How well does a model generalize to unseen real-world forecasting tasks?
Official reference·50 datasets·Official leaderboard linked directly
TIME
TIME evaluates forecasting tasks built from fresh data, with human-reviewed task definitions and pattern-level analysis. Its official leaderboard provides task and pattern views; scores depend on the selected evaluation surface.
What this benchmark answers
How do models behave across realistic tasks and different temporal patterns?
Methodology
The benchmark evaluates 12 foundation models across 98 tasks, combining dataset quality checks with forecasting-relevant pattern analysis. Consult the official leaderboard for current submissions and selected metrics.
Explore the official interactive leaderboard and published results.
Explore the official interactive leaderboard and published results.
Fresh tasks and pattern analysis
TIME addresses limitations of reused datasets and static domain labels. Its 50 datasets support 98 tasks, and the paper evaluates 12 foundation models. Pattern-level analysis helps distinguish performance across intrinsic time-series characteristics rather than relying on one aggregate score.
How to interpret it
- —Task-level and pattern-level rankings answer different questions; retain the selected metric and task scope when comparing results.
- —This page links directly to the official interactive results; it does not present an independently synchronized leaderboard.
Frequently asked questions
- Where are the TIME rankings?
- Use the official TIME leaderboard for current rankings, metric selection and pattern-level comparisons. The accompanying result repository exposes the published evaluation files.
Related reading
Compare with other TSFM benchmarks
Which models stay strong across heterogeneous datasets and probabilistic settings?
How do models behave on observability telemetry instead of academic datasets?
Does model performance hold up as real time passes and the data distribution shifts?
Which multimodal models can reason about anomalies, timing, magnitude, and cross-series structure in production telemetry?
Does a model remain strong across easy and difficult forecasting regimes?