Pricing a non-stationary world
Published on July 24, 2025
This analysis presents a comprehensive analysis of the escalating crisis in property insurance, driven by the increasing frequency and severity of climate-related catastrophes. The industry is at a critical inflection point where traditional actuarial methods and current-generation catastrophe models are proving insufficient to manage the non-stationary nature of climate risk. This article deconstructs the key drivers of rising premiums, provides a technical deep-dive into catastrophe modeling, and evaluates the market-warping effects of state-level rate regulation.
Section 1: The New Economics of Property Risk
The P&C insurance market is undergoing a fundamental repricing of risk. The evidence is unequivocal, with rising losses driving premium hikes and market exits. This section establishes the empirical basis of the crisis and deconstructs its underlying causes.
U.S. Homeowner Premium Increases
Average premiums surged over 30% from 2020-2023. Source: U.S. Treasury.
Frequency of Billion-Dollar Disasters
The average interval has shrunk from months to weeks. Source: NOAA.
Deconstructing Loss Cost Inflation
Rising losses are a product of escalating physical hazards interacting with evolving socioeconomic landscapes. Click each driver to understand its role.
Physical Risk
Socioeconomic Factors
Reinsurance
Select a driver above to see the details.
Section 2: The Modern Actuarial Toolkit
Traditional ratemaking is challenged by a non-stationary climate. This has led to the adoption of sophisticated Catastrophe (CAT) models to simulate future possibilities. This section deconstructs these models and their statistical underpinnings.
Anatomy of a Catastrophe (CAT) Model
The scientific core. It generates a large set of "synthetic" but physically realistic catastrophic events (hurricanes, wildfires, etc.) based on meteorology, geology, and geography. This is not a forecast, but a representation of what *could* happen over a long period, including events more severe than any in the historical record.
A detailed inventory of the properties being analyzed. Ideally, this includes precise geocoded locations, full replacement values, and physical characteristics like construction type and age, along with specific insurance policy terms (deductibles, limits).
Connects the hazard to the exposure using "damage functions." These mathematical relationships, derived from engineering and claims data, quantify the expected damage a specific property will sustain from a given hazard intensity (e.g., a 120 mph wind causes 25% damage).
Translates physical damage into a monetary insurance loss. It applies the policy terms (deductibles, limits) to the damage estimate. Running this for thousands of simulated events produces key risk metrics like Average Annual Loss (AAL) and Probable Maximum Loss (PML).
Statistical Underpinnings for the Model Developer
| Technique | Interpretability | Strengths | Weaknesses |
|---|---|---|---|
| GLM | High | Industry standard, transparent | Misses non-linearities |
| ARIMA | Moderate | Models trends & seasonality | Struggles with non-stationarity |
| SVR | Low | Models non-linear relationships | Computationally intensive |
| Neural Networks | Very Low | Models high complexity | "Black box," needs large data |
| Gradient Boosting | Moderate | High predictive accuracy | Can overfit, less interpretable |
Section 3 & 4: Model Critiques & Data Deficits
A model is only as good as its data and its assumptions. Current approaches face significant challenges on both fronts, leading to critiques that they systematically underestimate the true scale of risk.
The IFOA's Warning: "The Emperor's New Climate Scenarios"
The Institute and Faculty of Actuaries (IFOA) argues that widely used climate scenarios are "deeply flawed" and "far too benign." They systematically exclude catastrophic risks like climate tipping points and cascading system failures, creating a dangerous disconnect between modeled risk and potential reality.
Key Sources of Model Uncertainty
| Source | Description | Impact |
|---|---|---|
| GCM Resolution Bias | Global models are too coarse to simulate localized extreme weather. | Understates frequency/severity of perils like thunderstorms. |
| Damage Functions | Functions linking hazard to damage are based on limited, often outdated data. | High uncertainty in loss estimates for a given event. |
| Tipping Point Exclusion | Models ignore abrupt, irreversible system changes (e.g., ice sheet collapse). | Grossly underestimates ultimate tail risk and potential for insolvency. |
| Exposure Data Inaccuracy | Portfolio data has errors in property value, location, and characteristics. | Directly causes mis-estimation of expected loss and capital needs. |
Section 5: The Regulatory Environment
The final premium is not a pure product of actuarial models; it is heavily mediated by a complex and politically sensitive system of state-level regulation. This creates market-warping effects, including a system of unseen cross-state subsidies.
The Regulatory Friction Factor
In "high-friction" states, regulation suppresses rates, decoupling price from risk. National insurers may compensate by raising rates in "low-friction" states, creating an invisible cross-state subsidy. Research from NBER.
Section 6: Charting a Course Through the Crisis
The industry faces the growing specter of uninsurability. The only viable long-term solution is to reduce physical risk through mitigation, innovation, and new forms of partnership.
Mitigation & Adaptation
The most effective path is reducing physical risk through stronger building codes and property-level hardening, incentivized by premium discounts.
Innovations in Risk Transfer
New products like parametric insurance can provide rapid, transparent liquidity after a disaster, bypassing slow claims adjustment.
Public-Private Partnerships
As mega-catastrophes exceed private market capacity, new risk-sharing layers are needed, with government as a potential reinsurer of last resort.
Concluding Theses for the Next Generation of Modelers
- Embrace non-stationarity: Develop models that don't depend on the assumption of a stable climate.
- Prioritize data engineering: A model's output is capped by its input quality. Master data sourcing, cleansing, and validation.
- Integrate system dynamics: Move beyond isolated perils to capture correlations, cascading failures, and feedback loops.
- Bridge the gap to business and policy: Translate complex analysis into clear, actionable insights for all stakeholders.
This article is based on a research project I have undertaken. After diving deep into the topic, I felt it was important to summarize my findings and analysis here.