use its deal-directory fields, then build the rows for bonds outstanding in 2017, a year of heavy losses.

  2. Done looks like

    At least 500 tranches since 1997, each with its modeler, expected loss and outcome linked to a source; for each modeler, the ratio of actual to expected loss with an uncertainty range.