Calibration Before Complexity: Forecasting Compound-Hazard Exposure for U.S. Freight Systems from Public Multi-Source Data

Freight systems face hazards that seldom arrive alone, yet predictive studies typically model weather, disaster, trade and financial shocks in isolation. We assemble a leakage-controlled state–day panel for the United States that aligns disaster signals (NOAA Storm Events and FEMA declarations), freight exposure (Commodity Flow Survey and Freight Analysis Framework version 5) and macro-geopolitical context (imports, the CBOE Volatility Index and the Geopolitical Risk index), and forecast whether a composite hazard shock index, a proxy for conditions under which freight disruption becomes likely, will exceed its training-period 90th percentile on the following day. Four learners were trained to December 2022 and evaluated once from January 2024 (22,617 and 11,296 state-days), with 2023 withheld. Discrimination was moderate and nearly indistinguishable across three models (ROC-AUC 0.725–0.733 for gradient boosting, logistic regression and histogram-based gradient boosting; 0.706 for random forest), but probabilistic accuracy diverged sharply. Only gradient boosting outperformed a constant base-rate forecast (Brier skill score ≈ 0.09); logistic regression and histogram-based boosting over-predicted risk at every level (skill ≈ −0.8 to −0.9). Event prevalence rose from 10% to 17.2% between windows. Ablations showed that hazard persistence carries most of the signal, freight exposure adds little, and a wider macroeconomic block reduces out-of-sample discrimination. At a seven-day horizon, logistic regression was the most stable learner. A newsvendor illustration shows how uncalibrated probabilities near a decision threshold translate into systematic over-stocking. The study provides a transparent public-data benchmark and shows that, for logistics risk screening, calibration and temporal validation matter more than model complexity.