Training-Free Method Predicts Trajectories with High Accuracy

Predicting spatiotemporal trajectories, such as maritime vessel movements, typically relies on complex, large-parameter transformer models that require extensive training data and significant computational resources. This paper introduces a novel, training-free method for multi-modal trajectory prediction that challenges this paradigm. The approach operates without learned parameters and does not require a GPU for training, yet achieves accuracy comparable to a 57-million-parameter transformer.The method constructs a simple transition table from historical state-to-next-position pairs. During inference, it retrieves neighboring states using a product kernel that considers spatial proximity, bearing, speed, and temporal context. Two inference modes are available: diversity-penalized sampling to generate distinct plausible routes, and beam search to find the highest-likelihood path.Evaluated on the TrAISformer benchmark (Danish Maritime AIS data), this non-parametric method demonstrated competitive accuracy when full data was available. Crucially, it dramatically outperformed the transformer in data-scarce regimes, maintaining stability even with only 10% of the training data, where the transformer's performance catastrophically degraded. This breakthrough enables deployment in new geographic regions with significantly less historical data and eliminates the need for GPU training.