The U.S. trucking industry, in numbers.

FMCSA's crash file is built from state police reports. About 22% of those reports leave the carrier's DOT number blank — the officer didn't have it, the carrier doesn't operate under a DOT, or the carrier is defunct. We deliberately drop these from per-carrier aggregates rather than guess (wrong attribution would be much worse than missing data). Industry-level totals on this page therefore undercount actual crashes by approximately the same percentage. State-level counts are not affected (we group by report-state, which is always populated).

Police reports take time to flow from the original officer through state DMVs to FMCSA. The most recent month in any crash dataset is typically 50-70% reported. Our state-crash forecast trims trailing months whose count is <50% of the trailing-12-month median before fitting, but the rolling industry chart at the top of this page shows raw monthly counts including the partial latest month.

Power-unit counts come from each carrier's most-recent MCS-150 filing. Carriers update voluntarily (or under enforcement pressure), so the number reflects what they told FMCSA, not what's actually on the road today. Small carriers update rarely. Treat all fleet-size figures as approximate.

Our carrier-survival model defines a "failure" as either formal authority revocation OR transition to inactive status (USDOT code I). Inactive carriers may have voluntarily ceased operations, retired, or been acquired — these are economically different from regulatory failures. We chose this combined definition because it captures the most common new-carrier exit modes, but the model can't distinguish between them.

Each predictive model is walk-forward backtested on the most recent 18 months. That's a small enough sample that observed metrics carry meaningful uncertainty. Wherever possible we report 95% bootstrap confidence intervals — read them. A Pearson r of 0.45 with CI [0.10, 0.75] is meaningfully different from r=0.45 with CI [0.40, 0.50].

The 120-month training window includes the pre-2020 freight regime (low-volatility expansion), the 2020-2021 COVID shock, the 2022-2023 freight boom + bust, and the current normalization. Stock-macro relationships are genuinely different across these regimes. We mitigate with time-decayed sample weights (recent samples weighted higher) and explicit regime-flag features, but the basket forecast in particular is sensitive to regime drift. The 1-month basket prediction is the noisiest target in the suite and should be treated more directionally than precisely.

We post-hoc calibrate the survival classifier with isotonic regression. After calibration, predicted probabilities match observed rates within ~13 percentage points across the validation cohort. That's much better than the uncalibrated baseline (~30pp gap) but still not perfectly calibrated, especially in the 50-70% predicted-failure bucket. Use the predicted survival rate as a directional indicator more than a precise probability.

The Cass indexes shown in features and the published predictions are © Cass Information Systems, Inc., and ACT Research Co., LLC. We use them under fair-use for analytical purposes with attribution. We do not republish the raw values, only model-derived predictions.

Per-carrier stock prices come from yfinance's scrape of Yahoo Finance. Yahoo's Terms of Service technically prohibit automated scraping, and the underlying API has historically been unstable. Migration to a licensed feed (Polygon, Alpha Vantage paid tier) is on our roadmap. None of the published predictions are investment advice — see below.

The Logistics Managers' Index (LMI), Manheim Used Vehicle Value Index, and Port of Los Angeles container-throughput data are paywalled or rendered client-side and could not be ingested cleanly. They are not in any of the predictive models on this page. If you have access to a paid feed for any of these, the forecast pipeline accepts arbitrary {month, value} series via a one-line addition.