Weir
Time-series foundation models, ensembles and statistical baselines, forecasting the same fresh real-world series twice a day. Scores are relative to a seasonal-naive forecast, so lower is better and 1.0 means no better than repeating the last season.
Leaderboard
↓ lower is better: CRPS, MASE and every rank (1 is best)↑ higher is better: beats naive
| Rank | Model | Size | Organisation | Type | CRPS rank ↓ | MASE rank ↓ | Licence |
|---|
Click a column heading to rank by it. CRPS scores the whole forecast distribution and MASE the median, both relative to seasonal naive. Beats naive is the share of challenges where the model's CRPS beats seasonal naive, with a dot where that is significant (Wilcoxon, Holm-corrected). Avg rank is the mean CRPS rank over the last 14 rounds the model took part in.
Over time
CRPS relative to seasonal naive in each round; lower is better. Select any number of models below to compare them.
What the round contained
Each round draws a fresh, jittered mix of live series so no model can be tuned against a fixed benchmark.
Domains
Sampling rate
How rounds work
Fresh data. TSBench-Forge scrapes live public sources and samples windows from after each model's training cutoff, so no model can have memorised the answer.
Same challenges for everyone. Every row forecasts the same windows through the live Ephemeris API, then is scored on the same seed. Five jittered draws per round are pooled before ranking.
Changes are marked. On 2026-10-06 the ensemble switched to the accuracy-weighted mixture it serves today (Ephemeris 0.7.5, PatchTST-FM r2 replacing r1). Charts mark the break.
Open data. Every round is published as JSON: rounds index, history, and the TSBench-Forge repository.