A practical decision-support system for robust post-training model selection in spatiotemporal forecasting
Knowledge-Based Systems, article 116673, 2026
← Research programme
01
Research area
Reliable spatiotemporal AI
Evaluation and selection methods for forecasting systems that must remain dependable under missing data, sensor failures, spatial misalignment and changing deployment conditions.
Rationale
The research problem.
A forecasting model is useful only when the decision to deploy it remains defensible outside a single clean benchmark. This area studies the full post-training decision: how evaluation references can fail, when that failure changes model rankings and which selection rule is justified under an explicit performance budget.
Research questions
Questions that guide this area.
- 01
Does the selected model remain the same when the evaluation reference is missing, shifted or degraded?
- 02
Is the observed selection instability larger than ordinary training-seed variability?
- 03
When does robustness justify a measurable clean-performance cost?
Methods, systems and sources
Concrete research elements.
- Stress testing
- Selection regret
- Decision rules
- Audit trails
- Robust evaluation
- Distribution shift
- Model selection
Evidence