Geodesic-based Forecast Evaluation Metrics
Background: Forecast reconciliation methods on nonlinear manifolds are typically evaluated using Euclidean-based scoring rules, which may not capture the geometry of the constraint space.
Question / Future Work: There is a need to develop proper scoring rules that are explicitly geodesic-distance-based or manifold-aware to provide theoretically valid evaluation of reconciled distributions on nonlinear coherent manifolds.
Why It Matters: Misalignment between the evaluation metric and the underlying geometry of the forecast space can lead to inaccurate performance assessment and suboptimal model selection.
Metadata & Links
- created_at
- 2026-05-02T05:49:29Z