Calibrated Arrival-Time Prediction for Indian Road Freight

Abstract

Routing engines predict how long it takes to drive a road. For Indian long-haul freight they are wrong by a factor of three to four, because a vehicle is not driving most of the time. We present an arrival-time engine assembled from three measured layers: a per-road-class speed correction derived from map-matched fleet traces, a stop model derived from the observed distribution of halt durations, and a weather layer combining published speed-reduction coefficients with two locally trained models. The output is a calibrated arrival distribution rather than a point estimate. Evaluated on 790 held-out journeys under a strict temporal split, the engine achieves a median absolute error of 2.85 h and a median absolute percentage error of 26.9%, against 7.79 h and 64.9% for the incumbent telematics estimate in production on the same fleet, and 3.50 h and 30.3% for a well-fitted constant-speed baseline. Nominal 50%, 80% and 90% quantiles achieve 44%, 77% and 88% empirical coverage. We report five hypotheses that were measured and rejected, two of which were our own models that passed conventional validation while being materially broken. The most consequential result is methodological: the definition of arrival in commercial telematics data is unreliable enough to invalidate model evaluation, and a behaviour-derived definition is required before any accuracy figure can be trusted.