The Credibility Report
AI-curated actuarial intelligence — Edition 1
Welcome to the very first edition of The Credibility Report — a curated newsletter for actuaries who care about what's actually happening at the intersection of insurance, machine learning, and quantitative risk.
This week: 1/1 renewal analysis, diffusion models for reserving, transformer-based dependent risk pricing, Colorado AI regulation deep-dive, and 355 articles scored for actuarial relevance.
📊 The 1/1 Renewals: A Soft Market Has Landed
The January 1st reinsurance renewals confirmed what the market has been whispering for months: the hard market is over. Capacity is back, and cedants are finally in the driver's seat.
"Soft market to drive lacklustre margins for P&C reinsurers in 2026" — J.P. Morgan
The CAS Actuarial Review published a thoughtful analysis: "An Insurance Soft Market Landing" — arguing this isn't a crash, but a controlled descent.
Howden Re: "Re-Balancing"
Risk-adjusted prices have returned to levels last seen around four years ago (2021-2022).
- • Risk-adjusted rate cuts across most major lines
- • Market softening but still "rational"
- • Attachments remain elevated vs pre-2020
Guy Carpenter
Global reinsurer capital: USD 660 billion
- • Launched Sidecar Center of Excellence
- • Property retrocession demand increases
- • Alternative capital selective but persistent
Aon
Global reinsurer capital: USD 735 billion (highest estimate)
- • Increased capacity, high overall demand
- • Favourable conditions for cedents in casualty
- • Strong profitability supporting capital formation
Gallagher Re
Global reinsurer capital: USD 710 billion
- • Record capital across traditional & alternative
- • Competitive pressure intensifying
- • Selective deployment in loss-affected programs
By the Numbers
Risk-adjusted prices
Back to 2021-2022 levels
Global reinsurance capital
$660B - $735B
Turkey property-cat
Rates fell 15%
Cedant wins
Double-digit cuts secured
The ILS Angle
Despite the soft primary market, alternative capital remains robust. Artemis reports that reinsurance returns remain "still-strong" — retrocession holding up better than primary reinsurance.
Also worth reading: Catastrophe bond pricing under the renewal process — Scandinavian Actuarial Journal
📚 Research Spotlight
Advancing Zero-Inflated Tweedie Models and Evaluating Gradient Boosting
North American Actuarial Journal
Tackles one of pricing's thorniest problems: datasets with excess zeros. Compares gradient boosting implementations (XGBoost, LightGBM) against traditional Tweedie GLMs for zero-inflated data.
Read the paper →Probabilistic Loss Reserving via Denoising Diffusion Model
Insurance: Mathematics and Economics
Diffusion models — the same technology powering image generators like Stable Diffusion and DALL-E — now applied to claims reserving. This paper shows how to generate full probabilistic reserve distributions using denoising diffusion. Unlike bootstrap or Mack, which make strong distributional assumptions, diffusion models learn the underlying distribution directly from data.
Why actuaries should care: Stochastic reserving is about to get a serious upgrade. If you're still relying on Mack or ODP bootstrap for your reserve ranges, this paper shows what's possible with modern generative methods.
Read the paper →Comparing Predictive Models for Dependent Risk Pricing
Variance
When risks aren't independent — and they rarely are — standard GLM pricing breaks down. This paper compares approaches for pricing dependent risks, including transformer architectures that can capture complex dependency structures.
Read →Related reading:
- Bivariate phase-type distributions for experience rating — European Actuarial Journal
- Simulations of Bivariate Archimedean Copulas — NAAJ
- Bayesian Additive Regression Tree Copula Processes — arXiv
Development of Telematics Safety Scores in Accordance with GLM Frameworks
Variance
The CAS journal delivers a practical guide to building neural network models that play nicely with traditional GLM interpretability requirements. If you've ever tried to explain a black-box model to your regulator — or your Chief Actuary — this paper is your blueprint.
Key insight: You can layer neural network features onto a GLM backbone, getting the predictive power of deep learning with the interpretability regulators demand.
Read the paper →A Hybrid Architecture for Multi-Stage Claim Document Understanding
arXiv
This paper combines vision-language models (think GPT-4V) with traditional ML for real-time claims document processing. The architecture handles the full pipeline: document intake, classification, extraction, and validation.
Why it matters: Claims processing is still drowning in paper and PDFs. This hybrid approach — using LLMs for understanding and ML for structured extraction — is exactly where the industry is heading.
Practical applications: FNOL automation, subrogation document analysis, medical claims extraction, fraud detection pipelines.
Read the paper →Assessing Regional Flood Risks Under Climate Change
Geneva Papers on Risk and Insurance Score: 30
One of our highest-rated papers this month. A machine learning approach combining neural networks and random forests for flood risk assessment under climate scenarios. The authors tackle the challenge of limited historical data for unprecedented climate events by using ensemble methods to quantify uncertainty.
Practical applications: Catastrophe model validation, climate-adjusted pricing, regulatory stress testing.
Read the paper →📄 From the arXiv
This week's most relevant preprints for actuaries — filtered from 140+ papers.
| Paper | Score | Keywords | Link |
|---|---|---|---|
| Stochastic Deep Learning: A Probabilistic Framework | 31 | deep learning, neural network, uncertainty | → |
| Forecasting the U.S. Treasury Yield Curve | 24 | ML, random forest, time series | → |
| Collapsed Structured Block Models for Community Detection | 21 | interpretable, GAM, Bayesian | → |
| Decision-Theoretic Robustness for Network Models | 21 | Bayesian, dependence, SCR | → |
| Bayesian Nonparametric Modeling of Dynamic Pollution | 20 | GAM, Bayesian, prior | → |
| Bayesian Multiple Multivariate Density-Density Regression | 18 | interpretable, MCMC | → |
| HUR-MACL: High-Uncertainty Region-Guided Multi-Architecture | 17 | deep learning, SHAP, risk | → |
| Hybrid Architecture for Claim Document Understanding | 12 | ML, claims, VLM | → |
Scoring: Papers ranked by actuarial relevance (ML methods + insurance/actuarial terms).
🏛️ Regulatory & Professional Updates
Colorado AI Governance: What You Need to Know
Colorado's SB21-169 is now the most comprehensive US state regulation of algorithmic decision-making in insurance. Originally focused on life insurance, it now extends to auto and health.
- Test algorithms for unfair discrimination before deployment
- "Proxy discrimination" explicitly covered
- External audits may be required for high-impact models
- Extensive documentation requirements
📅 Conference Calendar
Upcoming events for your calendar:
| Event | Dates | Location |
|---|---|---|
| CAS Ratemaking Seminar | March 23-25, 2026 | Orlando, FL |
| RIMS Annual Conference | April 2026 | TBD |
| CAS Spring Meeting | May 2026 | TBD |
| Monte Carlo Rendez-Vous | September 6-10, 2026 | Monaco |
| SOA Annual Meeting | October 2026 | TBD |
| InsureTech Connect | October 2026 | Las Vegas |
| Baden-Baden Meeting | October 2026 | Germany |
| ASSA Convention | Oct/Nov 2026 | South Africa |
| CAS Annual Meeting | November 2026 | TBD |
Dates updated as organizations announce.
🔮 What We're Watching
- How the soft market evolves through Q1 renewals
- IFRS 17 Year 2 results — what are insurers learning from full implementation?
- More diffusion model applications in actuarial contexts
- US state AI regulation — who follows Colorado's lead?