How We Score Safety

Every score starts with public data from the Federal Motor Carrier Safety Administration, downloaded weekly and loaded with header validation and per-file manifests. Every crash and inspection is placed by the date it happened — that date sets the windows and the decay — and each record also carries the date FMCSA added it, so a grade can always be rebuilt as it would have been seen on any past day.

The universe is every property motor carrier — for-hire, private and exempt; interstate and intrastate — with at least one power unit. Operation type is a feature of the class rate, not a reason to exclude. Passenger operations (buses, coaches, school buses, vans, limousines) are a separate stratum with their own class models still to be built, so they are shown without a grade. A motor carrier under an FMCSA out-of-service order at the cutoff is shown without a grade, as is one whose operating authority has been revoked with no roadside inspection or crash recorded since and no active authority on file (a motor carrier seen operating after the revocation, or reinstated, is graded). Motor carriers registered after the cutoff, and filings of more than 5,000 power units, are also shown without a grade; each carries its reason on the page.

Every motor carrier belongs to one of nine fleet-size cohorts. The cohort sets the credibility prior for the experience term and the reference rates for the violation relativities. Six cohorts of ten or more power units form the ranked book; the three smaller cohorts are graded but advisory, because a one-truck record proves little either way and the data cannot finely separate those motor carriers.

The credibility prior a is explained under Factor 3; the column shows its range from the thinnest to the best-observed records in each cohort. The ranked book holds 107,352 motor carriers in the latest weekly run. Every filing of 1,000 or more power units is checked against its own drivers and the roadside record before grading. A filing they contradict (for example 2,000 power units, two drivers and no truck inspected in two years) is not graded: its filed fleet is not credible. Motor carriers filing more than 5,000 power units are not graded yet. Where the roadside confirms the size, these fleets crash about 1.4 times the rate of a 5,000-unit fleet and their 12-month crash forecast holds, but their grade moved between bands too often from month to month to publish; the test is repeated in mid-2027. Truck leasing and rental fleets, and filings the roadside does not confirm, are shown with their own reasons.

Crash risk is measured per power-unit-year. The filed power-unit count is a poor measure on its own: the study found that motor carriers seen with more power-unit vehicles than they filed realize 1.71× the crashes their filing predicts, while motor carriers never seen in 24 months realize 0.50×. So exposure uses a floor: the larger of the filed count and the number of distinct power-unit VINs observed at inspections and crashes in the trailing 24 months. Trailers are excluded from the count. Tenure is measured from the motor carrier's registration date (or, when FMCSA has re-dated a motor carrier that was already operating, from its first crash or inspection on record), and exposure months are weighted by the same decay as crashes (next section), so a steady fleet contributes about 1.54 decayed power-unit-years per power unit over the 36-month window.

The class rate $\mu_{\text{class}}$ is the expected number of recordable crashes per power-unit-year for a motor carrier with this one's attributes, from a Poisson regression fitted once on 8.1 million motor-carrier-years and frozen in the model artifact. Its inputs are the facts the census and inspection files state about the motor carrier as of the cutoff:

Two findings from the study explain why the class rate carries so much of the prediction. First, whether the motor carrier actually operates is the dominant information in FMCSA data: a last MCS-150 filing more than five years old meant 0.18× the predicted crash rate in the study population, and motor carriers never seen at an inspection or crash in 24 months realized half of what their filed fleet size alone predicts — which is why the share of the filed fleet seen on the road is itself a class-rate input. Second, inspection counts measure activity plus what the inspections found, not risk in themselves: a traffic-enforcement stop carries a crash signal (relative risk about 2.4 per log unit of stops per truck) while a weigh-station pull-in is mostly exposure (about 1.15), and among motor carriers inspected often with no violation at all, more inspections mean fewer crashes. The class rate therefore carries activity by type, and the violation relativities below carry what was found.

Rate ratios are exp(coefficient) from the frozen artifact; each holds the other inputs fixed. The domicile state is deliberately not an input: it is confounded with how completely each state reports crashes, and will be revisited once a reporting-completeness factor exists.

Keeping the class rate on its own record

The class rate was fitted on dates when the roadside record behind two of its inputs was still short: VIN sightings, which measure how much of a filed fleet is seen at inspections, begin in October 2024, so the 24-month seen-share window held 1 to 10 months of them, and the inspection file held 12 to 21 of the 24 months used for activity. A live run sees full windows, places motor carriers in higher-rated bands than the fit ever did, and so expected about a third more crashes from fleets of 10+ power units than they go on to have. Until the class rate is refitted on windows every fitting date can supply, it is multiplied by a factor per fleet size that depends on how many months of VIN sightings the scoring date can see: neutral on average at the fit's coverage, at full coverage 0.77 for 1–5 power units, 0.84 for 6–9, 0.86 for 10–19, 0.85 for 20–49 and 0.73 for 50 and up. The factor fades out for records expecting almost no crashes (none below 0.02 expected over 36 months, in full from 0.2). With it, the out-of-time check predicts 112,394 crashes against 112,397 realized, where it predicted about 6% too many without it.

Three violation signals from the trailing 24 months enter as credibility-shrunk relativities and multiply into one violation relativity $V$. Each is a ratio of what was observed to what was expected, pulled toward 1.0 when the record is thin (Bühlmann credibility $Z$):

The standardisation matters: a motor carrier inspected mostly in a state that puts many trucks out of service is compared with what that regime does to everyone, not with a national average. In the study the standardised ratios beat raw per-inspection rates every time, decisively for out-of-service results. The credibility constants (10 driver inspections; 1 expected OOS event; 3 expected violations) and the exponents were chosen on a held-out split by forward crash prediction, and the violation block carries the right sign in every inspection regime.

The experience term is classical Gamma–Poisson credibility. $N$ is the motor carrier's recordable crashes over the trailing 36 months, each weighted by how recent it is; $\Lambda$ is the number of crashes the class rate and violation relativity predict for the same decayed exposure. The prior strength $a$ is the credibility constant, fitted by how well the resulting prediction explained the next year's crashes. It depends on the cohort and on how much expectation the record carries ($a = a(\text{cohort}, \Lambda)$, interpolated in $\log\Lambda$ between buckets at 0.1, 0.3, 1, 3 and 10 expected crashes): a thin record gets a diffuse prior, so a single crash counts for a lot and a clean one earns little, while a record with a few expected crashes gets a narrow one. A single constant per cohort was tested and rejected: right on average but wrong in both directions by record — under it a crash on a record expecting under 0.1 crashes was followed by 3.6× the class rate against the 1.26× it credited, and clean records expecting one to three crashes realized 21% more crashes than credited:

A crash loses one thirty-sixth of its weight each month and is gone at 36 months — there is no cliff at a window edge. The study tested every common decay shape on 24- and 36-month windows: every 36-month shape beat every 24-month shape, and the step-free ramp costs 0.19% of the best achievable fit. Credibility rises with exposure: $Z = \Lambda/(a + \Lambda)$. For a 20–49-unit motor carrier, 0.3 expected crashes in the window and none observed gives $a \approx 0.65$ and $\bar\theta \approx 0.68$; one crash on the same record gives 1.74. With 3 expected crashes, $a \approx 2.7$: a clean record sits at 0.47 and one crash at 0.65. A thin record moves a lot on one event; a well-observed one moves less. Nothing is floored or capped by hand: a crash record speaks through $N$, and the posterior says how sure we are.

The posterior for the motor carrier's true relativity is Gamma($a + N$, $a + \Lambda$), scaled by $V$. Its mean is the graded statistic, and its tail probabilities drive the screens. The 80% likely range shown on the motor carrier page is that posterior widened slightly (squared coefficient of variation $1/(a+N) + 0.01$): the range study found the true rates of the best-measured fleets scatter about ±15% around the forecast where the posterior alone claims ±11%. The page also shows the likely number of crashes in the next 12 months, the middle 80% of a negative binomial around the forecast with squared CV $1.8/(a+N) + 0.012$, which covered about 80% of outcomes for every fleet size. Neither width changes a grade or a score. The credibility tier is descriptive — it does not change the grade.

Exposure corroboration. $\Lambda$ counts the filed power units as exposure. How much of a filed fleet the roadside record supports is measured as a corroboration: full when at least half the filed units were seen in 24 months or the fleet draws at least 0.5 inspections per filed unit-year, and falling in proportion below that. The displayed credibility is scaled by it. Below half corroboration — under a quarter of the filed units seen and under 0.25 inspections per filed unit-year — a motor carrier of 10 or more filed units is flagged Exposure unverified, and an Excellent or Strong grade is held one band lower until the fleet is confirmed: the roadside record does not yet go back far enough to test whether the class rate fully prices such filings, so the favourable grade waits for it. The relativity and forecast are not scaled; the tier says how much of the exposure behind them the data has actually seen, and exposure trust enters the forecast itself through the roadside-fleet terms of the class rate and the record-dependent prior.

The motor carrier's relativity to class is $\bar R = V \cdot \bar\theta$: its predicted crash rate divided by its class rate. 1.00× means a typical motor carrier of this class; below 1 is safer, above 1 riskier. Expected crashes over the next 12 months are simply $\mu_{\text{class}} \cdot \bar R \cdot \text{PU}_{\text{floor}}$. The grade starts as a band from fixed cuts on $\bar R$. The four lower cuts were fitted so that realized crashes divided by class-expected crashes cross 0.80, 1.10, 1.40 and 1.90 at the band edges (an isotonic fit with every cohort weighted equally); the Excellent cut was set at 0.60 by decision — roughly half the crashes of a comparable fleet.

The 0–100 score interpolates linearly inside the band between anchors 100 / 90 / 75 / 50 / 30 / 10 / 0, so the score and the grade never disagree; a score of 0 is reached at twice the Critical cut. The last column is the out-of-time check of the latest weekly run, November 2, 2026 (the population as it stood on July 1, 2025 — the latest month-start whose following year ended at least 120 days before the data date — graded on the same artifact; crashes realized in that year): realized crashes divided by class-expected crashes rise monotonically from Excellent to Critical, and each grade realizes inside its own band.

The screens

The band is where the posterior mean falls. The screens ask whether the record actually supports that band — and they are one-directional: upside screens need proof, downside screens need observed events.

Satisfactory on the evidence, or not enough record

Every grade except Satisfactory has to pass an evidence test; Satisfactory is where a motor carrier lands when none passes. That covers two different situations: a record that shows the motor carrier runs close to its class, and a record too thin to say anything. When the 80% likely range still runs from below the Excellent cut (0.60×) to above the Poor cut (1.57×), the record can't tell a best-in-class fleet from a poor one, and the motor carrier is shown as Not enough record with no score, in lists sorted after graded motor carriers and in the grade-mix charts counted beside the bars rather than in them. Its expected crashes are still shown: on the select split (ten monthly origins, 2024-11 to 2025-08) these motor carriers realized 1.00× their class rate and 1.01× their forecast, the same as Satisfactory motor carriers with a supporting record. What differs is the spread: around the forecast, a too-thin record scattered by about ±68% in the ranked book against ±26% for a record that supports Satisfactory, and the model puts about one in six of them at Poor-level rates.

What the engine grades today

Displayed grades of the latest weekly run, November 2, 2026, after the governor and the caps; 5,050 Excellent or Strong model grades are held one band lower by the exposure-corroboration rule. Not graded, each with its reason on the page: 100,630 operating authority revoked, no roadside activity since; 34,209 passenger operation — outside the model for motor carriers of property; 8,017 registered after the scoring cutoff; 4,606 under a federal out-of-service order; 658 fleet above the model's fitted size range (5,000+ power units); 44 outside the scoring population at the cutoff — 148,164 motor carriers in all. Across the whole universe most motor carriers are Satisfactory because a one- or two-truck record proves little either way; the distribution opens up with evidence, which is why charts default to the 10+ book.

Every change has a named cause

Against the previous run, the change in each motor carrier's log relativity is split into six components, and the dominant one with the sign of the move names the cause shown on the motor carrier page:

Gates before every write

The run grades the population as it stood on July 1, 2025 (1,666,169 motor carriers) — the latest month-start whose following year ended at least 120 days before the data date, so the outcome year is fully reported — on the same artifact, and judges it on the crashes of that year. Hard gates abort the write; soft gates are reported in the weekly run report.

Verdict of the November 2, 2026 run: all nineteen checks ran, every one inside its bound, with no hard or soft failure. The weakest evidence bucket, 1.05 of forecast at 0–0.05 expected crashes, is recorded in the run's metrics record.

What Is Not in the Grade

The engine was chosen in a bake-off on ten monthly origins (2024-11 to 2025-08), each with the engine's windows — crashes over 36 months; inspections, violations and vehicle sightings over 24 months, as far back as the inspection file then reached — and a 12-month forward outcome. Motor carriers were split by DOT number into a fit set, a selection set, and an evaluation set that was read exactly once. Out of sample the engine explains 28.2% of the Poisson deviance in forward crashes with an exposure-weighted normalized Gini of 0.653 — 0.021 behind a gradient-boosted ceiling (0.674) while remaining three named factors. The locked evaluation set, read once on the specification frozen at the end of the bake-off, gave 27.4%, 0.647 and calibration 1.003 against the ceiling's 30.6%, 0.674 and 1.014 — the same picture.

The full study — data, folds, decay and credibility estimation, cleaning rules, variable importance, the bake-off and the grading simulation — is written up in the Empirical Validation Study; the formal specification with every constant is the Scoring White Paper. Scoring and gating run in a reproducible pipeline against the frozen model artifact, and each run's gate values are written to its metrics record.