The Structural Reason that Pet Insurance Loss Ratios Run Higher Than Other Lines

Author: Andrew Hellman (Stride Systems)

Pet insurance loss ratios run roughly 14 points above personal lines. Approximately half reflects veterinary cost inflation; the remainder reflects an addressable behavioral data gap. This note explains the structural reason that gap exists: risk priced on the animal alone, observed only at inception, when the risk belongs to the dog-owner pair and changes across the term.

At time of writing, pet lines loss ratios run roughly fourteen points above the personal lines average and have done so persistently for years. The standard explanation is veterinary cost inflation, which is real and largely outside the insurance industry's control. That accounts for part of the loss ratio gap. It does not account for the part that persists after cost trends are stripped out, and it does not explain why a line this mature still prices with so little visibility into what actually drives a claim.

The more useful explanation is structural, and it has nothing to do with how much animals cost to treat. It has to do with what a pet policy actually insures.

A pet policy underwrites an animal. It insures a pair.

A claim is rarely the product of the animal alone. It is the product of the animal interacting with a particular owner, a particular routine, a particular level of attentiveness and care-seeking and consistency. The risk belongs to the pair. The insurance industry prices one half of the pair and accepts the risk of the whole.

Pet lines is overwhelmingly a dog business, and that is where the structural problem is both largest and most solvable. Two reasons converge on the same conclusion. The first is scale: dogs account for roughly 80% of pet lines premiums written, so addressing the dog book addresses the overwhelming majority of the problem, and whatever noise remains in the rest can optimize slowly behind it, or never, without changing the result. The second is observability: a dog's risk is driven by its relationship with the household, and a relationship expresses itself as behavior that recurs and can be observed over time. The dog-human bond is not the complication here. It is the enabling factor, because it is the part of the risk that produces a continuous, observable signal. Solve for the dog and you have substantially solved pet lines.

That observable signal is precisely what underwriting does not capture. This is the behavioral data gap: the behavior that produces the claim is the one thing the underwriter never sees.

The behavior goes unseen in two distinct ways. Underwriting a dog begins with declared attributes captured once, when the policy is bound: breed, age, location, sometimes a short medical history. The model therefore sees one moment where the risk runs the length of the term, and it sees one body where the risk has two. The pricing approach is static where the risk is dynamic, and single-bodied where the risk is dyadic. Two separate blind spots, each closing off a different view of the same behavior.

Take the static blind spot first. A dog acquired as a settled, well-exercised member of a stable household is a different risk eighteen months later if the household has destabilized, the routine has collapsed, and care has turned reactive. Even incremental change, well short of any acute event, can produce a meaningfully different risk. None of it is currently observed. The policy was priced on a snapshot and renewed on net present rate terms against the same snapshot, while the actual risk moved underneath it. Every other personal line learned this lesson on its own timeline and long ago. The question each line eventually faced was the same: the risk we price changes during the term, so why are our inputs frozen at inception.

The single-bodied blind spot is the one the insurance industry has no existing tool for. Pet lines insures a living system with two participants. The dog supplies instinctive drive and physical predisposition. The household supplies context: how consistent the routines are, how early care is sought, how well the environment fits the animal, how stable daily life is. Outcomes are produced by the interaction of those two determinant factors, and a model that reads only the animal captures one input of a two-input system. The error this produces is not random noise that better animal data would reduce. It is a structural omission that no amount of animal-side data can close, because the missing variable sits on the other side of the pair.

Which is why the obvious technical fix is the wrong one. More sensors optimize a single body when the problem lives in the pair.

The instinct is to instrument the subject: put a device on the dog, capture activity continuously, and let volume stand in for the missing signal. Auto telematics is the usual precedent, and it is read backward. A connected vehicle generates nearly twenty-five gigabytes an hour across more than a hundred data points, yet the inputs that actually move a price are a short list: speeding, harsh braking, distracted driving, route patterns. Every item describes the human at the wheel. The data that underwrites is human-intermediated, because it confirms instrumentation within context rather than measurement on its own. The car generates the volume. The driver is the signal.

A collar instruments the dog independently, and the appeal of that independence is real: no owner has to stand there counting, and automatic capture looks more accurate than a self-report. The independence is also the defect. Instrumenting one body in isolation strips out the context that gives the reading meaning, so the effort optimizes a way to widen the gap while presenting itself as the way to close it. Passive sensing captures motion. It does not capture context, and context is where the underwriting signal lives. The result is volume from the reachable body and silence on the determining one, which is data quantity aimed at the wrong half of the unit.

The point holds across every mature telematics line, all of which are, on inspection, already human-intermediated. The auto sensor measures a person driving. The usage signal in commercial lines measures people operating equipment. The value was always the human behavior the device happened to capture. Pet is the one line where the dominant sensing approach points at the non-human participant and calls the output behavioral data. To price a dog's risk without the household, one would have to argue that a dog's outcomes are largely unrelated to how it is kept. Nobody who has owned a dog believes that.

So the gap is neither a data-volume problem nor a compute problem. The insurance industry's analytics investment is substantial and growing. Compute ran ahead of signal, which is a comfortable position only if the missing piece is processing power. The missing piece is the input itself: a view of the pair, captured as it changes, rather than the animal captured once or instrumented as a single body.

That reframing is what turns a complaint into an opportunity. The stubborn part of the pet lines loss ratio gap is not evidence that the risk is inherently wild. It is evidence that the risk is half-observed, and a risk that is half-observed can be made whole. The unobserved half is behavioral, it is dyadic, and a gap that traces to a missing input is a gap that closes when the input arrives. That has happened in every adjacent line, recently enough that the pattern is not in doubt. The work now underway at Stride Systems is the building of that input for pet lines. The full analysis, including how the addressable half of the gap is sized against the global dog book, is set out in the white paper Pet Insurance Loss Ratios: Addressing the Behavioral Data Gap, with an interactive Simulation that applies the recovery logic to a participating book alongside it.

- -

Stride Systems builds longitudinal behavioral data infrastructure for pet lines underwriting.
For technical data memo materials, contact andrew@stridesystems.io.