MarathonTS — synthetic long time series and benchmark results

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. Every series is evaluated by its fidelity to the real data (83 metrics against train, validation and an out-of-time split), its utility (TSTR against TRTR in 14 classification and regression tasks), memorisation and cost.

This dataset holds 3,220 synthetic long hourly series and 63,000 synthetic windows generated from seven open datasets by 30 methods — window assembly from generated pools and long-series generators — together with everything measured on them: per-series fidelity metrics, utility predictions, run records, the collected tables, the rankings and the reports of the two published runs.

Resource	Link
Paper	arXiv (the link is added after the announcement)
Code	https://github.com/zypl-ai/MarathonTS
Package	https://pypi.org/project/marathonts/ (pip install marathonts)
Documentation	https://zypl-ai.github.io/MarathonTS
Leaderboard	https://huggingface.co/spaces/zyplai/MarathonTS

Synthetic data

What	Folder	Size	Format
Long series — assembled from windows (stitched) or generated directly	<run>/series/<dataset>/<method>/seed<k>.parquet	2,100 (run_h200) + 1,120 (run_l4) series of 5,040 hours (30 weeks)	one row per hour, real units
Window pools — generated windows before assembly (not stitched)	<run>/pools/<dataset>/<generator>/seed<k>.parquet	140 pools (run_h200: VAE and zEDGE) + 70 (run_l4: VAE), 300 windows of 168 hours each	array (300, 168, 3) in [0, 1]

Datasets: beijing_pm25, bike, bitcoin, electricity, household, jena, metro (the synthetic data of ETTh1 and PJM are withheld for licence reasons, see below), 10 seeds each.

Long series. Columns: t (hour, 0 … 5039), value (the synthetic series, in the units of the real one), how (hour of week, 0–167), window (index of the pool window the point comes from; −1 for a long-series generator), piece (consecutive piece of the series), seam (true at the first point after a seam). The method name tells how the series was made:

• vae_<assembler>, zedge_<assembler> — assembled from the pool of the VAE or of zEDGE by an assembler (rand random, greedy, atsp, pctsp, mbb, knn, quilting, herding, zseries), and <pool>_zseries_{xfade,wsola,quilt} — the same path with the seams processed;
• diffusion_ts, chronos, timesfm, timestral, zedge_ts_long, zedge_ts_ar — long-series generators (no pool; the seams are the boundaries of their generation chunks).

The same in one table. synthetic/series/ and synthetic/windows/ repeat the series and the pools in long format, with the run, dataset, method (or generator), seed and window as columns; they are the synthetic_series and synthetic_windows configs of this dataset:

from datasets import load_dataset

series = load_dataset("zyplai/MarathonTS", "synthetic_series", split="run_h200")    # 2,100 series, 10.6 M rows
windows = load_dataset("zyplai/MarathonTS", "synthetic_windows", split="run_h200")  # 42,000 windows, 7.1 M rows
one = series.filter(lambda r: r["dataset"] == "bike" and r["method"] == "zedge_quilting" and r["seed"] == 0)

Other configs: leaderboard (the overall leaderboard of each run, default), fidelity and utility (every metric of every series of the main run), datasets (size, split, range and scaling of each dataset).

Window pools. Channel 0 is the value scaled by min-max on the training split, channels 1–2 the hour of week as (sin + 1) / 2 and (cos + 1) / 2. Back to real units: value = lo + x · (hi − lo) with lo, hi of the dataset in <run>/tables/datasets.parquet.

import json
import numpy as np
import pandas as pd
from huggingface_hub import hf_hub_download

repo = dict(repo_id="zyplai/MarathonTS", repo_type="dataset")

# one long synthetic series (stitched from zEDGE windows by image quilting)
series = pd.read_parquet(hf_hub_download(filename="run_h200/series/bike/zedge_quilting/seed0.parquet", **repo))

# one pool of 300 generated windows (not stitched), back to real units
row = pd.read_parquet(hf_hub_download(filename="run_h200/pools/bike/tsgm_vae/seed0.parquet", **repo)).iloc[0]
pool = np.frombuffer(row["data"], dtype=row["dtype"]).reshape(json.loads(row["shape"]))      # (300, 168, 3)
ds = pd.read_parquet(hf_hub_download(filename="run_h200/tables/datasets.parquet", **repo)).set_index("name")
windows = ds.at["bike", "lo"] + pool[:, :, 0] * (ds.at["bike", "hi"] - ds.at["bike", "lo"])  # (300, 168)

Every synthetic series has its fidelity metrics in fidelity/ and its utility in utility/ and tables/utility.parquet, so a series can be chosen by quality. All series of one folder: snapshot_download(**repo, allow_patterns=["run_h200/series/**"], local_dir="marathonts").

Two runs

	run_h200/ — main run	run_l4/ — replication run
Hardware	1× NVIDIA H200, computing infrastructure of darya.ai	1× NVIDIA L4 (Google Colab)
Methods	30 (24 ranked): the min suite on the VAE pool and on the zEDGE pool, Diffusion-TS, Chronos, TimesFM, Timestral and zEDGE_TS	16 (13 ranked): the min suite on the VAE pool and the four open long-series generators
Role	every result of the paper and the leaderboard	reproduction check on other hardware

Both runs use profile min of the same code, configuration, datasets, task thresholds and seeds. Ranks are computed within one run, and times measured on different hardware are not compared. On the 13 methods common to both runs, the orders agree with Kendall's τ = 0.64, every mean rank of one run lies inside the 95 % interval of the other (12 of 13 in the other direction), and the top tier is the same up to one borderline method; the places within the tier change. Sources of the differences: the VAE and Diffusion-TS are retrained on another GPU and software stack (the main source), Chronos samples other trajectories with the same seeds, the downstream models differ slightly even on identical series, and the prepared ETT series of the main run differs slightly at the value level from that of the replication run (the source today gives the latter; the cause is not established). Read the leaderboard by tier, not by place.

Overall leaderboard of the main run (all 24 ranked methods)

* proprietary (zypl.ai), evaluated as a black box: results published, code not released.

place	method	family	mean rank	95 % interval	P(first)	top tier	datasets
1	zedge_quilting *	assembly	8.27	7.35–9.20	0.85	yes	9
2	zedge_zseries *	assembly	8.79	8.28–9.47	0.08	yes	9
3	zedge_greedy *	assembly	8.95	7.97–10.00	0.01	yes	9
4	zedge_pctsp *	assembly	9.18	8.13–10.25	0.00	yes	9
5	zedge_knn *	assembly	9.33	8.32–10.48	0.00	yes	9
6	zedge_atsp *	assembly	9.39	7.83–11.10	0.03	yes	9
7	zedge_ts_ar *	long-series	9.40	7.99–10.91	0.02	yes	9
8	zedge_ts_long *	long-series	9.50	8.25–10.89	0.00	yes	9
9	zedge_herding *	assembly	10.48	9.15–11.88	0.00	yes	9
10	zedge_mbb *	assembly	10.58	9.10–12.25	0.00	yes	9
11	zedge_rand *	assembly (random)	10.75	9.12–12.49	0.00	yes	9
12	vae_herding	assembly	13.53	12.79–14.30	0.00	no	9
13	vae_zseries *	assembly	13.68	12.15–15.24	0.00	yes	9
14	diffusion_ts	long-series	13.78	11.98–15.39	0.00	no	9
15	timesfm	long-series	14.43	12.56–16.23	0.00	yes	6
16	vae_quilting	assembly	14.44	12.64–16.20	0.00	yes	9
17	vae_pctsp	assembly	15.27	13.80–16.67	0.00	no	9
18	vae_atsp	assembly	15.28	13.26–17.10	0.00	no	9
19	vae_greedy	assembly	15.57	13.90–17.19	0.00	no	9
20	chronos	long-series	15.78	14.04–17.45	0.00	yes	7
21	vae_knn	assembly	16.01	14.16–17.70	0.00	no	9
22	vae_mbb	assembly	16.30	14.24–18.20	0.00	no	9
23	vae_rand	assembly (random)	16.47	14.43–18.39	0.00	no	9
24	timestral	long-series	16.77	15.53–18.08	0.00	no	9

Contents

Each run folder is an artifacts folder of the benchmark (artifacts/<run>/ in the repository):

Folder	Content
tables/, reports/, overall/, leaderboard/, metric_values/, hypotheses/	collected tables, rankings, research questions, reports
fidelity/<dataset>/<method>/	fidelity metrics of each series
records/<dataset>/<method>/	time, memory, hardware and diagnostics of each series
utility/<dataset>/<method>/seed<k>/	metrics.parquet (utility metrics) and predictions.parquet (predictions on the real test rows)
labels/<dataset>/	real test labels and targets
series/<dataset>/<method>/, pools/<dataset>/	synthetic long series; pools of synthetic windows
generators/<dataset>/	weights of the VAE and Diffusion-TS

Beside the run folders, synthetic/series/ and synthetic/windows/ hold the same synthetic series and windows in long format, one file per run and dataset (configs synthetic_series and synthetic_windows).

Not included: the weights of the downstream models (models.parquet; retrained by the same code); for ETTh1 and PJM, everything derived from the data (labels, series, pools, generator weights, predictions — see the licences below; these are regenerated locally by the pipeline); the code and weights of the proprietary zseries, zEDGE and zEDGE_TS (their series, pools and metrics are included); the raw and prepared data, which the benchmark downloads from the original sources.

Download

from huggingface_hub import snapshot_download
snapshot_download("zyplai/MarathonTS", repo_type="dataset",
                  allow_patterns=["run_h200/tables/**", "run_h200/fidelity/**"], local_dir="artifacts")

The analysis needs only tables/: python -m marathonts.hypotheses.run --artifacts-dir artifacts/run_h200 (pip install marathonts).

Licences

Each dataset keeps its licence (LICENSES.md). In short: the metrics, tables and reports are published for all nine datasets; artifacts derived from the data follow the licence of their dataset (CC BY 4.0 for six datasets, CC BY-SA 4.0 for Bitcoin) and are withheld for ETTh1 (CC BY-ND 4.0) and PJM. The timesfm series are outputs of TimesFM 3.0, whose weights are licensed for non-commercial use only.

Citation

Please cite the paper and the original source of every dataset you use:

@misc{varshavskiy2026marathonts,
  title         = {{MarathonTS}: A Benchmark of Long Synthetic Time Series from Window Assembly and
                   Long-Series Generators},
  author        = {Varshavskiy, Ilyas and Shulgin, Sergey and Khalilbekov, Shuhrat and Boboeva, Bonu and
                   Azimi, Azizjon},
  year          = {2026},
  eprint        = {XXXX.XXXXX},
  archivePrefix = {arXiv}
}