How well does a model generalize to unseen real-world forecasting tasks?
Published snapshot·232,200 evaluation cases·March 2026 published study · Snapshot checked October 6, 2026
QuitoBench
QuitoBench organizes forecasting difficulty by temporal regime. Its March 2026 study evaluates 10 models over 232,200 instances drawn from application traffic. The table shows overall mean rank and mean MAE, averaged across context lengths, horizons and variate modes; inspect the authors’ breakdowns for context-specific choices.
Page data loaded 10/23/2026
What this benchmark answers
Does a model remain strong across easy and difficult forecasting regimes?
Methodology
The study compares statistical, deep learning and foundation-model approaches across eight regimes and multiple context lengths. Results are published study comparisons, rather than a continuously updated competition ranking.
CrossFormer ranks first with Overall mean rank 2.86, followed by Chronos-2 (3.36) and TimesFM-2.5 (4.21).
4 of 10 ranked entries in the model directory · Lower is better
Rankings
Overall mean rank vs Overall mean MAE
Directory model (filled)Not in directory (outline)
Full results
| Model | ||
|---|---|---|
| 1 | CrossFormer | 2.86 |
| 2 | 3.36 | |
| 3 | 4.21 | |
| 4 | 4.35 | |
| 5 | 4.36 | |
| 6 | iTransformer | 4.67 |
| 7 | TSMixer | 5.51 |
| 8 | DLinear | 7.26 |
| 9 | ES | 9.18 |
| 10 | SNaive | 9.25 |
Why temporal regimes matter
The benchmark separates trend, seasonality and forecastability into eight regimes. Its study finds context length can change the relative ordering of deep learning and foundation models. This makes a regime-specific result more useful than assuming one model wins across every series.
How to interpret it
- —Compare models at the same context length and temporal regime.
- —Published study results are not a live leaderboard; newer models may be absent.
Frequently asked questions
- Is QuitoBench a live leaderboard?
- This page reproduces the March 2026 study’s overall mean-rank and mean-MAE table, checked against the authors’ project on October 6. It is a published snapshot rather than a live submission ranking.
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?
How do models behave across realistic tasks and different temporal patterns?