Neural networks spot risk differences that ZIP codes miss

bilanol - stock.adobe.com

In every weather event, entire regions are described as "impacted" as if risk were evenly distributed across geography. However, on the same street, sometimes even next door, one property suffers significant loss while another holds up. 

This happens not because the surviving property was outside the storm path, not because it was spared exposure, but because something about that property made it behave differently.

If two properties experience the same event but produce very different outcomes, why would a model treat them the same simply because they share a ZIP code or territory? To understand why outcomes differ, insurers need to evaluate risk at the level where it occurs: the individual property.

In the first article in this series, we traced how insurance data evolved from manual review of raw reports to structured attributes and predictive models, improving consistency and scalability along the way. Attributes allowed insurers to standardize how they described risk, but attribute generation still relies on human decisions about which signals matter. In the second article, we introduced neural networks and embeddings as the next evolution in risk modeling. Rather than relying entirely on predefined attributes, neural networks learn patterns directly from data. During training, they create embeddings: compact numerical representations that capture relationships among observations.  These embeddings can reflect how risks relate to one another across time and space, helping models recognize complex patterns that are difficult to represent through predefined attributes alone. 

The real value of neural networks and embeddings becomes clear when applied to problems where risk is not easily simplified, such as property risk.

Traditional approaches to model property risk often rely on coarse segmentation based on ZIP codes, metropolitan statistical areas and census blocks. Models then layer on static attributes like roof age or construction type and occasionally incorporate isolated signals such as weather or claims history. While useful, these approaches simplify location into a set of predefined categories and variables. They assume risk is relatively uniform within those boundaries.

However, property risk is shaped by a combination of factors that extend far beyond coarse labels:

• The structure itself. How was it built, what materials were used, how has it been maintained and what losses has it experienced? 
• The surrounding environment. What is nearby, both physically and functionally?
• Weather exposure. What are the historical weather patterns? What are the emerging patterns? How is the weather changing?
• Local loss experience. What losses have already occurred in the immediate area? How is that loss history changing over time?

Traditional machine learning approaches can struggle to capture the nonlinear, interdependent and time-varying aspects of property risk. Neural networks are designed to learn these relationships directly from data, identifying properties that behave similarly, and those that differ in meaningful ways. The resulting embeddings act as a multidimensional coordinate system, positioning each property within this broader "risk space."
This approach to location intelligence applies to neural networks and risk embeddings to help assess risk at the property level rather than through broad geographic groupings.  It integrates claims history, weather exposure, property characteristics and surrounding context. It considers how these factors change over time to create a more detailed representation of risk, which is explainable from an actuarially sound and risk-based perspective. 

This approach can provide greater separation between highest- and lowest-risk properties than traditional modeling methods, helping insurers better distinguish variation within a territory. 

Greater visibility into property-level risk can influence underwriting, pricing, policy lifecycle and portfolio management.

• Underwriting improves because risks can be evaluated at a more granular level. Insurers can identify properties that carry hidden risk within otherwise attractive areas, while also recognizing lower-risk properties within challenging regions. 
• Pricing becomes more precise because it aligns with actual, property-level exposure. This helps reduce cross-subsidization, where lower-risk policies unintentionally offset higher-risk ones within the same territory.
• Across the policy lifecycle, insurers can make more informed decisions. They can detect when risk is changing, adjust strategies at renewal and intervene earlier when exposure begins to shift.
• Portfolio management can improve with greater stability and predictability. By more clearly distinguishing between risks, insurers can better manage concentration, reduce the likelihood of adverse selection and identify attractive new markets to grow smartly.

Summarizing this series, the first article described how insurance data evolved from raw reports to structured attributes and predictive models. The second explored how neural networks and embeddings learn richer representations of risk directly from data. In this article, we examined how those approaches can be applied to property risk modeling, helping insurers evaluate risk at a more granular level and better understand differences between seemingly similar properties. 
And that brings us back to where we started. Two neighboring properties may experience the same storm and produce very different outcomes. Insurers may not be able to control the weather, but they can improve how they understand risk. 

As more data becomes available and modeling techniques continue to evolve, risk assessment can increasingly be performed at the level where it occurs: the individual property. Neural networks and embeddings provide a way to capture the complex relationships that shape outcomes, helping move risk assessment beyond broad geographic averages toward a more detailed understanding of how risk behaves.