Predicting Departure Delays in Time Critical Transportation: A Study of E-Commerce Dispatch and School Bus Operations

Purpose: This synthetic simulation study verifies a protocol for evaluating departure-delay prediction in stylized parcel dispatch and school bus operations. It tests whether information available before scheduled departure can rank lateness risk and support alerts within limited human review capacity. Methodology: The benchmark retains the prior five-seed experiment and adds twenty-seed Monte Carlo analysis, paired model comparisons, formal calibration measures, fixed-workload alert curves, delayed and missing observations, and one-factor sensitivity analyses. Each primary seed generates 9,000 stylized parcels and 9,000 stylized school-bus departures across 300 abstract operating days, with snapshots at 90, 60, and 30 minutes before schedule. The generator creates 90,000 distinct departures in the legacy five-seed analysis and 360,000 in the expanded analysis. No Chewy, carrier, school district, driver, employee, customer, student, route, or shipment records are used. Findings: In the shifted synthetic test period, late-departure prevalence was 19.1% (parcel) and 12.8% (school). At 60 minutes, the multivariable logistic model achieved mean average precision of 0.741 and 0.580, compared with 0.518 and 0.487 for the slack-only model; mean Brier scores were 0.088 and 0.079. Ranking was similar to one fixed gradient-boosting configuration. The shift worsened calibration and increased alert burden. At 10% review capacity, the model captured 43.8% and 48.1% of late parcel and school departures, respectively. Unique Contribution to Theory, Practice and Policy: The study formalizes departure readiness, information timing, calibration, and human review capacity within one evaluation protocol. It gives practitioners reproducible tests for alert performance and offers policymakers criteria for local validation and safety review. The synthetic results verify protocol behavior under the specified generator but do not establish real-world effectiveness or deployment readiness.