The Credibility Report
Edition 13: TabPFN v4, Temporal Fusion, and Credibility Networks
TabPFN v4, temporal fusion transformer, and credibility networks.
🔔 Opening Bell
This week’s signal is risk retention vs risk transfer.
Munich Re is explicitly choosing to keep more margin “in house” by cutting retrocession and walking away from collateralised sidecars. Meanwhile, specialty platforms are still leaning into transparent, index-based ILS structures (e.g. Lumen Re’s Photon Re cat bond) to manage peak tail risk.
Quietly, a reminder from workers’ compensation: you don’t need hype to make underwriting money — a decade of profitability in a line with falling frequency is exactly the kind of ballast you want when cat volatility flares.
📊 This Week's Headlines
Munich Re cuts retrocession and scraps sidecars to retain underwriting economics
Munich Re reported a record €6.1bn 2025 result and is signalling confidence in its balance sheet by reducing its 2026 retro program to $600m and not renewing its sidecar/cat bond components. The actuarial angle isn’t market-timing — it’s how re/insurers are trading off IFRS earnings volatility, capital deployment, and margin retention.
Why actuaries should care: retro is becoming a governance artifact: volatility budget + capital narrative + earnings stabilisation logic.
Source: Artemis.bm
Convex: growth to ~$6bn GWP with an 89% combined ratio
Convex reported $5.9bn GWP (up 14% YoY), $711m net income, and a combined ratio of 89%. Useful calibration for what “good” underwriting looks like in specialty re/insurance as markets start talking about modest rate declines.
Why actuaries should care: a concrete yardstick for pricing adequacy vs cat attritional drift as rate momentum turns.
Source: Reinsurance News
Lumen Re’s Photon Re: $175m index-based cat bond as a “natural evolution” of retro strategy
Photon Re provides $175m of collateralised, index-based protection over U.S./Canada hurricane and earthquake across two classes, over a four-year term. Demand remains strong for transparent structures with clear triggers.
Why actuaries should care: index triggers are a live trade-off between basis risk and speed/certainty of settlement — increasingly part of mainstream retro programs.
Source: Reinsurance News
Cat bond secondary market stayed “bid-heavy” in 2025; spreads tightened
Swiss Re Capital Markets describes a secondary market that was often seller-friendly with excess investor cash and limited offers. The result: tightening spreads and episodic liquidity around primary issuance bursts.
Why actuaries should care: changes the implied cost of ILS capacity and the expected stability of renewal pricing vs traditional CatXL.
Source: Artemis.bm
Workers’ comp: decade-long underwriting profit, frequency down, severity up (mostly inflation)
Triple‑I highlights workers’ comp as an outlier line with sustained underwriting profit: 2024 net combined ratio 87.8, frequency declining at -5.6% CAGR (2015–2024), severity rising, but severity growth looking more like economy-wide inflation than worsening safety.
Why actuaries should care: in portfolio steering, workers’ comp behaves like a stabiliser line and a clean case study in separating frequency vs severity drivers.
Source: Triple‑I
📚 Research Spotlight
Expectiles as basis risk-optimal payment schemes in parametric insurance
A clean link between parametric payout design and basis-risk minimisation. The paper frames basis risk using an asymmetric squared-error loss and shows how the basis-risk-optimal payout can be characterised via conditional expectiles of the true loss given the index/trigger state. The governance angle is strong: expectiles are elicitable (principled backtesting), which makes them unusually defensible when product committees ask “how do we validate that this payout rule is fair and stable?”
Additional papers worth skimming
Simulations of Bivariate Archimedean Copulas … for Loss Reserving under Flexible Censoring
A reserving-adjacent contribution: copula simulation under censoring constraints (useful when dependence matters and data is incomplete).
Validation of machine learning based scenario generators
If you’re using ML scenario generators (economic, cat, synthetic portfolios): this is about validation—exactly the part governance will ask you to defend.
Could Large Language Models work as Post-hoc Explainability Tools in Credit Risk Models?
Practical question for model risk: can LLMs help translate model behaviour into human-readable explanations without hallucinating?
💡 Practical Takeaways
📈 Pricing / ERM
- Treat retro spend as an explicit volatility budget decision. Make it auditable: what volatility is being bought down, at what margin cost?
- When rate momentum turns, add an “attritional drift” lens: how much of the combined ratio is cat vs underlying adequacy?
- For parametric covers: basis risk needs an explainable metric. Expectiles are interesting because they’re principled and backtestable.
🧠 Cat / ILS
- When comparing cat bonds vs CatXL: include basis risk, settlement mechanics, and renewal stability — not just RoL.
- Secondary market “bid-heavy” conditions can compress implied ILS costs and change renewal dynamics.
- Index triggers belong in the same governance lane as cat models: assumptions, validation, and post-event review.
🔮 What We're Watching
- Retro in softening markets: do other large reinsurers follow Munich Re in retaining more risk?
- ILS breadth: more first-time sponsors vs repeat sponsors pushing size.
- Workers’ comp trend: does frequency improvement persist while severity inflation continues?
- Parametric governance: do basis-risk metrics move from “academic footnote” to product approval gate?