Which model for the next data point?
The next observation, scored alone. Every model paired against a reference on the same series.
Series
Frequency
Reference
Series types
| model | series | win / loss* | ΔLL | CRPS | cov90 | licence |
|---|
laplace++ is trained on one half of each panel (split by series family) and scored only on the other half, so its row covers fewer series; tick held-out series only to put every model on that same half.
* By log score: a per-series Diebold–Mariano test on the log predictive density, 64 one-step points per series. Click a column to sort.
foundation model
classical, refit every step
laplace and laplace++
open commercial use allowed
restricted free below a revenue threshold
non-commercial research and evaluation only
TimeH2O: h2oai/timeh2o-v1 at Hugging Face revision c556990 (weights unchanged since 60b68fa, 2026-10-22), evaluated 2026-11-03 to 2026-11-04 under its non-commercial licence. Method: methodology.
Built 2026-11-09 from skaters commit 6c5da15 (2026-11-06).