Trajectory History Drives Short-Horizon Vessel Forecasting, and Gridded Wind Encodes Position: A Controlled Study of METOC Feature Ablations

Abstract

Vessel trajectory forecasting increasingly fuses meteorological and oceanographic (METOC) observations with AIS kinematics. We report that the standard method for valuing those inputs, the feature ablation, is confounded whenever the environmental field is gridded and sampled at the vessel's own position: the sampled value is a deterministic function of location, so withholding the feature also withholds location. Vessel-local gridded HRRR wind appears to reduce final displacement error by up to 4.57%, but supplying latitude and longitude symmetrically to both arms collapses that advantage to at most 0.28%. We establish this under matched, vessel-held-out evaluation across four U.S. coastal regions and 92 summer days, comparing snapshot and history-controlled LightGBM, a GRU consuming 25 minutes of trajectory history, and a fold-tuned steady-state CV Kalman baseline, each with and without public METOC features. The same protocol identifies what does drive skill. The GRU reduces 60-minute forecast error by 17-49% over constant-velocity extrapolation, and history-controlled LightGBM outperforms the tuned Kalman baseline in every region-horizon cell (by +6.09-+45.63%) while producing sharper calibrated 90% radial error bounds at matched coverage; matched comparisons attribute the gain to trajectory history rather than model family. Point METOC gains are small and model-dependent: an initially promising San Francisco sailing-vessel result (+2.69%, 95% CI [2.02, 3.38]) erodes under history, multi-seed, and strict forward-day controls. Across 7,454 vessels and 593,780 forecast origins, trajectory history is the durable source of improvement; environmental gains should be credited only after symmetric spatial, temporal, and model controls. This record contains the author's preprint. It has not undergone peer review and may differ from the subsequently accepted or published version.