MarathonTS leaderboard

MarathonTS is a benchmark of long synthetic time series. It compares the assembly of a long series from a pool of generated windows (which windows, in what order, how the seams are processed) with long-series generators, including pretrained forecasting models, on 9 open hourly datasets with 10 seeds.

• Overall leaderboard: core fidelity metrics against train, val and OOT plus utility (TSTR against TRTR) in classification and regression; lower ranks are better. top_tier marks the methods not significantly worse than the best (Wilcoxon over datasets, Holm). Read the table by tier, not by place: the tests do not separate the places within the tier, and these places change between runs on different hardware.
• Runs: the main run (1× NVIDIA H200, computing infrastructure of darya.ai) opens by default; the replication run (1× NVIDIA L4) repeats the open methods of the VAE pool on other hardware. Ranks and times are compared only within a run.
• Methods marked proprietary (zseries, the zEDGE pool, the zEDGE_TS long-series generator; zypl.ai) are evaluated as black boxes under the same protocol: their results are published, their code is not released.