In brief
Choosing the best forecasting model from one clean validation set can produce a fragile deployment decision. This work tests whether that decision survives realistic changes to the evaluation reference and turns the diagnosis into an auditable recommendation.
Contribution
A post-training workflow that reevaluates fixed predictions, measures selection instability relative to ordinary training variability, and compares robust decision rules under a declared performance budget.
Key finding
Robust selection is not automatically better. It is useful when the diagnosed instability is material and the assumed perturbations are credible; otherwise, clean selection may remain the justified choice.
Models and data
Distinctive research elements
Models and methods
- Post-training decision-support system
Data and evaluation
- EEA-PM25-ES
- Six real-world datasets
- Synthetic benchmark
Author-written overview
Research overview
Model selection in spatiotemporal forecasting is commonly based on performance against one evaluation reference, although that reference may be incomplete, spatially displaced or degraded in operation. This work introduces a post-training decision-support workflow that reuses fixed prediction bundles to test a selected model across declared reference perturbations. It compares selection instability with variability across training seeds, measures selection regret, evaluates alternative decision rules and reports a deployment recommendation constrained by a clean-performance budget. Experiments cover six real datasets and a synthetic benchmark. The results show that no robust rule is universally preferable: the justified choice depends on the measured instability, the credibility of the perturbations and the available budget. On EEA-PM25-ES, robust selection changes the recommendation and reduces median decision regret from 0.1403 to 0.0000 under the declared perturbation family. The resulting workflow makes the assumptions, trade-offs and deployment recommendation traceable.
Cite this work
BibTeX
@article{semper2026robust,
title = {A practical decision-support system for robust post-training model selection in spatiotemporal forecasting},
author = {Marc Semper and Manuel Curado and Jose F. Vicent and Leandro Tortosa},
journal = {Knowledge-Based Systems},
year = {2026},
eid = {116673},
doi = {10.1016/j.knosys.2026.116673}
}Always check the publisher record for final volume, issue and page information before citing.