Behavioral Claim Velocity Attribution

What is this report? When a physical hazard (hail, wind) damages a neighborhood, the actual insurance claims filed often differ from what the damage alone would predict. This analysis measures that difference — called claim velocity — and traces it back to specific neighborhood characteristics.

Baseline = 1.0×. A neighborhood at 1.0× files exactly as many claims as the physical damage model predicts. A neighborhood at 1.5× files 50% more claims than the damage alone would suggest. A neighborhood at 0.7× files 30% fewer.

Velocity is clipped at 3.5× to prevent extreme outliers from distorting pricing. The model is calibrated on Texas homeowner wind and hail claims (2018–2023) from the TX Department of Insurance.

Scale of the model: 19 data-ingestion pipeline stages (Census, CFPB, HMDA, Zillow, court records, state insurance-department complaints, and more) feed roughly 95 distinct demographic/behavioral signal columns. Those signals are combined into 4 composite indexes (defined in the glossary below), which in turn drive a 2-part statistical model (frequency + severity) with 4 fitted coefficients.

Four neighborhood types (cohorts) are identified by unsupervised clustering on those composite indexes plus the ~90 other underlying signals: Stable · Transitional · Stressed · Assertive
TX TDI calibrated multipliers: Assertive 1.344× · Stressed 1.310× · Transitional 1.031× · Stable 0.865×

Glossary — what each term means

Dispute Culture Index — how likely residents are to formally push back on an outcome rather than accept it: CFPB complaints, state insurance-department complaints, court filings, and dispatch disputes. Higher = more likely to escalate a claim. Raises claim frequency (β=+0.836).

Claiming Culture Index — density of claim-adjacent businesses (public adjusters, roofing contractors, chiropractors, diagnostic imaging) net of stabilizing institutions (churches, credit unions, home-improvement stores). Its model effect is negative (β=−0.598) — a known multicollinearity offset against Dispute Culture and Attorney Density, not a standalone "claims go down" signal.

Attorney & Adjuster Density Index — concentration of law offices and public adjusters serving the neighborhood, a proxy for how likely a claim is to become represented/litigated. Raises claim frequency (β=+0.688).

Maintenance Index — how well-kept and stable the housing stock is: owner-occupancy, long-tenure residents, home-improvement loan activity, permit activity. The single strongest driver in the model — well-maintained neighborhoods file far fewer claims than damage alone predicts (β=−1.216).

Cohorts (Stable / Transitional / Stressed / Assertive) — four neighborhood types found by unsupervised clustering on the above indexes plus ~90 other demographic/behavioral signals. Stable = high maintenance, low dispute culture, lowest claim velocity. Assertive = the opposite — high dispute culture and attorney density, lower maintenance, highest claim velocity. Transitional and Stressed fall in between, in that order.

Forecast (Part 4, below) — separate from everything above. Forecast predicts whether a tract's cohort itself is likely to shift in the next 2-3 years (e.g. Transitional → Assertive), with a label ("Unlikely" → "Will"), a predicted destination cohort, a horizon year, and a confidence score. It is exploratory and not yet validated against outcomes — read it as a directional early-warning signal, not a calibrated probability.

The model uses four composite neighborhood indexes, each built from dozens of raw signals. This table shows how much each index raises or lowers claim velocity for each cohort — relative to the median-level neighborhood (which anchors at 1.0×). Hover over any signal name for a full explanation. Click any column header to sort.

Each composite index is built from raw neighborhood signals. This table traces velocity contributions down to individual signals via chain propagation: a signal's effect = its weight in the parent index × the parent index's model coefficient. These are approximate estimates — not directly calibrated.

Source: Dallas–Fort Worth (1,559 census tracts). Hover signal names for definitions.

Every census tract in the national dataset is run through the model to get a predicted velocity. This section shows how those velocities are distributed within each cohort — not just the average, but the full spread from the 5th to 95th percentile. The 90% confidence interval on the average is computed by resampling 1,000 times (bootstrap).

For each signal, all other signals are held at their median value while this one moves from low to extreme. This shows the standalone velocity curve for each signal — which one has the most pricing leverage, and where the non-linearity kicks in.

Pricing Leverage is Q95 velocity ÷ Q50 (median) velocity. A signal with leverage of 3.0× means an extreme-level neighborhood is priced at triple the velocity of an average one, based on this signal alone. The ⌈ marker indicates the model ceiling (3.5×) was hit — the true uncapped value is higher.

Protective signals (negative β) appear with leverage below 1.0× — meaning the Q95 tract has lower velocity than the median. Maintenance Index (β=−1.216) is the single strongest driver in the model, but its leverage ratio is inverted: a neighborhood at the 95th percentile for maintenance behaves far better than average, not worse. The leverage floor is just as valuable as the ceiling of risk signals.

Everything above measures a tract's current claim velocity. This section is different: it's the model's Step-18 output predicting whether a tract's behavioral cohort itself is likely to change in the next few years — e.g. a Transitional tract drifting toward Assertive as dispute activity rises and maintenance investment falls. This is a directional early-warning signal, not a claim-velocity number, and it has not been validated against real outcomes — treat the confidence scores as relative ranking, not calibrated probabilities.

Aggregated across 198 of the model's city/state configs (83,008 tracts — a handful of configs don't yet have forecast output).