TabPFN against a NARX baseline on mooring-line tension

Benchmarked zero-shot TabPFN against a domain-specific NARX recurrent network on an offshore mooring-line tension regression task of roughly 100,000 rows. TabPFN fits in about 1.4 seconds and then spends about 166 seconds predicting — the wrong shape for this workload. Its prior is built for small tabular problems, not a large physical time series.

The write-up leads with the threats to validity rather than the headline number: the NARX baseline metrics are missing, the NRMSE normalisation is ambiguous in a way that inverts how the headline number should be read, and sea-state distribution shift is unaddressed. The comparison is not settled and the write-up says so.