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 naturally encounters more events. Exposure is how the model accounts for that: crash burden is measured per power-unit-year and each violation signal per inspection, so a carrier is compared on a rate, not a raw count. We use power-unit-years rather than self-reported mileage on purpose — MCS-150 mileage is frequently blank, stale, or implausible, so it is never the exposure base for the grade.

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 $Z = n/(n+K_{\text{band}})$, where $n$ is the carrier's accumulated evidence — the count of real safety observations (roadside inspections and observed crashes / out-of-service results, plus a small fleet-size-and-tenure floor). Credibility comes from how much of a carrier we have actually seen, not from the mileage it self-reports: a one-truck fleet that merely claims 100k miles has almost no evidence and is shrunk to its size-band prior, while a fleet with a heavy inspection record (or real adverse events) is governed mostly by its own history. Carriers with lots of evidence ($n \gg K$) ride their own record; thin-record carriers ($n \ll K$) are pulled ("shrunk") toward the band prior — so a strong grade has to be earned with observations, never granted for the mere absence of data. Because adverse events add to the evidence count, a genuinely bad record still has the credibility to grade a carrier down.

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:

Fatality and injury terms count casualties (capped) and stack — a two-fatality crash weighs more than a one-fatality crash. The carrier's own severity-weighted crash burden, and its vehicle out-of-service, severe-violation, hours-of-service / fatigue, unsafe-driving, and speeding rates, are each turned into a cohort-relative Empirical-Bayes relativity (1.0× = typical for its size cohort), shrunk toward the cohort average by how much real exposure backs that signal — crash burden by power-unit-years, violation rates by inspection count. These combine into one transparent, credibility-weighted risk relativity; a carrier's observed crashes set a floor on that relativity, so a real crash record can never be averaged away.

The grade comes from a transparent geometric blend of the per-signal relativities $r_{s}$ (no black box) — each raised to a weight $w_s$ learned out-of-time from how strongly that signal predicts the following year's crash burden. In order of learned weight, the signals are:

Weights are log-relativity exponents learned out-of-time by a non-negative Poisson fit; refit each cycle. Signals that don't predict forward crashes (e.g. generic equipment / maintenance write-ups) are dropped rather than given face-value weight.

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 its total predicted crashes (and burden) match that band's own recent observed totals — 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:

A parallel Poisson model is fit the same way on fatal crashes (crashes involving a fatality) to estimate each carrier’s forward fatal-crash rate, surfaced for underwriting as the probability of at least one fatal crash in the next 12 months. Fatal crashes are rare, so this estimate leans heavily on exposure, fleet-size band, and behavioral / speeding signals, and is calibrated band-by-band like the burden model:

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 this credibility-weighted risk relativity on fixed thresholds calibrated to realized forward crash burden — below 0.52× the carrier's size cohort earns “Excellent,” 0.52–0.60× “Strong,” 0.60–0.90× “Satisfactory,” 0.90–1.20× (at/just above the cohort average) “Marginal,” 1.20–1.60× “Poor,” and above 1.60× “Critical.” The thresholds sit on the actual forward-burden gradient, so each grade realizes materially higher next-year crash burden than the one above it. Because the relativity is measured within the carrier's size cohort — the in-scope book is banded 10–19, 20–49, 50–99, 100–199, 200–499, 500+ power units, with 1, 2–5 and 6–9 held as advisory — a fleet is judged against genuinely comparable peers, never against carriers many times its size. The bands are set finely wherever the population supports a stable per-cohort estimate. A single-power-unit carrier is capped at “Strong” — one power-unit-year of crash exposure is too thin to credibly earn the top grade (see grading owner-operators). A carrier with no inspections and no crashes on record has no observed safety record to rank, so it is marked “Not Rated” rather than assigned a fabricated grade.

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 (every fifth DOT number is held out of fitting). One hard gate blocks publication — grade monotonicity (below) — and the refit aborts before any score is written if it fails. Discrimination and calibration metrics are recorded for every refit and reviewed; the figures below are from the latest validated run (outcome year ending 2026‑03‑08).

Top-decile lift, per-cohort Gini, and a grade-monotonicity gate — realized forward burden must rise from Excellent to Critical, or the run aborts rather than ship a contradictory score — are computed and logged to a per-run metrics artifact for review. All scoring, calibration, and out-of-time validation is performed by the reproducible pipeline fred_score_v6.py.