How We Score Safety

Every score starts with public data from the Federal Motor Carrier Safety Administration. We pull seven distinct datasets, join them on DOT number, and filter to the population of active for-hire property carriers.

Not every FMCSA-registered entity is a for-hire trucking carrier. We apply hard exclusions that remove non-carriers entirely — but beyond those, nearly every carrier with power units and usable exposure is graded. Thin-data carriers are not dropped to N/A; their estimate is shrunk toward their size-band prior and tagged with a confidence tier. This is why the FRED Score covers about 1.1 million carriers, versus roughly 475k under the prior approach that excluded thin carriers outright.

Public FMCSA records — crashes, inspections, and violations — are aggregated for each carrier over its most recent observation window. But comparing a 3-truck fleet to a 500-truck fleet by raw event counts is inherently unfair. A larger fleet drives more miles and naturally encounters more events. Exposure is how the model accounts for that: it enters the prediction as a log-exposure offset, so the model predicts a rate of crash burden, not a raw count.

Exposure is the carrier's mileage over the scoring window, in 100k-mile units. It enters the model as log(E) — an offset that puts every metric on a per-100k-mile footing so small and large fleets are predicted on the same scale:

Even on a per-mile basis, a tiny fleet with 1 crash in 50k miles looks far worse than a large fleet with 10 crashes in 5 million miles — even though the small fleet's rate is mostly just noise. One lucky or unlucky year can swing it wildly. Rather than throwing thin carriers away, the model blends each carrier's own signal with its size-band prior using Bühlmann-Straub credibility.

Each carrier gets a credibility weight $Z = E/(E+\beta_{\text{band}})$ — the share of the estimate that comes from the carrier's own record, with the size-band prior filling the rest. Carriers with lots of exposure ($E \gg \beta$) are governed almost entirely by their own history; thin carriers ($E \ll \beta$) are pulled ("shrunk") toward the band prior.

That same credibility weight drives each carrier's confidence tier:

The FRED Score targets a carrier's severity-weighted crash burden — not just how many crashes, but how bad. Each crash is weighted by its outcome, so a fatal collision counts far more than a minor tow-away:

That severity-weighted burden is modeled as frequency × severity with a Poisson / Tweedie GLM and a log-exposure offset, so the model predicts the carrier's expected burden per mile driven over the forward 12 months. A gradient-boosted XGBoost-Tweedie model is run alongside as a cross-check.

The coefficients $\beta_k$ are fit from data, not hand-set. In order of predictive strength, the strongest predictors are:

Relative-strength illustration; exact coefficients are refit each cycle from the most recent year-over-year history.

A model can rank carriers well yet still systematically over- or under-predict for a given fleet size. To prevent that, each fleet-size band is calibrated so the total predicted burden matches the total observed burden — the observed-over-expected ratio lands at ≈ 1.0:

The forward expectation is then surfaced for underwriting as two figures per carrier — expected crashes and expected severity-weighted burden over the next 12 months:

Each carrier's predicted burden is first expressed as a risk relativity — its predicted burden divided by the level that is typical for its size band. 1.00× means typical for its size; below 1 is safer than peers, above 1 is riskier.

Grades are then assigned from the credibility-shrunk within-band burden-rate percentile — where the carrier sits among same-size peers. Because the comparison is within band, a small fleet is graded against other small fleets, never against the national giants.

The model is refit on the most recent complete year→year pair — coefficients are learned from one year of carrier history paired with the crash burden that actually followed. Every carrier is then scored on its latest 12 months of history to predict the forward 12 months. Because the model is never scored on the same data it learned from, the FRED Score is genuinely out-of-time.

Beyond the statistical model, a set of deterministic rules handle edge cases where the math alone isn't sufficient. These fall into three categories: score overrides, data-quality adjustments, and informational flags.

Every refit is validated out-of-time: the model is fit on one year, then judged on whether it ranks and calibrates the following year's crash burden on a carrier-disjoint holdout. Before a score update goes live it must clear a battery of discrimination and calibration gates — if any fails, the update is held.

All scoring, calibration, and out-of-time validation is performed by the reproducible pipeline fred_postgres_v4.py.