The Credibility Report
AI-curated actuarial intelligence — Edition 3
Welcome to Edition 3 of The Credibility Report — curated intelligence for actuaries who care about what's happening at the intersection of insurance, machine learning, and quantitative risk.
This week: Swiss Re & Munich Re's sobering 2025 nat cat reviews, Travelers buying down retention, Lemonade's autonomous car insurance, AI-powered hurricane forecasting, and climate resilience bonds.
📊 This Week's Headlines
🔥 Swiss Re & Munich Re: 2025 Nat Cat Losses Exceed $140bn
Sixth consecutive year of $100bn+ insured natural catastrophe losses. California wildfires and intense hurricanes drove a devastating 2025. Climate risks are "redrawing the boundaries of the insurance market."
🛡️ Travelers Buys $1bn Lower Layer Cat Reinsurance
The largest US P&C insurer is reducing retention and increasing protection — a notable shift signaling primary insurers are buying more coverage despite the softer market.
Read more → Artemis.bm🚗 Lemonade + Tesla: Autonomous Car Insurance
Lemonade launches the first insurance product designed specifically for self-driving vehicles, starting with Tesla FSD. Per-mile pricing with significantly lower rates for FSD users.
Why actuaries should care: Traditional auto pricing assumes human drivers. Autonomous vehicles introduce fundamentally different risk profiles — potentially much lower frequency, but different severity characteristics and new liability questions.
Read more → Reinsurance News🌀 AI Hurricane Forecasting Goes Mainstream
Gallagher Re reports the 2025 hurricane season marked a "new era" — official agencies now use AI models for storm formation, track, and intensity prediction. This isn't experimental anymore.
Read more → Reinsurance News🔐 Cyber Reinsurance: Buyers Secure Favorable Terms
Cyber reinsurance buyers secured favorable terms at early 2026 renewals. The market continues to mature with improving loss experience and growing capacity.
Also this week: Howden Re launches H2 2025 Cyber Threat Report • AI risks bring new scrutiny to cyber insurance policies
📚 Research Spotlight
A Quantitative Model for Climate Change Adaptation Resilience Bonds
European Actuarial Journal
This paper introduces a model for resilience bonds — a new instrument designed to support climate change adaptation rather than just transfer risk.
The key insight: As climate risks become increasingly uninsurable through traditional means, alternative instruments that fund adaptation measures become critical. The paper proposes a quantitative framework for pricing and structuring these bonds, focusing on the interaction between insurance companies and public authorities.
Why it matters: This is where cat bonds meet climate adaptation. The model addresses the fundamental question: how do we price instruments that reduce future risk, not just transfer current risk?
Read the paper →Deep Learning-Copula for Climate Insurance
arXiv
A Canadian Prairies case study (2002-2011) combining deep neural networks with copula-based multivariate analysis for precipitation-driven claims.
What's new: The framework models precipitation intensity/duration and claim severity jointly — capturing threshold effects where claims escalate nonlinearly above certain precipitation levels.
Read paper →Graph Neural Networks for Credit Default
arXiv • Score: 16
Credit default risk arises from complex interactions among borrowers, financial institutions, and transaction-level behaviors. This paper proposes heterogeneous graph neural networks that explicitly capture cross-entity dependencies — something traditional tabular models miss.
Why actuaries should care: Credit risk modeling is becoming a network problem. Graph-based approaches may outperform feature-engineered tabular models for portfolios with complex counterparty relationships.
Read paper →Bayesian Cost-Effectiveness Analysis
arXiv • Score: 16
Cost-effectiveness analyses (CEAs) compare costs and health outcomes to inform medical decisions. This paper develops a Bayesian framework that handles nonrandom treatment assignment, administrative censoring, and irregularly spaced data.
Applications: Health insurance pricing, disease management program evaluation, value-based care contracting.
Read paper →Tail Risk: Empirical VaR Bias
arXiv • Score: 15
Analytical bounds on the bias of empirical TVaR estimators. Essential reading for anyone implementing internal models under Solvency II or validating regulatory capital calculations.
Dynamic Insurance Market Games
arXiv • Score: 12
A dynamic insurance market model with two competing insurers and strategic underreporting by insureds. The paper examines equilibrium outcomes — relevant for understanding adverse selection in competitive markets.
Read paper →🔬 Deep Dive: The 2025 Nat Cat Reckoning
Both Swiss Re and Munich Re released their 2025 natural catastrophe reviews this week. The numbers are stark:
Key takeaways:
- $140bn+ insured losses in 2025 — The California wildfires and intense hurricane season pushed the total well above the $100bn threshold that's now become the norm rather than the exception.
- Sixth consecutive year above $100bn — This is no longer an outlier year. The baseline has shifted. Actuaries building cat models need to recalibrate their view of "average" loss years.
- Thunderstorms as emerging peril — Munich Re specifically calls out intense thunderstorms (including severe convective storm) as an exacerbating factor. This is harder to model than hurricanes and often excluded from traditional cat definitions.
- Wildfire as uninsurable risk? — California wildfire losses continue to challenge insurability. Several major insurers have exited the California homeowners market, and remaining capacity is priced at levels many homeowners can't afford.
Implications for reinsurers:
The industry has adapted — higher retentions, tighter terms, expanded ILS capacity. But if $140bn becomes the new normal, even current pricing may prove inadequate.
Implications for primary insurers:
Travelers' purchase of additional $1bn in lower-layer cat reinsurance signals a shift in appetite. Expect more cedants to buy down their retentions in 2026.
🏛️ Market Intelligence
Fitch: Softer US P&C Pricing Expected in 2026
Fitch sees softer US P&C insurance pricing ahead, with reinsurer outlook stable but challenged. The rating agency expects:
- Continued rate moderation in property lines
- Casualty reserve development to remain a key watch item
- Social inflation pressures persistent
Life Reinsurance Market to Reach $904bn by 2035
A new market sizing report projects the global life reinsurance market will grow to $904.43 billion by 2035, driven by aging populations and increased demand for longevity risk transfer.
Cat Bond Pricing Rises
Aetna's Vitality Re XVII health cat bond saw guide prices rise across all tranches. SafePoint returns with a new $150m Nature Coast Re 2026-1 named storm cat bond.
Zurich Lloyd's Syndicate
Zurich confirms plans to launch a Lloyd's syndicate, joining the wave of major insurers establishing London market presence.
🌍 Global Markets
🇮🇳 India: The Growth Outlier
Swiss Re forecasts India's insurance market to outpace all major markets in the coming years. With a low insurance penetration rate and a rapidly growing middle class, India represents the largest greenfield opportunity in global insurance.
- Life insurance penetration still below 4% of GDP
- Motor insurance growing with vehicle sales
- Health insurance expanding with government schemes
- Agricultural insurance modernizing with parametric products
🤖 AI & Technology
The End of Human-Only Meteorology
Gallagher Re's analysis of the 2025 hurricane season marks a turning point: AI models are now part of official forecasting. This isn't a research project anymore — it's operational.
- Cat model validation — If official forecasts use AI, cat model vendors need to incorporate the same or better methods
- Real-time exposure management — AI models provide faster updates, enabling more dynamic risk management
- Model risk — AI forecasts introduce new model risk considerations — how do you validate a neural network's hurricane track prediction?
Generative Models: When Pretty Isn't Enough
The Stylized Facts Alignment GAN (SFAG) paper (arXiv) tackles a critical problem: generative models for financial time series often look realistic but fail completely in backtesting.
The insight: Standard GANs optimize for distributional matching — making generated data look like historical data. But financial applications care about tail behavior, asymmetries, and trading signals. SFAG introduces structural constraints that preserve these properties.
Why actuaries should care: If you're using synthetic data for stress testing, scenario generation, or model training, this paper is essential reading. Your synthetic data may look great but be worthless for risk assessment.
💡 Practical Takeaways
For Pricing Actuaries
- Graph neural networks are showing promise for credit risk — worth experimenting if you have network/relationship data
- Copula + deep learning hybrids continue to prove their value for climate-related risks
For Cat Modelers
- AI-enhanced forecasting is now mainstream — update your validation frameworks
- Sixth consecutive $100bn+ year should inform your long-term view assumptions
For Reserve Actuaries
- Cost-effectiveness analysis methods are evolving — the Bayesian CEA framework handles real-world data messiness well
- Time-varying treatment effects are the norm in health insurance data
For ERM
- Resilience bonds are an emerging asset class — worth understanding even if not ready for portfolio inclusion
- Cyber reinsurance market maturing — favorable buyer terms available
🎯 One Thing to Try This Week
Graph-based credit analysis
If you have borrower relationship data (co-signers, employers, guarantors, transaction networks), experiment with a simple graph embedding:
- Build a network: nodes = borrowers, edges = relationships
- Run node2vec or a basic GNN to get embeddings
- Add embeddings as features to your existing XGBoost model
- Compare performance
The research suggests this can unlock predictive signal that traditional feature engineering misses — especially for thin-file borrowers where relationship signals matter more.
📄 From the arXiv
| Paper | Score | Focus | Link |
|---|---|---|---|
| Deep Learning-Copula for Climate Insurance | 32 | Climate risk, claims modeling | → |
| Inverting Self-Organizing Maps | 28 | Clustering, dimensionality | → |
| Field-Space Autoencoder for Climate Emulators | 23 | Climate, neural networks | → |
| Beyond Visual Realism: Financial Time Series | 22 | Synthetic data, backtesting | → |
| Intermittent Time Series: Local vs Global | 22 | Inventory, forecasting | → |
| Bayesian Cost-Effectiveness | 16 | Health, CEA | → |
| Graph Neural Networks for Credit Default | 16 | Credit risk, GNN | → |
🔮 What We're Watching
📊 By the Numbers
🎓 Conference Spotlight
1st ASTIN Bulletin Conference
January 14-16, 2026 | ETH Zurich, Switzerland
The inaugural ASTIN Bulletin Conference brought together actuarial researchers and practitioners at ETH Zurich — building a stronger community of young scholars in actuarial science.
Keynote Highlights
Research Themes
Chaired by Mario Wüthrich (ETH Zurich). The ASTIN Bulletin, published by the IAA, is the premier journal for mathematical methods in insurance.