A comparison of wildfire models underscores the need for the insurance industry to develop and utilize a more advanced wildfire model that accurately assesses wildfire risk, provides transparency for consumers, and creates the ability to determine actuarially sound rate levels, even for properties with extreme wildfire exposure.

Tammy Nichols Schwartz, CPCU

Founder and CEO
Black Swan Analytics LLC

Wildfire Insurance Market

The availability crisis in the California insurance market can happen anywhere catastrophe premiums do not match risk levels and the industry is lulled into a false sense of security by a few better-than-average years. Surplus built up over decades was wiped away in a matter of months. California property reinsurers increased premiums almost immediately and withdrew capacity, hardening the market like a new ice age as carriers struggled to stay afloat.

One small regional carrier, Merced Property and Casualty Co., collapsed under the weight of the losses from the 2018 Camp Fire in Paradise, and a new sense of urgency, maybe even panic, swept the industry. Primary carriers rushed to increase premiums and reduce capacity, issuing mass non-renewals to policies in zip codes with wildfire exposure, until precluded from doing so, temporarily, by statute.

The Western Fire Chiefs Association (WFCA) called for change at the 2019 WFCA Wildfire Summit, recognizing insurance policy non-renewals, cancellations, and complete withdrawals from the Wildland Urban Interface (WUI) as a market failure. The following statement was taken from the Action Plan Recommendation for WFCA Wildland Fire Policy Committee:

“There exists a disconnect between public policy WUI mitigations and insurance industry risk assessment. What continues to connect successful public policy and successful implementation of risk mitigations is public behavior. Until public policy and insurance practices in the WUI are aligned, the desired public behavior of mitigation implementation and maintenance will not be consistently realized.”


As the VP of Underwriting and Operations of the FAIR Plan at the time, I witnessed firsthand thousands of policies being referred to the FAIR Plan every month because they were in a “high risk” area. Many of these policies had little to no wildfire exposure and/or they had just completed mitigations ordered by the Fire Department, yet their carrier non-renewed them anyway. Carriers, in general, lack the tools needed to determine accurately who is at risk and who isn’t, and even if they had the tools, they don’t have the human resources to consider the mitigation actions policyholders took and the impact that may have had on the property’s risk level.

There is a need for consumer advocacy to help property owners and risk managers who are displaced by their insurance company because of their wildfire exposure. Individualized Structure Ignition Zone assessments, complete with applicable mitigation strategies and fuel maintenance plans, are necessary to assist consumers in establishing and maintaining insurability so they can continue to find coverage in the standard market. Understanding the models being used in the market today and offering consumers insight into the accuracy of the underlying information is possible with tools and data currently available.

Rather than a game of cat and mouse where carriers try to catch the best risks and avoid the worst with secret underwriting tools and black box rating algorithms, carriers should enlist the assistance of the homeowner as the front-line risk manager through the education and information that wildfire models can provide.

This paper examines the elements of wildfire models used by carriers today and calls attention to the differences between these models so that we can make better use of model results. This is important. Policyholders are not going to address wildfire risk that is identified by their carrier if they can just transfer their policy to another carrier who uses a different model. We lose credibility as an industry when we claim to have it right, and yet no two models agree. More important than our credibility is market stability. The property insurance market in California will not stabilize until carriers are comfortable with their ability to underwrite according to their risk tolerance and price according to expected loss.

Carriers can demand transparency for their policyholders, and they can supplement wildfire models that do not employ all the relevant criteria with inexpensive, hard data from insure-tech companies like HazardHub2 . They have this capability today. As relevant underlying data becomes readily available to consumers, and property owners arm themselves with information, carriers will improve in their ability to consider it.

If you are asking yourself, “What’s relevant?”, then this article is for you. It will also help you find answers to the following questions:


A wildfire pricing model is used to determine risk classes for properties with different propensities of loss that can be assigned, so that discrete rate levels can be established. It is not a stochastic model that derives loss probabilities of individual locations based on simulations of hypothetical wildfires or past events. However, a stochastic model is used to evaluate several pricing models in the Evaluation Methods section. The A.I.R. Touchstone 7.0 model is used to compare models by examining the expected Average Annual Loss (AAL) as a percentage of the Total Insured Value (TIV) by model output, or risk score.

The purpose of a pricing model is to create homogeneous classifications of risk so that premiums can most accurately be determined for one group relative to another. A group is considered homogeneous if it is made up of members which are all similar according to some characteristic.

When determining risk classifications for ratemaking, there are three important considerations:

For example, the type of paint used on the exterior is more relevant to building performance than the color of the paint. Since the color of a home does not impact expected losses, color should not be used to determine a risk category even though a group of red homes may be homogeneous.

Statistically, we know that empirical means will cluster about hypothetical means more densely when homogeneity is attained. Conversely, heterogeneous groups will likely see empirical results that are more dispersed about the hypothetical mean and predictions based on historical results will be less accurate.

Homogeneity accomplishes two important things:

The following section discusses the risk classifications used by available pricing models, the homogeneity of those risk classes, or lack thereof, and their influence over expected claims patterns.

Traditional Risk Factors

The use of the same variable by two competing models can produce entirely different results, so it is not just about what variables are used; it is also about how they are used. The following sections identify the risk factors used, and the differences between the classification methodologies employed by leading providers.

1 Fuel

Without fuel, a wildfire cannot exist. Since fuel directly influences the damage caused by wildfire, it is a relevant variable to include. While vegetation is a component of every wildfire model, the models use fuel in different ways. The goal is to identify vegetation classes that are expected to generate similar loss patterns, in keeping with the homogeneity requirement.

More sophisticated models use both type and amount of fuel, and they incorporate both the fuel at the insured location, on the residence premises, as well as the fuel in the surrounding area. See Table A below for a comparison of six wildfire models regarding their treatment of the fuel parameter.

In the final row, each model is credited with including the Fuel statistic, as indicated by a “Yes” for each of the providers. The cell is shaded light green if the parameter is utilized and bright green when the provider is using the most appropriate statistic in the most effective way.


2 Surface Composition

Surface Composition, included only in CoreLogic’s model, is based on the mentality that history is bound to repeat itself. According to CoreLogic,4 “Areas that have burned before carry a certain proclivity to burn again.” Density of regrowth after a wildfire is used to estimate the likelihood of the event recurring over time. Higher densities are correlated with higher frequencies. However, at a certain point, regrowth continues without perceivable changes in density.

3 Slope

Wildfire experts agree that slope is a relevant factor since wildfires move up hill faster than they move across level ground. The flame height is higher and setback requirements greater on a slope according to National Fire Protection Agency Standards. Therefore, slope meets the predictive requirement for inclusion. Only two providers omit slope from the model: HazardHub and Zesty.AI.

Policies can be grouped according to ranges of slopes with similar expected rates of spread, flame heights, and/or setback requirements. For example, homes with slope greater than 25% will tend to perform differently than homes with slope below 25%. Any number of discrete categories could be created, provided there is a distinguishable difference in the performance.

The group must be sufficiently large to provide enough statistical credibility for the empirical results to justify a distinct rate level. Hypothetically, carriers with different risk profiles could create separate categories of slope for a customized version of the model which maximizes the number of rating segments, the accuracy of the factors generated, and is unique to the distribution of that carrier’s portfolio. This creates the ability for carriers to differentiate themselves according to their individual wildfire risk appetite and create competitive advantages even if everyone uses the same variables in their models. This feature applies to any variable with a continuous distribution.

4 Access

Tragically, more than 80 lives were lost and many more injured, including firefighters, as residents attempted to evacuate using the only escape route. Burning vegetation along both sides, with fallen power lines and burning telephone poles spilling over onto the roadway, adversely affected access and evacuation efforts.

Dead-end roads and roads with smaller cul-de-sacs present great risk to firefighter safety, often preventing fire department equipment and personnel from reaching a threatened structure.

Verisk incorporates a component of access, identifying properties located on dead-end roads, or roads with cul-de-sacs that prevent firefighting equipment from navigating in and out of the community safely.

To take it a step further, neighborhoods with secondary road endings should also be identified. Homes in these communities are more likely to experience delayed response times and urban conflagration, further increasing risk of loss due to wildfire.

When the property is not located on a dead end or cul-de-sac, the number of access roads should be considered. The more options available for fire departments to reach a community and exit safely, the more able they are to defend it.

5 Aspect

Aspect is the direction which the slope is facing when the property is not on flat ground. Aspect affects vegetation according to the sun’s path. Hillsides facing the south are drier and warmer because they are in the sun longer; therefore, wildfire is more likely to ignite and spread more quickly.

A summary of traditional risk factors is shown below with the same color scheme as Table A.


Before contemplating the newer models and what they bring to the table, some additional background on the traditional models is considered.

Verisk and CoreLogic

As Verisk and CoreLogic have been in this space for many years, they have the majority of market share and remain the companies to beat for any new entrants to this space. Newer entrants have created additional variables to use; some with and some without the black box of secrecy.


Verisk derives two Fuel scores as part of Location®: one for the actual insured location and one for the surrounding area. Both Fuel scores are provided but only one is used in the FireLine® score calculation. Neither of these scores represent the type of fuel, but rather the amount present at the time of the latest evaluation date.

Verisk’s FireLine® product is based on Fuel, Slope, and a version of Access which evaluates the primary road terminus. The score is determined as the product of Fuel and Slope added to the Road Terminus score for a final score ranging from zero to thirty. There is an add-on for the identification of Special Hazard Interface Areas, in California only. It is a binary indicator of interfaces that are susceptible to wind-blown embers. FireLine® is only available in thirteen states.


CoreLogic’s wildfire ranking translates quantitative scores from five to one hundred, into qualitative risk categories from Low to Very High. Like Verisk, it is also based on landscape features such as Fuel and Slope. However, differentiating themselves, CoreLogic includes both type and amount of fuel at the insured location, noting that only the 30×30 grid containing the insured location is considered. As indicated in Section 1, more advanced models also incorporate the fuel of the surrounding area.

New Risk Factors

The next section discusses the latest additions to traditional wildfire pricing models and their importance in moving the industry forward in determining optimal wildfire risk classifications.

6 Distance to High Fuel Area

Frequently, wildfires start in high fuel areas. Therefore, distance to those areas is predictive. Like slope, categories can be defined using discrete ranges and adjusted according to carrier distributions if needed. All four of the new providers include Distance to High Fuel Area in their model.

7 Distance to Prior Burn Area

There are many places in the United States with high fuel areas that have never burned. Equally important as the distance to a high fuel area is the distance to prior burn perimeters. Statistically, the closer a property is to a prior burn perimeter, the more likely it is to be close to another one.

The map on the left shows burn perimeters of all major fires from 1998 to 2008, highlighting the national impact of wildfires. This is not just a west coast issue.

8 Number of Vegetation Burn Points

There are numerous occasions when vegetation is sparked, and fire ignites only to be extinguished shortly thereafter without becoming a wildfire. These could be intentional or fortuitous events caused by human error, motor vehicles, lightning strikes, or prescribed burns and debris burning. The more times vegetation ignites near a property, the greater the likelihood that a wildfire event may occur. Black Swan Analytics also captures lightning strikes that do not become vegetative burn points and has shown lightning frequency to be correlated with expected wildfire loss.

9 Historical Wildfire Frequency

Naturally, properties that have never burned and never been near a wildfire will be less likely to be near one in the future. Similarly, those that have been near a burn area once are lower risk than those near multiple burn areas or areas that have burned repeatedly. Consequently, in addition to the distance to prior wildfires, the number of recent wildfires nearby is an important consideration for predicting future wildfire behavior. Historical frequency should be evaluated quantitatively when possible with a history of at least twenty years. RedZone provides a narrative of fire history which is not incorporated into the model score. HazardHub and Black Swan Analytics are the only two providers that incorporate historical frequency in the determination of the location’s risk score

10 Proprietary “Fire Frequency”

An alternative to including actual historical wildfire frequency is to incorporate an expected wildfire frequency through some other proprietary estimate. While one cannot comment on the effectiveness of any proprietary components without a thorough understanding of them, it is feasible that the proprietary fire frequency is an acceptable alternative for the historical frequency. Conceding this to be case, and all else being equal, the ability to explain the model and its inner workings to the public is a critical component to driving changes in behavior in wildfire-prone areas.

Carriers and modelers should make every effort possible to avoid “black box” models that keep consumers in the dark about their risk level and the factors which contribute most to that exposure. For this reason, it is preferable to use an actual historical number with the necessary adjustments to reflect future risk in lieu of a proprietary “expected” wildfire frequency.

11 Fire Season Rainfall

HazardHub and Black Swan Analytics both include average rainfall during the wildfire season. Rainfall can be an important consideration as an indicator of fuel levels and moisture content. It can also be an important differentiator when considering two properties with the same type of fuel and the same number of years since the last wildfire. The one experiencing drought conditions may be more flammable while the one with more rainfall during the last wildfire season may have significantly more fuel. Black Swan’s model also incorporates the presence of drought conditions with the last 12 months.

12 Surface Fire Behavior Potential

Surface Fire Behavior potential is a function of Fire Regime Group6 (map on right) which determines mean return intervals and is not contingent on regrowth rates which can be difficult to discern. This information can be used to classify risks, in lieu of Surface Composition, according to the behavior potential of the region as observed over centuries.
In section 2, density of regrowth after a wildfire is used to estimate the likelihood of the event recurring over time since higher densities are correlated with higher frequencies. We referred to this as Surface Composition, and in Table B, CoreLogic is identified as the only provider using Surface Composition to capture differences in expected fuel behavior. Fire Regime Group and Surface Fire Behavior Potential are more stable choices based on historical fuel patterns and topographies in the area.

13 Housing Density

This may also be referred to as parcel density or structure density. Zesty.AI excludes housing density, which could be an important variable correlated with the urban conflagration risk that exists when housing density is high, though not technically considered wild fire.

14 Road Density

It is unclear how road density is material other than as an alternative to road access. Since fuel percentages and housing density are already utilized, road density is already inherently included in other models as the difference between unity and the sum of housing density and vegetation percentages.

If Road Density is intended to be an alternative to fire department road access, that would be an appropriate use although maybe not the most accurate method since roads can be cul-de-sacs or throughways, while road densities could be the same.

15 Risk of Loss due to Flying Embers

An approach by RedZone classifies expected wildfire loss in three zones:
These zones are defined as discrete areas that remain unchanged regardless of wind patterns and other seasonal effects. The potential for large rate differentials between homes in the same neighborhood or potentially on the same street must be monitored carefully. This approach precludes a continuous transition of rate as the risk dissipates from one zone to the next and could possibly become as problematic as territorial rating has been for insurers historically. Only three zones are used to identify wildfire risk levels across the entire country.

Apart from CoreLogic and RedZone, providers incorporate risk of loss due to flying embers as part of the overall expected loss due to wildfire. There is no differentiation between locations at risk of loss due to ground spread versus those caused by wind-blown embers. As seen in Santa Rosa in the 2017 Tubbs Fire, once an ember fall zone ignites and becomes a direct wildfire damage area, the previous evacuation zone becomes an ember fall zone and additional evacuation zones need to be identified.

16 Extreme Wind Areas

This statistic captures risk of loss caused by horizontal fire spread in addition to vertical spread, including expected loss caused by flying embers. These areas need their own mitigation strategies. Fuel must be treated and maintained differently than similar vegetation in non-wind areas. Extreme wind areas, therefore, are important to identify when calculating risk of loss due to wildfire.

17 Proximity to Fire Stations

The proximity to the nearest fire station seems an obvious choice to include in any wildfire model. However, if a carrier is already utilizing fire protection rating elsewhere in the Rate Order of Calculation, or ROC, then they have already accounted for this variable and it should not be included here as well.

Since wildfires require the response of more than one fire station, the model could contemplate the distance to multiple fire stations even if distance to nearest responding station is a separate rating variable. If the distance calculations for both are identical, the actuary’s job is complicated, but it can be done. Ideally, the two statistics would use different definitions of “distance” reducing the potential for a direct correlation.

Whatever the selection, attention must be paid to the definition of the term, “distance.” It is less relevant for the property to be close to the fire department as a crow flies, than as the truck drives. The single most relevant statistic for the “distance” calculation is the length of time it takes for fire department equipment to arrive.

18 Miscellaneous Building Features

Like fire station proximity and other variables used for rating, the construction features of the home should not be incorporated into the wildfire model if they are accounted for elsewhere in the carrier’s ROC. Most carriers already have different rates by construction class. However, neighborhoods like the one immediately to the right may warrant additional consideration.
If an entire neighborhood is comprised of the same fire resistive construction features, the risk of wildfire loss is diminished compared to neighborhoods with mixed construction types, a small percentage of which may be fire resistive.

Since wildfires often spread from home to home, including construction types common in the neighborhood can be appropriate, specifically since it cannot be used as a rating variable.

19 State-Specific Variances

California is an example of a state that has unique topography and wind patterns that exacerbate the wildfire issue. Since katabatic winds do not exist in every state, it is important that models have the flexibility to adjust to state by state differences. These special wind areas would impact the “ember fall zones” used in RedZone’s model. However, RedZone does not vary by state so the katabatic wind areas may not be included.

Other important differences by state exist from western U.S. to eastern zones with different climates, fuel types, and wildfire seasons.


Having examined each variable contemplated by leading wildfire models according to public information available at the time, the findings are summarized in Table C on the following page.

With each carrier having its own unique risk tolerance, premium goals, and distribution challenges, what is best for one carrier might not be best for another. Unfortunately, it is not as simple as selecting the column that is greenest.

Carriers looking to optimize the risk portfolio and write policies across the risk spectrum, will need a more robust model with many categories, or pricing points. That can be best accomplished by using as many relevant variables as possible, broken into as refined a segment as can be defined while still in keeping with the requirements for classifications: homogeneity, size, and correlation to loss

Other carriers using models for risk selection may be satisfied with a model that has only a few categories. This can also be sufficient provided the categories are homogeneous.


For an extreme example, consider a model that segments the business into two categories, “Good” and “Bad.” The carrier will want to test the model to ensure the groups are homogeneous, i.e., that there are no “good” policies in the “bad” group or vice versa. They will want to ensure the categories are sufficiently large, and that one category performs consistently better than the other.

Evaluation Methods

To evaluate homogeneity, review the final output of the model and examine each category for outliers. Groups with high standard deviation characterized by large disparities in performance may be poorly defined. For groups with many outliers or no seemingly typical property, consider splitting the group into multiple parts. If possible, comparing one model to another is the easiest way to identify outliers, provided the comparative model is reliable.

One way to evaluate a wildfire model’s effectiveness is to use the expected Average Annual Loss from wildfire (AAL) for each location from a stochastic model; compare that to the Total Insured Value (TIV) for that location and summarize the results by Score, the output from the model being tested.

Summarizing results for all policies by category, and sorting the results in increasing order, it becomes evident rather quickly if the model is predicting higher expected loss per policy for higher scores, as illustrated in the examples that follow.

Consider a case study of 21,000 policies grouped according to Verisk’s FireLine® score on the following page. The expected AAL according to A.I.R. Touchstone v7.0 is expressed as a percentage of Total Insured Value for each FireLine® Score and displayed graphically, as a blue line.

Graph 1

The blue line represents the expected AAL. The orange columns represent the first 90% of policies with the lowest risk scores 0, 1, or 2. Black represents the next 5% at FireLine® 3, and light grey represents the remaining 5% of policies with highest risk.
These results show that the Verisk model is stable and increasing for FireLine® scores 0-3. At 4 and higher, the results are erratic and there are significant reversals (circled in red) with higher FireLine® scores having less than half the expected loss of the FireLine® 4 group. For a detailed analysis of the FireLine® 0 category and an actuarially sound pricing strategy to split the category into two groups, see CDI Filing #18-2336.7 While not shown here, there is a significant lack of homogeneity in the FireLine® 0 category.

Since the goal is to price according to risk, a model which places 69% of business in one nonhomogeneous category is less effective than one with a more even distribution of scores, and a different price point for each one.

Graph 2

Using more risk factors on this same set of policies provides greater dispersion for the lowest 90% of policies while still providing generally increasing rate levels for about two dozen rating categories.
When evaluating the tail end of a model with large fluctuations, it can be helpful to examine the low points separately from the high points, as illustrated below. The high points are circled in green and low points in red. Connecting the high points with a green line and low points with a red line, the slopes can be compared. Similarly increasing or decreasing slopes are a favorable indication of the model’s ability to accurately predict wildfire risk for a given location. In these situations, fitting a curve to the results directly will provide the most effective rating factors, maximizing both accuracy and efficiency. Other models may need to employ additional grouping or capping.

Black Swan Analytics considers more factors known to be correlated to wildfire loss and, consequently, produces more rate segments, or price points, with higher expected loss for increasingly higher scores. Again, reversals can be seen in the tail end of the distribution, on the policies with the highest risk. See Graph 3 below. This is common.

Graph 3

Capping and grouping individual scores can smooth the output and hide reversals that occur because of low volume. When evaluating a model, it is important to look at the raw data before any grouping or capping is applied.

Each carrier has a different distribution of risks, underwriting rules, and rate structures so they need the flexibility to create their own groupings. If modelers provide Low, Moderate, and High categories or some other qualitative grade, ask if the underlying numbers are available. Once a model has been selected, grouping can be applied as appropriate for each carrier. It is not reasonable to assume every carrier will need the same groupings; therefore, more flexibility in this area is better.

When faced with the question of whether to use a model for underwriting or for rating, a good rule of thumb is to price for everything that can be priced and underwrite for everything else. When capping or grouping is required to smooth results, it may be an indicator that underwriting will be needed to control losses in the groups that have been capped. When the model is a good fit, rating factors based directly on the fitted curve provide the most accurate matching of risk to premium, reducing the need for additional underwriting.

A “Goodness of Fit” test is performed to determine how much of the variability in expected wildfire loss can be explained by the model. The statistic used for this test is the Coefficient of Determination, or R2 . An R2 of one means the model explains the expected loss perfectly and completely, while an R2 of zero indicates the model has no predictive value at all.

Graph 4, below, shows a fitted curve to model output for Black Swan Analytics. The R2 of 0.9437 indicates expected wildfire loss can be explained by the model with 94.37% accuracy.

Graph 4


A good wildfire pricing model will create many homogeneous risk classifications with enough statistical credibility to capture meaningful differences in performance from one classification to the next.

The best wildfire models go further, capturing all variables shown to be correlated with expected loss in order to provide the most accurate rates. If the model is transparent, insight into loss drivers can be used to reduce exposure and/or risk of loss through mitigation. These models also have the added benefit of being able to compare the impact of various mitigation strategies.

When choosing a wildfire model for pricing, a “Goodness of Fit” test is often performed. The most important statistic is the Coefficient of Determination, or R2 . As previously explained, R2 is used to measure how much variability in expected wildfire loss can be explained by the model. An R2 of one means the model explains the expected loss perfectly and completely, while an R2 of zero indicates the model has no correlation whatsoever to expected loss.

To maximize the R2 and get the most out of the pricing model, apply the following guidelines:

All told, nearly two dozen factors with a correlation to expected wildfire loss were discussed, yet the most popular wildfire models on the market use only four of them, and only one model effectively uses more than ten. Carriers agree they need more robust output than their current wildfire model can provide. Until now, they lacked the information needed to make the most informed decision.

With a clear understanding of the differences and similarities between models, carriers have the option of adopting the model that works best based on scientific evidence. Through renewed confidence that comes from better accuracy, a thorough understanding of their wildfire model, and access to critical supplemental data, the “black box” era can come to an end, and transparency can become reality for consumers.

Knowledge is power. Carriers are in the best position to put the power in the hands of the consumer where it can do the most good, creating a win-win for carriers and consumers. Carriers win with more accurate rates and insured properties with lower risk, while consumers win with knowledge and control of wildfire exposure, lower premiums and less risk of wildfire injury or damage.

Thank you for subscribing.
Don’t worry, we hate spam too.