Research Paper

Key Results

1. Introduction

Motor carrier safety is a critical concern for insurers, regulators, and the public. The Federal Motor Carrier Safety Administration (FMCSA) collects extensive data on carrier operations, including crash records, roadside inspections, and violations. A fundamental question for the industry is: can this data predict which carriers will crash in the future?

Previous research has shown correlations between safety metrics and crash history, but correlation is not prediction. A carrier's past crashes may simply reflect its size (more miles = more opportunity for crashes) rather than underlying risk. To be useful for underwriting, loss prevention, or regulatory targeting, a predictive model must demonstrate temporal validity—the ability to identify future crashes, not just explain past ones.

This study addresses this question using a rigorous temporal split methodology: we use only 2024 data to score carriers, then measure how well those scores predict 2025 crash outcomes. This approach ensures we're testing genuine predictive power, not hindsight.

Research Questions

1. Can publicly available FMCSA data predict which carriers will crash in subsequent years?
2. Which data sources provide the strongest predictive signal?
3. Do behavioral violations (driver decisions) predict crashes better than equipment violations (vehicle condition)?
4. Does predictive power vary by carrier size?

2. Data and Methods

2.1 Data Sources

We obtained the following datasets from FMCSA's Safety Measurement System (SMS):

2.2 Study Population

To ensure statistical reliability, we restricted the analysis to carriers with:

• At least 50,000 annual miles (sufficient exposure for rate calculation)
• At least 1 inspection in 2024 (observable activity)
• Active census registration

This yielded a study population of 243,333 carriers.

2.3 Temporal Split Design

All predictor variables (crash counts, violation counts, rates) were calculated using only Year 1 data. The outcome variable—whether a carrier crashed in Year 2—was completely withheld during model development.

2.4 Violation Categorization

We classified violations into two categories based on the FMCSA Group_Desc field:

2.5 Evaluation Metrics

• AUC (Area Under ROC Curve): Measures discriminatory power. 0.5 = random, 1.0 = perfect.
• Gini Coefficient: 2 × (AUC - 0.5). Standard insurance industry metric.
• Lift: Ratio of crash rate in top decile vs overall rate.
• Capture Rate: Percentage of Year 2 crashes found in top X% of risky carriers.
• Relative Risk (RR): Crash rate with factor / crash rate without factor.

3. Results

3.1 Overall Predictive Performance

The combined model achieved strong predictive performance on the held-out 2025 data:

3.2 Crash Capture by Risk Decile

Sorting carriers by their 2024 risk score, we measured how 2025 crashes distributed across deciles:

3.3 Predictor Comparison: What Predicts Best?

We compared the predictive power of different data sources using relative risk (RR) and lift metrics:

3.4 Size-Stratified Analysis: Hidden Predictive Power

A critical finding is that violation predictive power varies dramatically by fleet size. When we stratify by annual mileage, behavioral violations become much stronger predictors:

3.5 Cumulative Effect: Multiple Behavioral Violation Types

Perhaps the most striking finding is the cumulative effect of having multiple types of behavioral violations. The count of distinct violation categories is itself a powerful predictor:

4. Discussion

4.1 Answer to Research Questions

4.2 Why Behavioral Violations Predict Better

We hypothesize several mechanisms:

• Driver behavior is causal: Speeding, distraction, and impairment directly cause crashes. Equipment defects are often incidental findings during inspections.
• Behavioral violations reflect culture: A carrier with multiple behavioral violation types likely has weak safety management, poor driver selection, or inadequate supervision—systemic factors that persist over time.
• Equipment can be fixed quickly: A brake defect found today may be repaired tomorrow. A driver who speeds today will likely speed tomorrow.

4.3 Practical Applications

4.4 Limitations

• Selection bias: Only carriers with inspections are observable. Carriers avoiding inspections may differ systematically.
• Mileage accuracy: MCS-150 mileage is self-reported and may be unreliable for some carriers.
• Crash reporting: Minor crashes may be underreported, especially for small carriers.
• Two-year window: With only 2024-2025 data, we cannot assess longer-term stability or multiple validation periods.

5. Conclusion

This study demonstrates that motor carrier crashes can be predicted with strong accuracy using publicly available FMCSA data. The key findings are:

1. Prediction is possible and practical. AUC 0.848 and 5.8x lift indicate that risk scoring can meaningfully differentiate carriers. The top 10% of risky carriers account for 58% of future crashes.
2. Behavioral violations are the key signal. Driver decisions (speeding, reckless driving, distraction) predict crashes far better than equipment conditions. A carrier's behavioral violation profile reveals its safety culture.
3. Size stratification unlocks hidden power. Violations that appear weakly predictive overall become 3x+ predictive within fleet size bands. One-size-fits-all models leave significant value on the table.
4. Multiple violation types compound risk. The number of distinct behavioral violation categories is itself a powerful predictor. Carriers with 3+ types have 65% crash probability; 5+ types have 91%.
5. Past crashes remain the best single predictor. But they are size-biased. Combining crash history with behavioral violations and size stratification yields the strongest models.

Data Availability

The underlying data for this study is publicly available from FMCSA's Safety Measurement System (SMS) data downloads. The carrier-level dataset used in this analysis (243,333 carriers with Year 1 predictors and Year 2 outcomes) is available upon request for research purposes.

For methodology details on specific violation categorizations and rate calculations, see our Violation Type Study.