Does Architecture Preference Transfer in Time-Series Forecasting?
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Forecasting benchmarks are often reused to infer which architectures should work well on related tasks, but global architecture strength need not imply transferable task-specific preference. We isolate architecture preference as the residual interaction in log forecasting risk after removing architecture-wide strength and task difficulty. Across ten architectures and 41 tasks from eleven datasets, preference is strongly local to dataset identity: residual profiles correlate at across horizons of the same dataset but at across datasets. When an entire horizon is held out from a previously observed dataset, nearest-horizon reuse recovers of held-out pairwise preferences, compared with for exchangeable matrix factorization. Matched random-group controls show that this gain is specific to real dataset membership rather than generic grouped parameter sharing. The locality pattern reproduces on a separately executed panel of ten datasets and six fixed architectures (within/across ; slope ) and is stable across the tested update budgets. Strict leave-one-dataset-out prediction is substantially weaker, showing that observable metadata captures only part of preference on unseen datasets. A frozen-mask evaluation-budget audit further shows that the operational value of local reuse is budget-dependent. Overall, global architecture strength transfers broadly, whereas residual preference is predominantly dataset-local and most reusable across related configurations within an observed dataset.
- Primary area:
- learning on time series and dynamical systems