Methodology
Motor Truck (For-Hire) Carrier Ranking Model
1) Executive Summary
This model produces a single, auditable safety ranking for for-hire property carriers. It combines five components—BASIC-based safety, crash rate, critical violation rate, experience, and behavioral risk— into a combined_score (0–100), then ranks carriers and assigns a score-based grade (Exceptional to Critical).
The score is a predictive safety indicator within the eligible population. Our temporal validation study demonstrates strong predictive power: AUC 0.848, with the top 10% of risky carriers capturing 58% of the following year's crashes.
2) Data Sources & Key Fields
The model uses publicly available FMCSA data (conceptually):
- •Carrier registration / census — authority status, carrier type, power units, drivers, authority dates
- •BASIC snapshot — percentile measures and alerts across the 7 BASIC categories
- •Crashes — reportable crashes with dates and outcomes
- •Inspections & violations — inspection dates and violation severity weights
The latest BASIC record is selected deterministically by date (never an arbitrary row).
3) Population Definition / Filtering
The model targets for-hire, property/cargo motor truck carriers with active operating authority and at least one power unit. Exclusions are applied to avoid non-comparable operations:
- •Exclude passenger-focused operations and passenger equipment
- •Exclude private property carriers without for-hire authority
- •Exclude carriers with zero truck/tractor power
Carriers with authorized or exempt for-hire authority (including US Mail and governmental authority when present) remain in-scope.
4) Time Window & Exposure Definition
Events are counted over a fixed evaluation window of 24 months. Exposure is defined as window mileage (annual mileage × 24/12) and expressed in 100k-mile units for rate calculations.
Exposure-adjusted rate:
$$r_i = \frac{y_i}{E_i}$$Where $y_i$ is event count and $E_i$ is exposure (in 100k miles).
5) Empirical Bayes Shrinkage (Poisson–Gamma)
To stabilize rates for small fleets, crash and critical violation rates are shrunk toward the fleet mean using a Poisson–Gamma empirical Bayes model (credibility weighting).
Likelihood:
$$y_i \mid \lambda_i \sim \text{Poisson}(E_i \lambda_i)$$Prior:
$$\lambda_i \sim \text{Gamma}(\alpha, \beta)$$(shape $\alpha$, rate $\beta$)
Posterior mean (Empirical Bayes rate):
$$\hat{\lambda}_i = \frac{\alpha + y_i}{\beta + E_i}$$Fleet-level method of moments (trimmed for robustness):
$$\mu = \text{mean}(r_i) \qquad \sigma^2 = \text{var}(r_i)$$ $$\alpha = \frac{\mu^2}{\sigma^2} \qquad \beta = \frac{\mu}{\sigma^2}$$Extreme rate outliers are trimmed before estimating $\mu$ and $\sigma^2$ to reduce the impact of data anomalies.
Worked Example
Suppose a carrier has 2 crashes in 200k miles. Exposure $E_i = 2.0$.
If $\alpha = 1.2$ and $\beta = 3.0$, then:
6) Component Scoring
Crash Rate (EB)
45%Uses EB-shrunk crash rate per 100k miles and compares to fleet mean. Our validation study shows past crashes predict future crashes with 1.77x relative risk.
Safety (BASIC)
15%Uses the latest BASIC snapshot and active alerts across the 7 categories. Alerts reduce a 0–100 starting score.
Critical Violations (EB)
15%Uses EB-shrunk critical/severe violation rate per 100k miles and compares to fleet mean.
Experience
15%Years since authority date mapped to 0–100 via a saturating curve. Newer carriers are not overly penalized.
Behavioral Risk Caps
10% + CapsBased on our violation type study, specific behavioral violations trigger automatic score caps regardless of other factors:
- • Reckless Driving → Cap at 49 (Critical) — 1.49x relative risk
- • Drug/Alcohol → Cap at 59 (Poor) — 1.29x relative risk
- • Speeding 10+ → Cap at 59 (Poor) — 98% crash probability
- • Speeding 5-9 → Cap at 69 (Marginal) — 87% crash probability
7) Smooth Score Mapping Function
Rate ratios are converted to scores with a smooth, monotone sigmoid. This avoids hard clamps and makes score changes gradual and defensible.
Sigmoid Score Mapping:
$$\text{score} = \frac{100}{1 + \left(\frac{RR}{k}\right)^p}$$Where $RR = \dfrac{\hat{\lambda}_i}{\bar{\lambda}}$ is the rate ratio, with $k = 1.0$ and $p = 1.5$
Illustrative curve: RR=1 yields a mid-range score, RR<1 increases scores, RR>1 decreases scores.
8) Weighting, Behavioral Caps & Missing Data
The combined score is a fixed weighted average, then subject to behavioral caps:
Final Score with Behavioral Caps:
$$\text{Final Score} = \min(\text{Base Score}, \text{Behavioral Cap})$$Behavioral caps override high scores when critical violations are present.
- •Weights are fixed for auditability. Behavioral caps are applied post-calculation.
- •6,965 carriers currently have behavioral caps applied due to reckless driving, drug/alcohol, or excessive speeding violations.
- •Missing BASIC data defaults to a neutral safety score with no alerts.
9) Behavioral Violation Risk Analysis
Our violation type study analyzed 6.5 million FMCSA violation records and found that behavioral violations predict crashes far better than equipment violations. Driver decisions (speeding, reckless driving, distraction) are more predictive than vehicle conditions.
Critical Behavioral Violations
These violations trigger automatic score caps:
| Violation Type | Relative Risk |
|---|---|
| Reckless Driving | 1.49× |
| Driving Fatigued/OOS | 1.36× |
| Drugs | 1.29× |
| Alcohol | 1.29× |
High-Risk Behavioral Violations
These contribute to behavioral score component:
| Violation Type | Relative Risk |
|---|---|
| Dangerous Driving | 1.18× |
| Phone/Texting | 1.16-1.18× |
| Speeding (severe) | 1.17× |
| HOS Fraud (False Log) | 1.12× |
Dose-Response Evidence
Our research shows clear dose-response relationships between violation counts and crash probability:
0 speeding violations
5-9 speeding violations
10+ speeding violations
10) Predictive Validation
Our predictive validation study uses a temporal split methodology: we trained on 2024 data and tested on 2025 outcomes to ensure genuine predictive power.
Key Finding: Multiple Behavioral Violation Types Compound Risk
Carriers with 3+ behavioral violation types in Year 1 have a 65% probability of crashing in Year 2. With 5+ types, it rises to 91%. This suggests the count of distinct violation categories—not just total violations—captures something fundamental about organizational safety culture.
11) Ranking & Grading
Only eligible carriers receive a rank and grade. Eligibility requires window miles ≥ 100k or at least 1 inspection in the 24-month window. Low-credibility carriers are flagged but not ranked.
Rankings are deterministic: sort by combined_score (descending), then by window miles (descending), then by DOT (ascending). Grades are assigned based on the carrier's combined score value:
12) ISS Score (Complementary Safety Indicator)
In addition to the FRED Score, we provide the Inspection Selection System (ISS) score, implementing the FMCSA December 2012 ISS-CSA Safety Algorithm. ISS is used by roadside inspectors to prioritize which carriers to inspect.
Inspect (75-99)
High-priority carriers with multiple BASIC alerts or high-risk indicators. OOSO carriers receive ISS=100.
Optional (50-74)
Moderate-risk carriers with some alerts. Inspection at officer discretion.
Pass (1-49)
Low-risk carriers with no active alerts. Lower inspection priority.
ISS Algorithm Details
- •Cohort-based ranking: Carriers are grouped (1-13) based on alert patterns, then ranked within each group by sum of BASIC percentiles.
- •High-risk detection: Group 1 triggered by ≥4 alerts OR any Unsafe/HOS/Crash BASIC ≥85%.
- •Crash alert computation: Since FMCSA basic files lack Crash BASIC data, we compute crash alerts from our crash records (≥0.10 crashes per 100k miles OR ≥10 crashes).
- •NULL for insufficient data: Carriers without BASIC data receive NULL (not a guessed score).
- •Fully deterministic: No randomness in scoring — same inputs always produce same outputs.
FRED vs ISS: FRED Score is our comprehensive proprietary risk assessment using EB-shrunk rates and multiple factors. ISS is the official FMCSA inspection prioritization score. Both are valuable — FRED for underwriting decisions, ISS for regulatory compliance context.
13) Diagnostics, Sanity Checks & QA
- •Score bounds enforced to [0, 100].
- •EB priors validated for finite, positive parameters.
- •Deterministic BASIC snapshot selection for audit trails.
- •Summary distributions and grade counts logged for each run.
- •Export QA: row counts and index creation validated on output tables.
14) Limitations, Intended Use & Governance
- •Depends on data completeness and timeliness of FMCSA feeds.
- •Does not model cargo type, geography, or claim severity directly.
- •Designed as an underwriting aid; final decisions require human review.
Governance: Model inputs, weights, and thresholds are versioned. Recalibration is recommended on a regular cadence (e.g., quarterly or when data distributions shift), with back-testing and documentation of changes.