Munich Re
Natural Catastrophe Pricing with AI: Improving Cat Loss Models
Business Context & Strategic Drivers
Natural catastrophe losses have more than tripled in the past 30 years, driven by climate change and rapid urbanisation into hazard-exposed areas. Munich Re, as the world's largest reinsurer ($50B+ premiums written), bears enormous nat cat risk concentration. Improving pricing accuracy by even 1pp across its nat cat portfolio translates to hundreds of millions in underwriting profit.
Strategic Drivers
- Climate change making historical vendor cat model assumptions increasingly unreliable
- Primary insurers withdrawing from high-risk markets (California wildfire, Florida hurricane) creating reinsurance market opportunity for those who can price it
- Satellite imagery and climate data now available at resolutions unimaginable when vendor models were built
- Munich Re's $50B premium base means 1pp combined ratio improvement = $500M in profit
- ILS and capital markets pricing discipline requiring Munich Re to demonstrate proprietary data edge
The Problem
Traditional natural catastrophe (nat cat) pricing relies on vendor models (RMS, AIR, Verisk) that are updated infrequently, use historical loss data that may not reflect forward-looking climate risk, and cannot incorporate the full range of real-time earth observation and climate data now available. Munich Re's underwriters had limited ability to challenge or supplement vendor model outputs.
The Solution
Munich Re's Digital Partners and NatCat teams built a proprietary AI-enhanced catastrophe pricing layer - 'geo:unison' - that ingests satellite imagery, climate model outputs, real-time weather data, and Munich Re's proprietary 50-year nat cat loss experience to produce location-level risk scores that supplement and challenge vendor cat model outputs.
Technical Architecture
Tech Stack
Architecture Overview
A geospatial data pipeline ingests satellite imagery (Sentinel-2, 10m resolution globally, updated every 5 days), climate model projections, and elevation/topology data for every 30m grid cell globally. A convolutional neural network extracts hazard-relevant features (vegetation index, impervious surface, slope, proximity to water). These features feed into peril-specific gradient boosting models (wildfire, flood, wind) trained on Munich Re's 50-year loss history. Output is a location-level annual expected loss (AEL) and return-period loss curve that underwriters use alongside vendor cat model outputs.
Data Requirements
50 years of Munich Re proprietary nat cat loss data (global, peril-specific). 5 years of Sentinel-2 satellite imagery (petabyte scale - processed via Google Earth Engine). CMIP6 climate projections to 2100. Third-party property exposure databases (CoreLogic, Verisk). All proprietary data processed within Munich Re's private cloud.
ROI & Financial Analysis
Investment
$50–70M over 4 years (data infrastructure, model development, satellite data licensing, actuarial validation)
Annual Return
$400–500M attributable combined ratio improvement on nat cat portfolio
Payback
~12 months post full integration into underwriting
ROI Multiple
15–20x over 5 years
ROI Breakdown
Nat cat underwriting profit improvement
3–4pp combined ratio on $10B+ nat cat premium book
$400M+/year
Retrocession cost optimisation
Better risk stratification allows more efficient retrocession purchasing and ILS structuring
$60M/year
Market opportunity capture
Ability to selectively write risks in markets competitors have exited (CA wildfire) at adequate pricing
$40M/year
Implementation Journey
Total timeline: 48 months from research project to production underwriting integration
Research and Proof of Concept
12 monthsMunich Re's research team validated that satellite-derived features could predict nat cat losses better than terrain data alone. Proof of concept on California wildfire using 2017–2020 fire seasons.
Global Data Infrastructure
12 monthsBuilt petabyte-scale satellite data pipeline on Google Earth Engine and Databricks. Ingested and normalised 50-year proprietary loss database. Built PostGIS spatial database for global property exposure.
Peril Model Development
12 monthsDeveloped and validated models for wildfire, riverine flood, coastal flood, and wind. Each model validated against held-out historical loss events and reviewed by Munich Re's actuarial council.
Underwriting Integration & Rollout
12 monthsIntegrated geo:unison outputs into Munich Re's treaty and facultative underwriting platforms. Underwriter training on interpreting AI risk scores alongside vendor models. Rollout to European, US, and Asia-Pacific books.
Challenges Overcome
- 1Petabyte-scale satellite data processing: Processing global Sentinel-2 imagery required significant investment in Google Earth Engine and Databricks infrastructure
- 2Actuarial validation rigour: Munich Re's actuarial council required the model to meet the same standards as Solvency II internal models, requiring 12 months of backtesting
- 3Vendor model comparison: Demonstrating proprietary model outperformance required agreed-upon methodology - contentious as vendor models are black boxes
- 4Underwriter workflow integration: Underwriters were accustomed to vendor model outputs and needed guidance on how to use and weight an additional proprietary score
- 5Climate model uncertainty: CMIP6 scenarios have wide uncertainty ranges for 2050+ projections, requiring the model to express uncertainty rather than point estimates
Governance & Oversight
Governance Controls
- Proprietary model used as supplementary tool - vendor cat models still required for all treaty pricing
- Annual actuarial council review of model performance vs. actual loss emergence
- Independent external validation every 3 years
- Underwriter override documentation when deviating from AI-indicated pricing
- Catastrophe event post-mortems: every major event triggers model performance review and recalibration
Data Privacy Measures
- No individual policyholder data used in model training - only aggregate loss event data
- Satellite imagery is public domain (Sentinel-2, Landsat) or licensed without privacy constraints
- Property database data licensed from CoreLogic/Verisk under enterprise agreements with data security provisions
- GDPR-compliant for any European property data
Human-in-the-Loop
All underwriting decisions are made by qualified cat underwriters. geo:unison provides an additional data point for underwriter judgement, not a replacement. Treaty pricing above €100M requires Head of NatCat Underwriting review regardless of model output. All model outputs reviewed by actuarial team before any pricing authority limits are adjusted.
Regulatory Considerations
- Solvency II internal model governance if geo:unison outputs are used in SCR calculations
- IAIS standards on climate risk in insurance supervision
- TCFD-aligned climate risk disclosure requirements
- EU Taxonomy Regulation on sustainable finance (for climate risk reporting)
- Lloyd's Market Oversight on cat model usage requirements
Lessons Learned
Key Lessons
- Publishing research findings externally (academic papers) accelerated credibility with regulators and internal actuarial teams
- Petabyte-scale geospatial ML requires a dedicated platform team - this cannot be done with standard data science infrastructure
- The model's value is highest at the sub-grid level where vendor models are weakest - target integration to those use cases first
- Climate model uncertainty must be expressed explicitly in outputs, not hidden - underwriters need to understand the forward-looking uncertainty range
- Validating against individual historical events (Hurricane Harvey, 2018 California wildfires) was more persuasive than aggregate statistics for underwriter adoption
What Worked Well
- Google Earth Engine handled the global satellite processing workload without requiring Munich Re to build custom distributed computing infrastructure
- Publishing wildfire PoC as a research paper generated external credibility and attracted top geospatial ML talent
- Integrating proprietary data as a supplement to (not replacement for) vendor models reduced underwriter resistance significantly
The Outcome
Location-level flood and wildfire risk scores now available for 1 billion+ properties globally. Underwriters can price individual risk locations that vendor models cannot resolve at sufficient granularity. Portfolio aggregation losses improved, contributing to a combined ratio improvement of 3–4 percentage points on nat cat-exposed books.
Key Metrics
- Risk scores available for 1B+ properties globally
- Combined ratio improvement of 3–4pp on nat cat books
- Wildfire risk score outperforms vendor models by 15% on out-of-sample loss prediction
- Flood model validated against 50 years of Munich Re's proprietary loss data
- Pricing cycle for complex nat cat layers reduced from 5 days to same day
References & Further Reading
Quick Stats
Company
Munich Re
Industry
Team Size
45 engineers, 20 data scientists, 12 geospatial specialists, 15 actuaries, 8 NatCat underwriting domain experts
Timeline
48 months from research project to production underwriting integration
Investment
$50–70M over 4 years (data infrastructure, model development, satellite data licensing, actuarial validation)
Annual Return
$400–500M attributable combined ratio improvement on nat cat portfolio
Payback Period
~12 months post full integration into underwriting
Key Metrics
- Risk scores available for 1B+ properties globally
- Combined ratio improvement of 3–4pp on nat cat books
- Wildfire risk score outperforms vendor models by 15% on out-of-sample loss prediction
- Flood model validated against 50 years of Munich Re's proprietary loss data
- Pricing cycle for complex nat cat layers reduced from 5 days to same day
Tech Stack
ROI figures and metrics are based on publicly available data, company disclosures, and reasonable estimates. Always conduct your own due diligence for strategic decisions.