Skip to main content
Wildfire Insurance → Investigation 01

When Does the Risk Tier Change?

An insurer classifies each WUI property as LOW, MEDIUM, HIGH, or EXTREME wildfire risk. The classification determines the premium. We ran 316 properties through four model fidelities — from simple zone lookup to 200-draw Monte Carlo.

316
Properties
4
Fidelity Levels
200
MC Draws
91%
Reclassified

This is the foundation of the study. The property-level classifications produced here feed directly into Q2’s portfolio loss estimation, Q3’s correlated season risk, and Q4’s honest boundary analysis. Every downstream finding depends on what this simulation produces.

The Problem

Zone-Based Classification Doesn't Differentiate

Most insurers classify wildfire risk using zone-based systems like CAL FIRE's Fire Hazard Severity Zones (FHSZ). Every property in the WUI gets the same classification — typically HIGH — regardless of terrain, vegetation, wind exposure, or defensible space. The premium is set by the zone, not the property.

A more detailed fire behavior model could differentiate risk within the zone. But does it change the classification enough to matter? If 95% of properties stay in the same tier, the added complexity isn't worth the cost. If most properties move, insurers are mispricing risk at scale.

We tested this by running 316 WUI properties through four progressively detailed models and tracking how many properties changed risk tiers at each step.

Fidelity Ladder

Four Levels of Classification

L0 — Zone Lookup

Distance to Wildland

Binary: inside the WUI or not. Uses CAL FIRE FHSZ boundaries. Every property near wildland vegetation gets the same classification.

Result: 96% classified HIGH. No differentiation.
L1 — Screening Model

Terrain + Fuel + Wind

Feature-based scoring using slope, fuel type, wind speed, and aspect. Adds an EXTREME tier for properties with compounding risk factors.

Result: 76% EXTREME. 80% reclassified from L0.
L2 — Deterministic Simulation

Fire Behavior Model

Physics-based fire spread model (Rothermel + wind vectors). Catches directional effects: properties upwind of fuel loads are safer than those downwind.

Result: 31% reclassified from L1.
L3 — Monte Carlo

200-Draw Uncertainty

Samples over wind direction, speed, fuel moisture, and ignition location. Classification based on the probability distribution of outcomes, not a single run.

Result: 13% reclassified from L2. Reveals uncertainty.

Each level builds on the previous. L1 adds features to L0's zone. L2 adds physics to L1's features. L3 adds uncertainty to L2's deterministic answer.

Risk Distribution

How Classification Shifts Across Fidelity

Risk Tier Distribution by Fidelity Level (316 Properties)
Finding
Zone-based classification treats 96% of properties the same. Monte Carlo reveals a 4-tier distribution where 86% are EXTREME, 9% HIGH, 4% MEDIUM, and 2% LOW. The 91% total reclassification rate means premium assignments change for 9 out of 10 properties.

The biggest shift happens between L0 and L1: the screening model alone reclassifies 80% of properties by adding terrain and fuel features. The deterministic simulation (L2) reclassifies another 31% by catching directional wind effects. The Monte Carlo (L3) reclassifies 13% more — smaller, but these are the properties where uncertainty in wind and moisture conditions changes the answer.

Reclassification = property changes risk tier relative to the previous fidelity level. Total 91% is L0 to L3.

Transition from L0 to L3 (316 Properties)
Decision Impact

What If You Stop Earlier?

If You Stop At Properties Differentiated EAL Estimate Time to Compute What You Miss
L0 (zone) 0% (all HIGH) $54M <1 second Everything — no differentiation
L1 (screening) 80% reclassified $118M Seconds Directional fire effects, weather uncertainty
L2 (deterministic) 91% reclassified $127M ~10 seconds Tail risk from weather variability
L3 (Monte Carlo) 91% reclassified $127M ~12 minutes Nothing — this is the full answer
Finding 4
The marginal value of simulation drops sharply after L1. The screening model (seconds to compute) captures 80% of the reclassification that Monte Carlo (12 minutes) provides. But L3 is essential for one thing L1 cannot do: produce the burn probability distribution that Q2 needs for PML estimation.
Validation

How Well Does Each Level Detect Actual Fire?

We validated each fidelity level against NIFC fire perimeter data for the Kincade Fire. Sensitivity measures how many properties that actually burned were classified HIGH or EXTREME. Specificity measures how many properties that didn't burn were correctly classified LOW or MEDIUM.

Level Sensitivity Specificity Interpretation
L0 — Zone 1.00 0.04 Catches everything, flags everything
L1 — Screening 1.00 0.04 Same detection, same over-classification
L2 — Deterministic 1.00 0.07 Slightly better specificity
L3 — Monte Carlo 1.00 0.06 Perfect detection, conservative

All four levels achieve perfect sensitivity — no property that actually burned was classified LOW. But specificity is uniformly poor (4–7%), meaning the models over-classify nearly every property as high risk. This is expected in the WUI: these properties are all near wildland fuels. The value of higher fidelity isn't in catching fires the zone model misses — it's in differentiating risk within the high-risk zone.

Finding 2
The jump from L0 to L1 captures 80% of the total reclassification — and L1 requires zero simulation. A screening model based on terrain features, wind exposure, and defensive space captures most of the differentiation that a full Monte Carlo provides. For insurers with limited computational budget, L1 is the highest-value upgrade.
The Honest Caveat

Everything Is High Risk in the WUI

At Kincade, 96% of properties are classified HIGH under zone-based pricing. Under Monte Carlo, the distribution shifts — but it shifts from HIGH to EXTREME, not from LOW to HIGH. In a WUI area, almost nothing is safe. The model’s value isn’t finding safe properties — it’s differentiating between “high risk” and “catastrophic risk.” That distinction is where the premium differential lives: the difference between EXTREME (86% of properties, 5% annual loss rate) and HIGH (9%, 2% loss rate) is a 2.5× premium multiplier. For an insurer, that’s the decision that matters.

This finding generalizes: in any WUI area, zone-based classification will always lump most properties together. The value of higher-fidelity modeling is within-zone differentiation, not safe-vs-unsafe classification.

Finding 3
The 2.5× premium multiplier between HIGH (9% of properties, 2% loss rate) and EXTREME (86%, 5% loss rate) drives $73M in annual premium differentiation across this 316-property neighborhood. Zone-based pricing leaves this $73M invisible — every property pays the same.
Cross-Fire Comparison

Reclassification Across Different Fires

L0 to L3 Reclassification Rate by Fire

The reclassification effect is not specific to Kincade. Camp Fire — a faster, more destructive fire in different terrain — shows an even higher reclassification rate of 88%. The more extreme the fire environment, the more the zone-based model fails to differentiate.

Camp Fire

When Everything Burns

Camp Fire destroyed Paradise, California in November 2018. With 85% of properties inside the fire perimeter, almost every property burned. The zone-based model says they're all HIGH risk — and it's not wrong. But even when nearly everything burns, the Monte Carlo still differentiates.

87% classified EXTREME, 12% HIGH, 1% LOW. Even in the worst-case fire, properties are not equally at risk. The ones classified HIGH instead of EXTREME had slightly better terrain position, more defensible space, or less exposure to the dominant wind direction. That 12% matters for pricing — they should pay less than the 87%.

Camp Fire's 88% reclassification rate confirms the pattern: zone-based classification consistently misprices WUI properties. The properties at the extreme end of the risk spectrum subsidize those with moderate exposure, and everyone pays the same premium.