Forecasting leaderboard
Forecasting methods in sktime, from naive baselines to foundation models, evaluated on the same datasets with the same cross-validation. Each method is ranked on every dataset, and the table orders them by their average rank.
- Catalogue
- Sample task
- Methods
- 12
- Datasets
- 2
- sktime
- 1.2.0
- Benchmark run
- Oct 18, 2026
Mean absolute error, lower is better. Ranked across 2 datasets, 12 of 12 methods shown.
- Rank on one dataset
- Average rank
| Rank | Method | Rank on each dataset | Average rank | Fit time | Predict time |
|---|---|---|---|---|---|
| Leading method: 1 | AutoARIMA | 3.00 | 165 ms | 3.28 ms | |
| 2 | Toto 2.0 - 22m | 4.00 | 2.95 ms | 40.6 ms | |
| 3 | Theta | 5.50 | 4.78 ms | 1.47 ms | |
| 3 | TimesFM 2 - 200m | 5.50 | 15.9 ms | 692 ms | |
| 5 | Chronos Bolt Tiny | 6.50 | 882 ms | 10.1 ms | |
| 5 | Damped Exponential Smoothing | 6.50 | 6.21 ms | 1.39 ms | |
| 5 | Holt's Exponential Smoothing | 6.50 | 4.77 ms | 1.14 ms | |
| 5 | Naive Drift | 6.50 | 0.791 ms | 0.757 ms | |
| 9 | Naive Last | 7.00 | 0.801 ms | 4.44 ms | |
| 9 | TiRex 2 | 7.00 | 28.3 ms | 1.34 s | |
| 11 | Simple Exponential Smoothing | 8.00 | 4.27 ms | 1.19 ms | |
| 12 | Naive Mean | 12.00 | 0.802 ms | 0.809 ms |
On each dataset, rank 1 is the best mean absolute error, and tied scores share the average of the ranks they span. A method’s average rank is the plain mean of its dataset ranks. Raw mean absolute error values are never averaged across datasets, because datasets have different scales. Methods with the same average rank share a place.
How the ranking works
The leaderboard reads precomputed benchmark results from sktime’s benchmarking module. It does not train or run any model on this page.
- Rank per dataset
- Every method is evaluated with the same cross-validation folds. On each dataset, methods are ranked by their mean score across folds.
- Average the ranks
- The overall order is the mean of those dataset ranks. Ranks are comparable across datasets; raw errors are not.
- Ties and gaps
- Tied scores share a rank. A method missing a dataset is left out of that dataset’s ranking instead of getting a filled-in score.
Datasets in this catalogue
| Dataset | Frequency | Length | Evaluation |
|---|---|---|---|
| Airlineunivariate | Monthly | 144 points | ExpandingWindowSplitter |
| Longleyunivariate | Yearly | 16 points | ExpandingWindowSplitter |
Source data
Results, metadata, and the reference ranking code are published with the leaderboard repository. View the leaderboard repository
Exported Oct 20, 2026 · 40a19e6