The Credibility Report
Edition 14: Deep Learning Reserving, ICN Attention, and the Future of Pricing
Deep learning for reserving, ICN attention networks, and future pricing trends.
🔔 Opening Bell
This week’s signal is retrocession as a capital narrative.
We’re seeing multiple players emphasise balance sheet resilience via tighter, more deliberate retro structures. In parallel, the research stack keeps pushing on tail risk (flood), ruin, and mortality forecasting — the quantitative backbone that ultimately justifies those capital decisions.
The practical takeaway: treat retro not as a one-off purchasing exercise, but as an auditable volatility budget that has to map cleanly to model evidence.
📊 This Week's Headlines
Swiss Re reports strong profits, stable renewals, and reduced external nat cat retro
Swiss Re’s FY2025 communication is a useful read-through for how a major reinsurer is framing results and renewals while also signalling a shift in retro posture. The actuarial angle: what’s being bought (or not bought) in retro shows up later as a volatility and capital management choice, not just a market view.
Why actuaries should care: retro strategy is increasingly inseparable from earnings volatility management, model governance, and the story told to investors/regulators.
Source: Swiss Re (press release)
Pool Re successfully completes renewal of a £2.75bn retrocession placement
Pool Re announced completion of its retro renewal sized at £2.75bn. Even if you don’t work in terrorism pools, the structure is a reminder that the “retro supply chain” has become a core part of public-private risk mechanisms.
Why actuaries should care: helps calibrate how governments and mutual structures buy volatility protection (and what scale of capital they need to warehouse tail risk).
Source: Pool Re (press release)
Active Re completes Global Retrocession Programme; highlights broader protection & AI integration
Active Re reported completion of its Global Retrocession Programme with a focus on more comprehensive scope/territory and an emphasis on execution quality. Notably, it also describes progress on AI integration into technical processes (a reminder: automation is becoming part of operational resilience).
Why actuaries should care: retro is being positioned as an all-cycle capital protection layer, while AI is moving from “innovation” to control and consistency in operations.
Source: Active Re
📚 Research Spotlight (arXiv • last 14 days)
Contributions of geolocated weather and building related data for insurance assessment of flood risks
Flood is one of the most expensive and model-sensitive perils: this paper sits in the practical intersection of geospatial exposure data, hazard features, and insurance assessment. It’s worth skimming if you’re calibrating vulnerability, pricing, or accumulation views for flood across heterogeneous portfolios.
Additional papers worth skimming
Asymptotics of Ruin Probabilities in a Subordinated Cramér-Lundberg Model
arXiv • Score: 26
Classic actuarial DNA: how heavy-tailed / subordinated dynamics change the tail behaviour of ruin probabilities (and therefore capital intuition).
Hidden multistate models to study multimorbidity trajectories
arXiv • Score: 18
Potentially useful for health/life: multi-state modelling choices matter when transitions are partially observed and latent state is real.
Enhancing Mortality Forecasting with Ensemble Learning: A Shapley-Based Approach
arXiv • Score: 12
Mortality forecasting meets explainability: ensemble performance plus Shapley-style contribution analysis can help governance and model risk sign-off.
A Dirichlet-Multinomial-Poisson framework for the coherent analysis and forecast of cause-specific mortality
arXiv • Score: 12
Cause-of-death modelling is a recurring pain point: coherent multi-cause frameworks reduce “inconsistent totals” when you forecast by cause and then aggregate.
Bayesian Profile Regression using Variational Inference to Identify Clusters of Multiple Long-Term Conditions Conditioning on Mortality in Population-Scale Data
arXiv • Score: 12
Segmentation with mortality conditioning: relevant for health pricing and longevity studies where comorbidities drive heterogeneous risk.
💡 Practical Takeaways
🧾 Reinsurance / Capital
- Write down a retro “volatility budget”: what volatility is being reduced (earnings, SCR, rating agency), and what margin is being traded away.
- Stress test retro decisions using scenario families (not single events): e.g., clustered flood years, or multi-peril seasons with correlated loss creep.
- When using new tooling (AI/automation) in technical operations, treat it like a model: document controls, error rates, and fallback procedures.
🧬 Life / Health Analytics
- If you’re doing mortality work: test whether ensemble methods improve stability, but insist on explainability artifacts you can show to governance.
- For multi-cause mortality modelling: validate coherence by checking that cause-level forecasts re-aggregate correctly under perturbations.
- For multimorbidity / comorbidity segmentation: sanity-check clusters against clinical intuition and ensure anti-leakage in feature engineering.
🔮 What We're Watching
- Retro appetite vs retention: how many large players choose to buy less external nat cat protection as pricing normalises?
- Public-private tail risk: continued reliance on retro for pools and quasi-government structures (scale, structure, and counterparty concentration).
- Flood model evidence: do insurers materially upgrade exposure data (buildings + geolocated weather) to reduce tail uncertainty?
- Mortality & health segmentation: more multi-state and multi-cause models in production — and the governance burden they imply.