Test the physical mechanisms behind long-context skill in pooled buoy data

Investigate through controlled ablation of buoy subsets whether swell propagation across the buoy network and altered wave-climate sampling explain why the 47-buoy corpus exhibits long-context skill recovery that is absent from the single-buoy experiment.

Background

The single-buoy experiment found that persistence skill peaked at 12–24 hours and declined at longer context lengths, whereas the multi-buoy evaluation showed skill recovery between 48 and 96 hours and a plateau through 168 hours. The paper proposes two possible explanations: pooled buoys may expose models to implicit long-range swell-propagation patterns, or buoy mixing may broaden the training distribution and shift the optimal context window.

These explanations are presented as hypotheses rather than established mechanisms. Controlled removal or grouping of buoy subsets is needed to determine whether cross-buoy propagation or distributional heterogeneity causes the different context-length behavior.

References

These hypotheses remain to be tested through controlled ablation of buoy subsets.

— On the Limits of Univariate Deep Learning for Significant Wave Height Forecasting  (2609.30688 - Zhai et al., 25 Sep 2026) in Section 4.2, “Context-length scaling”