How ZestyAI Models Work: A Deep Dive into Property-Level Risk

At ZestyAI, we’re often asked:
What Does “Property-Level” Mean? 
What Makes ZestyAI Different from Traditional Risk Models?
Are ZestyAI Models Approved for Use in Underwriting and Rating?

This post answers the most common questions we receive from underwriters, actuaries, regulators, and technology partners—using wildfire, hail, wind, water, and storm perils as examples of how we turn complex data into actionable insights.

What Does “Property-Level” Mean?

Traditional risk models often rely on ZIP codes, territories, or broad regional averages to assess hazard and vulnerability. Stochastic models may support property-specific analysis, but they typically require external data sources, which adds cost, complexity, and inconsistency.

At ZestyAI, we assess each structure based on its physical characteristics and how it interacts with the surrounding environment.

We assess:

• Parcel boundaries and building footprints
• High-resolution aerial and oblique imagery
• Topography, slope, and vegetation
• Structural details like roof shape, materials, and defensible space

By integrating climatology with the built environment, we generate contextual risk scores that capture how each property’s physical characteristics, regional climatology, and historical loss experience interact to shape real-world risk.

That drives smarter decisions in underwriting, pricing, and mitigation, without relying on assumptions or manually sourced data.

What Makes ZestyAI Different from Traditional Risk Models?

Traditional risk tools often rely on:

• Broad hazard zones that can’t distinguish risk within a ZIP code
• Infrequent model updates that fail to reflect current conditions
• Over-simplified proxies—often relying solely on factors like year built—without accounting for deeper structural nuance
• Manual inspections that are slow, inconsistent, and costly

ZestyAI takes a fundamentally different approach.

Our models:

• Use gradient boosted machines that capture complex interactions between property features and environmental conditions
• Are trained on millions of actual insurance claims, not simulations, ensuring outputs reflect real-world loss experience
• Leverage both imagery (e.g., high-resolution aerial and oblique photos) and non-imagery sources (e.g., permits, climatology, topography)
• Continuously incorporate new data to reflect changing exposures
• Provide parcel-level risk scores with full transparency and regulator-ready documentation

With 97%+ U.S. property coverage, ZestyAI delivers national models with localized precision, helping carriers segment risk, price accurately, and respond to today’s evolving climate risks.

Are ZestyAI Models Approved for Use in Underwriting and Rating?

Yes. ZestyAI’s models, including Z-FIRE™, Z-HAIL™, Z-WIND™, Z-WATER™, and Z-STORM™, have been approved for use in underwriting and rating across the U.S.

Our regulatory approach is grounded in a few key principles:

• Transparency: We provide clear, regulator-ready documentation, including model methodology, variable selection rationale, and statistical validation.
• Collaboration: We work directly with carriers and state regulators throughout the filing process, from pre-submission briefings to objections.
• Responsible Innovation: Our models are trained on real-world claims, regularly updated with new data, and built with fairness and explainability in mind.

We support filings with:

• Detailed methodology and input documentation
• Variable importance rankings and validation studies
• Pre-built regulatory summaries to streamline the review
• Ongoing support throughout the regulatory lifecycle

ZestyAI has a track record of success navigating regulatory review. We help carriers adopt cutting-edge risk models with confidence and compliance.

Roof Age vs. Roof Condition: What’s the Difference?

ZestyAI’s models distinguish between roof age and roof condition, treating them as complementary signals that together provide a more accurate picture of roof-related risk, especially for hail and wind.

Roof Age is validated using a combination of building permit data and aerial imagery, analyzed through multiple proprietary methods simultaneously. We assign confidence scores to each roof age and apply minimum roof age rules to avoid false positives, ensuring the data is robust, even in jurisdictions with limited permitting records.

Roof Condition is assessed through computer vision models applied to high-resolution aerial imagery. These models detect visual signs of degradation, such as discoloration, wear, patching, and debris, that may not correlate with official replacement dates.

Why does this matter?

Because many insurers rely solely on reported roof age, which is often missing, outdated, or self-reported.

Our approach captures:

• Properties with older roofs that are still in good shape (and may be lower risk)
• Properties with newer roofs already showing signs of wear (and may be higher risk)
• Up-to-date roof vulnerability that static datasets can’t match

Together, roof age and condition power smarter decisions in underwriting, pricing, and mitigation—grounded in observable reality, not assumptions.

Modeling Approach: Why Gradient-Boosted Machines (GBMs)?

ZestyAI's risk models use gradient boosted machines (GBMs), a machine learning technique that delivers powerful predictive performance while remaining transparent and regulator-ready.

We use GBMs because they:

• Achieve high predictive accuracy by combining many simple decision trees into an ensemble that learns from its own errors over time, ideal for capturing complex insurance risk signals.
• Model non-linear interactions between variables, such as how roof complexity, condition, and regional climatology jointly influence risk, something traditional models or GLMs often miss.
• Enable transparency and explainability, with tools like feature importance rankings, partial dependence plots, and SHAP values that help underwriters, actuaries, and regulators understand what’s driving risk scores.
• Support a wide range of input types, from imagery-derived features to structured data like permits, topography, and property characteristics, all in one unified framework.

The result: a modeling approach that delivers real-world impact, supporting smarter underwriting, better pricing, and confident regulatory adoption.

How Z-FIRE Evaluates Wildfire Risk

Z-FIRE is ZestyAI’s structure-level wildfire risk model, built to capture both traditional wildland fire exposure and the growing threat of urban conflagration. Unlike traditional hazard maps that apply uniform risk zones across ZIP codes or counties, Z-FIRE delivers granular, property-specific risk scores for every structure in the U.S.

The model includes two levels of scoring:

1. Level One: Exposure Risk – Evaluates how likely a structure is to fall within a future wildfire perimeter, based on vegetation, slope, elevation, proximity to the wildland-urban interface (WUI), historical burn patterns, and regional climatology.
2. Level Two: Structure Vulnerability – Assesses how likely that structure is to be damaged if a wildfire occurs nearby. This score factors in structural characteristics like building materials, defensible space, and surrounding fuels, extracted from aerial imagery using computer vision.

Together, these scores provide a more complete view of wildfire risk: not just where fires may happen, but how individual structures are likely to perform.

Z-FIRE also captures non-traditional wildfire scenarios, including embers and wind-driven fires that jump the WUI and ignite dense suburban and urban neighborhoods. This makes the model particularly valuable for identifying concentration risk, urban conflagration, and managing PML across books of business.

Z-FIRE is validated on millions of insurance claims and performs reliably across all geographies—from the forests of California and the grasslands of Texas to emerging risk zones in Colorado, Oregon, and the Eastern U.S.

How Z-HAIL Accounts for Roof Vulnerability

Z-HAIL is a property-specific hail risk model designed to assess not just the likelihood of hail, but how damaging it will be to a specific structure.

Unlike traditional models that rely on historical hail frequency alone, Z-HAIL captures the interaction between local climatology and structural resilience by analyzing:

• Hail climatology: storm frequency, hailstone size, and intensity at a hyperlocal level
• Roof geometry and materials: pitch, complexity, covering type, and other features that influence how hail impacts a roof
• Property-specific vulnerability factors: including building height, exposure, and roof condition (derived from imagery and computer vision)

By modeling how hail behaves in a given location and how a specific roof is likely to perform under those conditions, Z-HAIL delivers precise risk segmentation at the parcel level. 

Carriers using Z-HAIL have seen significant improvements in underwriting performance. In an independent third-party review, Z-HAIL demonstrated a 20× lift in loss ratio segmentation between high- and low-risk properties—enabling more accurate pricing, better risk selection, and actionable mitigation strategies.

How Z-WIND Analyzes Wind-Driven Damage

Z-WIND is a property-specific model that evaluates vulnerability to both straight-line winds and tornadic activity by analyzing how wind climatology interacts with structure-level characteristics.

The model captures:

• Roof geometry: including shape, pitch, and surface area, which influence uplift forces
• Building elevation: to assess exposure to wind at various heights
• Local terrain and land cover: which impact wind speed, turbulence, and exposure to flying debris
• Historical wind climatology: including storm frequency and intensity
• Real-world claims data: to ensure outputs reflect actual loss performance

Z-WIND generates property-level frequency and severity scores, helping insurers move beyond broad wind zones to more precisely identify risk at the structure level. By understanding how specific buildings respond to local wind conditions, Z-WIND enables more accurate pricing, underwriting, and mitigation strategies across both inland and coastal regions.

How Z-WATER Tackles Non-Weather Water Losses

Z-WATER is an AI-powered model that predicts the frequency and severity of non-weather water and freeze claims at the property level, covering every structure in the contiguous U.S. 

While many traditional models depend heavily on basic indicators such as “year built,” Z-WATER combines those inputs with a broader set of property, climate, and infrastructure features to capture the interaction between three core dimensions of risk:

• Construction & Architecture: Property-specific features that influence vulnerability and claim severity, such as number of bathrooms, number of stories, presence of a pool, and overhanging vegetation (a signal for potential tree root intrusion).
• Climatology: Environmental stressors like temperature swings and freeze/thaw cycles that contribute to pipe bursts and system strain.
• Local Infrastructure & Hydrology: How local plumbing systems and electrical grids perform when real-world cold snaps or heat waves exceed what regional codes anticipated, exposing systemic weak points that lead to burst pipes and interior water damage.

These variables are derived from aerial imagery, tax assessment data, and regional climate and infrastructure datasets, all processed through ZestyAI’s proprietary AI framework.

Z-WATER helps insurers:

• Set fair and adequate rates based on true exposure
• Target high-risk homes for mitigation (e.g., water sensors or shutoff valves)
• Streamline operations by automating low-risk decisions and focusing resources where they matter most

What Is Z-STORM?

Z-STORM is a predictive, property-specific model built for carriers that rate hail and wind as a combined peril. It provides structure-level risk scores across the contiguous U.S., enabling more accurate pricing, underwriting, and mitigation decisions.

Unlike traditional territory-based approaches, Z-STORM models how storm climatology interacts with the built environment—capturing the real-world conditions that drive loss at the individual property level.

The model incorporates:

• Storm climatology: frequency and severity of wind and hail events at a hyperlocal scale
• Structural features: roof shape, material, pitch, and condition—key factors in a structure’s vulnerability
• Environmental context: including open terrain and nearby vegetation, which can amplify damage

Z-STORM predicts both:

1. Claim frequency: the likelihood a property will experience a storm-related claim
2. Claim severity: the expected loss as a percentage of Coverage A, providing a more precise view of financial exposure

This dual prediction enables carriers to:

• Accurately rate combined wind and hail risk at the property level
• Target mitigation strategies (e.g., roof improvements that reduce exposure to both hazards)
• Improve risk segmentation and pricing

Z-STORM offers a single, AI-powered solution for capturing the true complexity of convective storm risk—from climate data to construction detail to expected loss outcome.