Leakage-controlled benchmarking of kinematic and machine-learning models for ten-minute buoy drift forecasts

Reliable short-horizon drift estimates can support maritime search planning when environmental forcing fields are incomplete. However, trajectory-derived predictors can inadvertently contain information from the forecast interval, random splitting can mix strongly related observations across data partitions, and complex models can appear skillful without outperforming simple motion extrapolation. Rather than proposing a new predictor, we establish a fair, leakage-controlled evaluation framework based on chronological validation and explicit kinematic baselines. The framework was applied to 2,317 Taiwan Strait buoy observations. Each forecast used the current position, velocity calculated only from the preceding observation interval, and a requested lead time restricted to 9.5–10.5 min. Persistence and time-normalized constant-velocity extrapolation were compared with ridge regression, radial-basis-function support vector regression, random forest, extremely randomized trees, and a multilayer perceptron (MLP). Learned models predicted a correction to the kinematic forecast; all hyperparameters and the MLP seed were selected only within the chronological development partition. Final evaluation used 691 eligible forecasts in the untouched held-out test block, with four post hoc trajectory segments examined descriptively. Constant velocity achieved the lowest full-test coordinate RMSE and mean position error (17.02 and 20.60 m); Extra Trees ranked second (18.22 and 21.95 m). Learned corrections produced a small improvement only on the short single-curvature segment and no consistent advantage elsewhere. The MLP was less accurate than constant velocity on the full test and every segment. These findings identify local velocity persistence, rather than model complexity, as the dominant source of ten-minute forecast skill in this deployment. Because all observations came from one drifter, independent deployments, longer horizons, and forcing-aware evaluation are required before operational use.