Disaster Metrics and Risk Assessment — An Overview of Athena’s Wildfire Model
No one wants large natural catastrophes to occur, but to insure risks, the insurance industry needs tools for assessing possible and probable losses from severe weather events. Athena Intelligence, when talking to insurers, is occasionally envious of history and clarity of hurricane risk models. The company is named for the Goddess of Wisdom, but it takes more than wisdom to explain the distinction between the conditions of the land (conditional risk) that fuel wildfires, and likelihood of structural damage and economic loss (probability risk).
For risk modelers working on Florida hurricane risk, there is a good understanding of the difference between frequency and severity. Any number of hurricanes might exist, but what matters to a risk modeler is how many achieve landfall (frequency) and how much damage they do (severity).
Unfortunately, no similar concepts exist for wildfires — yet!
Most underwriters want a Yes/No decision for their underwriting of wildfire risk. You can learn more about using Athena for underwriting decisions here where we discuss the ease of implementation.
The rest of this article is for those that want to look under the hood at how Athena gains the insights from Voice of the Acre®. All the screen prints will be from Athena’s founder, David Sypnieski’s public Tableau account, Historic Profile Targeting.
All the information in this article is based an example, Clark County in Washington State. This is an area that historically has had moderate risk. Athena’s analysis is based on the entire state that have occurred within the last 6 years.
Why Six Years?
Data quality. Athena Intelligence is the first conditional, geospatial, AI/ML risk probability model for wildfires. Most of the data used is public — primarily from government agencies, academic research into wildfire behavior, property damage reports, etc. With any software analysis, the quality of the data is key. So, while we have stories of wildfires over the past 100 years, the data needed by the machine learning must be clean and usable. That only exists for the past 6 years.
The raw information has been sourced from dozens of databases, including records of other fires that have occurred in the area and property damage reports. The data is disaggregated and unstructured, ranging from PDF files from CalFire damage reports to weather and vegetation that arrive in 100 square feet pixels. These are blended, indexed and used to create a statistical assessment of both the wildfire potential and the potential property damage.
At the end of the day, the resulting analytics are only as good as the initial data. (Garbage-in, Garbage-out, as the saying goes.) Our goal is analysis good enough to bre utilized in driving operational decision making. While there is A LOT of environmental and wildfire related data available, in its originating forms it is virtually impossible to be combined and used quickly to drive real business decisions. Athena’s wildfire prediction algorithms are based on geospatial artificial intelligence.
Why is Geospatial important?
The conditions that drive a wildfire’s behavior and ultimate destructive potential are dynamic and fundamentally geospatial. The originating data primarily comes in various geospatial formats. Geospatial formatted and structured data is different from traditional non-geospatial data as it is indexed to a specific location on the Earth’s surface. Geospatial data allows for spatial relationships to be utilized in Athena’s models, analytics, and predictions.
Athena has years of experience in geospatial data processing for use in industrial scale commercial decision making.
The Voice of the Acre® (VoA) aggregates all terrain data. Because many businesses are operationally tied to the ground, the long-term goal is to give the land a voice — to make it as easy to query the land as a Google search. While VOA will, in the future, be optimized to answer many questions, the current optimization module is wildfire risk.
At a very high level, these are the factors that go into Athena’s wildfire prediction models.
Conditional Risk and Probability Risk
Below is the first spreadsheet in our discussion of the multiple perspectives needed to understand wildland fire risk. This shows Conditional and Probability Risk for Washington State, as well as all 6 years of wildfire history.
Conditional Risk refers to the underlying risk of wildfire on the land. It reflects the fuel load (now and projected, like a crop harvest) and the expected behavior of the fire.
Conditional Risk, shown here on the Y-Axis, at the highest level is a Green, Yellow, Red.
Wildfires come in multiple flavors — flame length, temperature and intensity, for example, and the software uses multiple axis for factors, which roll up into Green/Yellow/Red. The reality of Conditional Risk is more nuanced, which is described as an Aegis Score which ranges from Very Low to Very High.
Similarly, Probability Risk, shown here on the X-axis, is similar to but different from the Conditional Risk. Probability Risk reflects the potential property damage if a wildfire occurs nearby. The same factors are included in the analysis, but the indexing, scoring and optimization is tilted slightly differently.
At a high level, Probability Risk is described by words, from Low to Very High. The more nuanced version of Probability Risk is described by colors. These go from Green to Yellow to Orange to Red to Purple.
It looks like the highest “Risk of Destruction” occurs where both the Conditional Risk and the Probability Risk are high.
Athena’s Voice of the Acre® for Wildfire is of interest to decision makers in property insurance and electrical utilities, as well as people writing Community Wildfire Prevention Plans (CWPP) because of its predictive capabilities. The characteristics of the land and structures in wildfires are then applied to current and future conditions, and these are changing with increased weather volatility.
Climate change is changing solar degree days, plant growth and evapotranspiration. These agricultural concepts can be modeled in the context of geospatial data, something that the Athena developers have more than a decade of experience doing.
We don’t know where wildfires will strike, but they only strike under specific circumstances.
At the beginning of the year, we make predictions about risk. We can, with high confidence,
At the end of the year, you will be able to look back and 70% or more of all the wildfires will have occurred in areas that Athena categorized as High Risk (Red on Conditional Risk, and Very High Probability Risk).
But what are the numbers in the boxes?
This is where people get confused. We have described both an X and Y axis with words and colors, expressed as words. Athena is a synthetic data vendor, so for cat modelers pulling our data into their proprietary analysis, those words are numbers. But for the purpose of explaining Athena to people, we use words.
But here is our promise, At the end of the year, you will look back … The numbers shown in the boxes reflect all of the fires that have occurred in the data we used.
The red box on the upper right corner is the highest risk, with a Conditional Risk score of Red/Very High, and a Probability Risk score of Very High/Purple. Over the 6 years, 40.03% of all wildfires occurred in land with these characteristics. 21.97% of all wildfires occurred in land only slightly less risky — with a Conditional Risk of Red/High and a Probability Risk of Very High/Red.
At a less nuanced level, 62% of all wildfires in the State of Washington, over the 6-year period occurred in areas with a Conditional Risk score of Red and a Probability Risk score of Very High.
At the end of the day, wildfires have a random component. Land with a Probability Risk score of Very High/Orange has higher risk than locations with a Probability Risk Score of Low/Yellow. But in this time period, few fires occurred in the Very High/Orange (0.09% of the area). A larger (but still small number, less than 1%) of area burnt over this small number of years was categorized based on conditions at the beginning of the year, as Low/Yellow.
Over longer time periods, with more experience, the randomness will fade, and the statistical analysis will become more accurate. But even then, each new year will have some randomness.
This is a good place to stop. More information will be forthcoming.
Athena is a next generation InsurTech data vendor which produces synthetic data. The earth’s essential data is refined to make it easy for enterprises to use environmental information for future contracts, proprietary business decisions and risk management.