Skip to main content
Wildfire Insurance → Investigation 04

Where Does the Model Fail?

Every model has a boundary. We ran ours on the Marshall Fire — a WUI conflagration that destroyed 1,084 structures on just 6,026 acres. The results are humbling.

100% Kincade Sensitivity
12% Marshall Sensitivity
6,000x Spread Rate Gap

Q1 through Q3 demonstrate that fidelity improves the insurance answer — for wildland fires. This investigation asks the harder question: where does the entire framework break down? The answer determines the scope of every Rothermel-based catastrophe model.

Validation — Three Fire Types

The Three Fires

We validated the model against three real fires that span the spectrum from pure wildland to pure WUI conflagration. Each fire tests a different failure mode.

Fire Type Properties Burned Sensitivity Specificity
Kincade (2019) Wildland 316 21 100% 6%
Camp Fire (2018) WUI + Wildland 155 131 99% 0%
Marshall (2021) WUI Conflagration 1,616 314 12% 68%

Sensitivity = burned properties correctly classified as high-risk. Specificity = unburned properties correctly classified as low-risk. L3 model used for all three.

Model Performance

Three Fires, Three Metrics

Model Performance Across Fire Types
Finding 1
The model catches every burned property at Kincade (wildland fire) and Camp Fire (WUI with wildland approach). At Marshall (pure WUI conflagration), it catches only 12% — missing 88% of the actual destruction.
Finding 2
Marshall’s 68% specificity is the highest of the three fires — the model correctly identifies unburned properties because Rothermel predicts low spread in suburban terrain. The model is right about what doesn’t burn. It is wrong about what does.
The Physics Gap

A 6,000x Spread Rate Gap

Rothermel models wildland fire: how fuel burns, how wind drives spread, how slope steepens the flame angle. It does not model how fire jumps from house to house through radiant heat, flying embers landing on decks, and gas line ruptures.

At Marshall, the fire spread at 30 km/h through suburban neighborhoods — 6,000 times faster than the model’s urban spread rate of 5 m/h. The houses were not burning because wildland fire reached them. They were burning because their neighbors were burning.

This is not a calibration problem. Structure-to-structure ignition is a fundamentally different physical process than wildland fire spread. No amount of parameter tuning closes a 6,000x gap. The model needs a different engine for dense WUI.

Rothermel’s equations describe fire propagation through continuous fuel beds — grass, brush, timber. In a suburban neighborhood, the “fuel” is discontinuous: houses separated by driveways, lawns, and concrete. Fire crosses those gaps through three mechanisms the model doesn’t represent: radiant heat flux from burning structures, lofted firebrands (embers) carried by convective columns, and infrastructure failures (gas lines, power lines) that create new ignition points.

Fire Type Peak Observed (km/h) Model Urban Rate (km/h) Gap Factor Cause
Kincade Wildland 15 0.005 3,000× Wind-driven grass/shrub spread
Camp Fire WUI + Wildland 27 0.005 5,400× Canyon-channeled wind + ember cast
Marshall WUI Conflagration 30 0.005 6,000× Structure-to-structure radiant heat

The model’s 5 m/h urban spread rate (from Rothermel fuel parameters) is calibrated for isolated wildland-urban interface structures. Marshall’s 30 km/h spread was driven by 100+ mph wind gusts and structure-to-structure ignition — a fundamentally different physical process.

Insurance Implication

What This Means for Cat Models

Rothermel-based catastrophe models are appropriate for exposure screening — will fire reach this neighborhood? They should not be used for structural loss estimation in dense WUI without a structure ignition overlay.

For Kincade-type events, the model works. Wildland fire approaching scattered rural properties is exactly what Rothermel was designed to predict. The fire reaches the property or it doesn’t. Terrain, fuel, and wind determine the outcome.

For Marshall-type events, it doesn’t. Once fire enters a dense neighborhood, the dominant spread mechanism switches from wildland fuels to structure-to-structure ignition. The model has no representation of this process and cannot be fixed by recalibration.

Finding
Both answers are useful — but only if you know which one you’re getting. An insurer using this model for a rural wildland portfolio is well-served. An insurer using it for a dense suburban WUI portfolio is missing the dominant loss mechanism. The model’s boundary is not a weakness to hide — it’s information that determines where the model should and should not be trusted.
Finding 3
The model’s failure at Marshall isn’t gradual — it’s categorical. Sensitivity drops from 100% (Kincade) to 12% (Marshall) with no intermediate behavior. This means an insurer cannot simply “adjust” a Rothermel model for WUI conflagration. The physics is fundamentally different. A separate structure ignition model is required.

The honest recommendation: use Rothermel-based models for wildland exposure and fire arrival probability. For dense WUI loss estimation, supplement with a structure ignition model (e.g., Pathfinder, NIST BFIRES) that explicitly represents radiant heat transfer and ember transport between structures. The fidelity gap is not in the parameters — it’s in the physics.

The Path to a WUI Model

The Marshall limitation is not a dead end — it's a defined engineering problem with public data available at each step. Here is what a full WUI-capable model requires:

Step What's Needed Data Source Status
1. Structure ignition overlay Parcel footprints, construction year, roof material, wall cladding County assessor data + CoreLogic; NIST BFIRES-2 or Pathfinder ignition model Data publicly available; IBHS provides structure vulnerability functions
2. Ember transport Higher-resolution wind fields (sub-km), structure-to-structure ignition probabilities IBHS research on ember accumulation; WRF wind downscaling over complex terrain Research-grade; no off-the-shelf implementation for rapid deployment
3. Validation data Structure-level destroy/damage/intact outcomes from real WUI fires Cal Fire DINS records — post-fire damage inspection, publicly available by incident Available now for recent fires (see below)

Available DINS Validation Fires

Cal Fire's Damage Inspection (DINS) program provides structure-level destroy/damage/intact records for each inspected parcel. Two recent fires with DINS data and dense WUI exposure:

  • Mosquito Fire (Sept 2022, El Dorado/Placer County) — 76,788 acres, significant residential exposure in Foresthill and Oxbow communities
  • Oak Fire (July 2022, Mariposa County) — 19,244 acres, structures in communities adjacent to Yosemite National Park

Both fires have DINS inspection records available through fire.ca.gov/incidents/. A structure ignition model validated against either fire closes the Marshall gap directly.

The limitation is scoped, not open-ended.

This model handles terrain-driven wildland fire accurately (validated against three fires). What it cannot do is model ember-driven ignition in dense residential areas — the Marshall scenario. The upgrade path above converts that gap from "unknown" to "three engineering steps with public data." The most expensive step (Step 2) requires research-grade wind modeling; Steps 1 and 3 are achievable with existing public datasets.