Skip to content

Datasets for the experiments

Open datasets on which the methods from BASELINES.md are compared by the metrics from METRICS.md. Licences are listed in DATASET_LICENSES.md.

Principle: the same datasets for all tasks

All utility tasks (U2.1–U2.9) are built from a single target series of a dataset, so every dataset below is used in all 14 tasks (nine task groups U2.1–U2.9) — there is one set of datasets for the whole experiment. To this end, for all datasets:

  • a single target column is chosen (channel 0 of the pipeline);
  • the step is brought to one hour (aggregation by mean or sum), the calendar is the hour of week (BENCHMARK.md, Protocol: data);
  • the length after aggregation is ≥ 17,000 hours: with a 60 / 20 / 20 split (train / val / OOT, profile max, the full series) and T = 168 this gives K ≥ 60 windows, enough for h = 168 and for events with q = 0.99. Profiles min and mid take the last 10,000 hourly points of every dataset with a 50 / 25 / 25 split: 9,664 points after the embargoes, train 5,000, val = OOT = 2,332, K = 30 (artifacts/run_h200/tables/datasets.parquet).

Tasks (from METRICS.md, U2)

ID Task Type
U2.1 Upper tail: change over h = 24 above the 0.90 quantile binary classification
U2.2 The same at horizons h ∈ binary classification
U2.3 Rare events: q ∈ binary classification (strong imbalance)
U2.4 Lower tail: change below the 0.10 quantile binary classification
U2.5 Down / neutral / up multiclass classification (3 classes)
U2.6 Value x[t+h], h ∈ regression
U2.7 Quantiles 0.1 / 0.5 / 0.9 at horizon h = 24 quantile regression
U2.8 Vector x[t+1 : t+24] multi-step regression
U2.9 std over the next 24 hours regression (volatility)

Every dataset below is used in all tasks U2.1–U2.9.

Summary table

Columns — in the source file. Hours — length of the target series after conversion to an hourly step. K — approximate number of assembled windows with train = 60% and T = 168 (profile max; min and mid: K = 30 for every dataset).

# Dataset File Columns Rows in file Source step Hours K Target column Tasks
1 ETTh1 ETTh1.csv 8 17,420 1 h 17,420 ≈ 62 OT U2.1–U2.9
2 Electricity Load Diagrams LD2011_2014.txt 371 140,256 15 min 35,065 ≈ 125 sum over 370 clients U2.1–U2.9
3 Household Electric Power Consumption household_power_consumption.txt 9 2,075,259 1 min ≈ 34,590 ≈ 124 Global_active_power U2.1–U2.9
4 PJM Hourly Energy Consumption (PJME) PJME_hourly.csv 2 145,366 1 h 145,392 ≈ 519 PJME_MW U2.1–U2.9
5 Beijing PM2.5 PRSA_data_2010.1.1-2014.12.31.csv 13 43,824 1 h 43,800 ≈ 156 pm2.5 U2.1–U2.9
6 Jena Climate 2009–2016 jena_climate_2009_2016.csv 15 420,551 10 min ≈ 70,000 ≈ 250 T (degC) U2.1–U2.9
7 Metro Interstate Traffic Volume Metro_Interstate_Traffic_Volume.csv 9 48,204 1 h (with gaps) ≈ 29,000 (after skipping 2014–2015) ≈ 103 traffic_volume U2.1–U2.9
8 Bike Sharing (hourly) hour.csv 17 17,379 1 h 17,544 by calendar ≈ 63 cnt U2.1–U2.9
9 Bitcoin Historical Data btcusd_1-min_data.csv 6 7,767,566 1 min ≈ 85,000 (2017-01 – 2026-10) ≈ 305 Close (log price) U2.1–U2.9

Columns and rows in the files have been checked against the downloaded data (data/raw/, October 2026; Bitcoin is updated daily, so the number of rows grows). Hours runs from the first to the last valid hourly value, gaps included. Values marked ≈ are rounded.

The files are stored in data/raw/<dataset>/ as Parquet (zstd, lossless: all source columns, float64): ett/, electricity/, household/, pjm/ (together with the other PJM regions), beijing_pm25/, jena/, metro/, bike/ (together with day.parquet and Readme.txt), bitcoin/. In total ≈ 200 MB versus 1.3 GB of source CSV. The data are not included in the repository: the stage python -m marathonts.bench.run raw downloads the source files (Kaggle — with ~/.kaggle/kaggle.json) and converts them to Parquet. Column names in the table above are the original ones; the number of rows matches the source files.

Descriptions

1. ETTh1 — Electricity Transformer Temperature

Loads and oil temperature of a power transformer in China, July 2016 – June 2018, 1-hour step. Columns: date, six loads (HUFL, HULL, MUFL, MULL, LUFL, LULL) and the oil temperature OT. A standard long-horizon forecasting benchmark (Informer, Autoformer, etc.), with no missing values. The shortest in the set — the lower bound on length.

  • Source: github.com/zhouhaoyi/ETDataset, file ETT-small/ETTh1.csv
  • Paper: Zhou H. et al. Informer: Beyond efficient transformer for long sequence time-series forecasting. AAAI 2021

2. Electricity Load Diagrams 2011–2014

Electricity consumption of 370 clients in Portugal, 2011–2014, kW per 15 minutes. The first column is the time, the rest are clients; separator ;, decimal comma. Clients connected later have zeros at the start. The target series is the total load aggregated to hours: strong daily and weekly seasonality.

  • Source: UCI #321

  • Author: Trindade A., 2015

  • Leakage for Chronos: the Chronos training corpus contains this dataset (Electricity 15 min, hourly and weekly, all 370 clients, without only the last 24 points; Ansari et al., Chronos: Learning the Language of Time Series, TMLR 2024, Table 3). The benchmark's target is the sum of the clients, so Chronos has seen the val and OOT periods of this series in aggregated form; its results on electricity are not out-of-sample.

3. Individual Household Electric Power Consumption

Consumption of a single household in France, December 2006 – November 2010, 1-minute step. Columns: Date, Time, active and reactive power, voltage, current, three sub-meters. About 1.25% of the rows are missing (marked ?). The target series is the hourly mean of Global_active_power. A noisy series with sharp peaks — a test on irregular dynamics.

  • Source: UCI #235
  • Authors: Hébrail G., Bérard A., 2012
  • Leakage for TimesFM: the TimesFM 3.0 training data (GiftEvalPretrain) include this dataset as sceaux (from BuildingsBench): one hourly series, 2007-01-01 – 2010-11-26, equal to the hourly mean of Global_active_power on every common hour after a 1-hour label shift. TimesFM is left out on household (Config.leaked).

4. PJM Hourly Energy Consumption (PJME)

Hourly consumption of the PJM East region (USA), 2002–2018, megawatts. The set also contains ~12 files for other PJM regions in the same format (Datetime, <REGION>_MW). The longest series in the set (largest K). There are isolated duplicates and gaps at the transitions to daylight saving time.

5. Beijing PM2.5

PM2.5 concentration at the US Embassy in Beijing and airport weather, 2010–2014, 1-hour step. Columns: No, year, month, day, hour, pm2.5, DEWP, TEMP, PRES, cbwd, Iws, Is, Ir. pm2.5 has about 2000 missing values. A heavy right tail and smog episodes (tail tasks U2.1–U2.4).

  • Source: UCI #381
  • Paper: Liang X. et al. Assessing Beijing's PM2.5 pollution: severity, weather impact, APEC and winter heating. Proc. Royal Society A 471, 2015
  • Leakage for TimesFM: the TimesFM 3.0 training data (GiftEvalPretrain, built on LOTSA) include "Beijing Air Quality" (UCI #501: PM2.5 and weather of 12 Beijing stations, March 2013 – February 2017). It is another file, but the same city, pollutant and hours as this series over 2013–2014 — the period of the min window. TimesFM is left out on beijing_pm25 (Config.leaked).
  • Leakage for Chronos: the Chronos training corpus includes KDD Cup 2018 (air quality of Beijing and London stations, January 2017 – March 2018; Benchmark I of Ansari et al., 2024). Same city and pollutant, other stations and years: a related pair, so Chronos is left out on beijing_pm25 too.

6. Jena Climate 2009–2016

Weather station of the Max Planck Institute for Biogeochemistry (Jena), 2009–2016, 10-minute step, 14 variables: pressure, temperature, humidity, wind, etc. The target series is temperature averaged to hours. Strong daily and yearly seasonality with weak weekly seasonality — a check that the "hour of week" calendar does not impose spurious structure.

7. Metro Interstate Traffic Volume

Hourly westbound traffic on I-94 (Minneapolis – St Paul), October 2012 – September 2018, plus weather and holidays. Columns: holiday, temp, rain_1h, snow_1h, clouds_all, weather_main, weather_description, date_time, traffic_volume. There are duplicate hours (several weather descriptions) and a large gap in 2014–2015. The duplicate hours are averaged, and the series starts after the last gap longer than 30 days (June 2015). Very strong weekly structure.

  • Source: UCI #492
  • Author: Hogue J., 2019

8. Bike Sharing (hourly)

Hourly number of Capital Bikeshare bicycle rentals (Washington, D.C.), 2011–2012, with weather and calendar. Columns of hour.csv: instant, dteday, season, yr, mnth, hr, holiday, weekday, workingday, weathersit, temp, atemp, hum, windspeed, casual, registered, cnt. Several hours are missing (zero demand) and are set to 0. Count data with a trend between years.

  • Source: UCI #275
  • Paper: Fanaee-T H., Gama J. Event labeling combining ensemble detectors and background knowledge. Progress in Artificial Intelligence 2(2–3), 113–127, 2014

9. Bitcoin Historical Data

Minute-level OHLCV data for BTC/USD (Bitstamp), from 2012 to the present. Columns: Timestamp, Open, High, Low, Close, Volume. Before 2017 many minutes have no trades; the series is taken from bitcoin_start = 2017-01-01 to bitcoin_end = 2026-10-01 (exclusive). The target series is the log of the hourly close price. A financial series: weak seasonality, volatility clustering, heavy-tailed increments.

  • Source: kaggle.com/datasets/mczielinski/bitcoin-historical-data
  • Author: Mark Zielinski (Kaggle display name "Zielak"; the full name is that of the GitHub account mczielinski linked from the dataset description)
  • Leakage for TimesFM: GiftEvalPretrain includes the Monash "Bitcoin" set (18 daily series, among them the Bitcoin price, 2009 – mid-2021). The same asset at a daily resolution: a related pair, so TimesFM is left out on bitcoin in every profile — also in min, whose window (2025–2026) lies after it (Config.leaked).

Leakage into pretrained models

A dataset whose object is in a pretrained model's training data is not out-of-sample for it: the method is left out on that dataset in every ranking, question and case-study test (Config.leaked, bench.config.LEAKED); the other methods keep every dataset. The analysis always applies the list of the current code (not the copy saved with the tables), so a pair added later takes effect on the next run of notebook 7 without recomputing anything; --leaked replaces it from the command line.

Rule. Left out are both - exact pairs: the same data, the same source series, in the training corpus; - related pairs: the same object — the same dataset, the same city and pollutant, the same asset — at other stations, years or resolution, whether or not the periods overlap with the benchmark window.

Kept are pairs that share only the phenomenon (bike rentals in another city, gridded weather reanalysis) and datasets that a model uses only for evaluation.

The audit (October 2026) checked every model × dataset pair against the primary sources: Chronos — the paper's Appendix B, Table 3 (pretraining = pretraining-only sets + Benchmark I; Benchmark II is zero-shot evaluation only); TimesFM 3.0 — the model card and the full list of GiftEvalPretrain sub-datasets (Hugging Face Salesforce/GiftEvalPretrain, LOTSA builders of uni2ts), with the candidate sub-datasets downloaded and compared with the benchmark's raw files; Timestral — the model card and the release note.

Model Training data (source) Left out on
Chronos T5 (small; base in max, same training corpus) Ansari et al., TMLR 2024, Appendix B, Table 3 electricity (exact), beijing_pm25 (related: KDD Cup 2018)
TimesFM 3.0 model card: GiftEvalPretrain (without GIFT-Eval test sets and fev-bench overlaps), Wikipedia Pageviews, Google Trends, synthetic household (exact: sceaux), beijing_pm25 (related: Beijing Air Quality), bitcoin (related: Monash Bitcoin)
Timestral 209M technical report (Petralia et al., 2026): pretraining on synthetic data only; post-training on a mix of real and synthetic data, the real data not listed — unknown, cannot be checked (kept)
Diffusion-TS, VAE trained by the benchmark on train only —

Every pair of the audit (exact / related / weak / none / unknown):

Dataset Chronos T5 TimesFM 3.0 Timestral 209M
ETTh1 none: zero-shot evaluation only (Benchmark II) none: GIFT-Eval test set, not in GiftEvalPretrain unknown
electricity exact: Electricity 15 min / hourly / weekly in Benchmark I none: GIFT-Eval test set unknown
household none (London Smart Meters only: other households) exact: sceaux = UCI #235, hourly 2007-01 – 2010-11, equal to the hourly mean of Global_active_power on every common hour (after a 1-hour label shift) unknown
pjm none none found: no GiftEvalPretrain series equals any of the 11 PJM regions; pdb (public Kaggle set of undocumented origin, 2013–2014) correlates with PJME (r = 0.945) — origin unresolved unknown
beijing_pm25 related: KDD Cup 2018 (Beijing and London stations, 2017-01 – 2018-03) related: Beijing Air Quality (UCI #501, 12 stations, 2013-03 – 2017-02; station correlations with our series 0.81–0.96) unknown
jena weak: WeatherBench gridded reanalysis only weak: gridded ERA5 / CMIP6 only (jena_weather is a GIFT-Eval test set) unknown
metro none (traffic sets are PeMS, evaluation only) none (California and China traffic sets) unknown
bike weak: Mexico City Bikes (another city) none unknown
bitcoin none related: Monash Bitcoin, daily price, 2009 – mid-2021 unknown

Timestral declares no GIFT-Eval test-data leakage on the GIFT-Eval leaderboard (zero-shot), which weakly covers ETTh1, electricity and Jena; the other pairs cannot be checked. Earlier TimesFM versions differ (1.0 / 2.0 also trained on Electricity and the Jena weather data): the benchmark uses 3.0 only.

How the datasets differ (why there are nine)

Property Datasets
Strong weekly seasonality Electricity, PJME, Metro, Bike
Strong daily, weak weekly Jena, ETTh1
Heavy tails, spikes PM2.5, Household, Bitcoin
Volatility clustering, weak seasonality Bitcoin
Trend / level drift Bike, ETTh1, Bitcoin
Gaps and duplicates requiring preprocessing Household, PM2.5, Metro, Bike, PJME
Short (K ≈ 60) ETTh1, Bike
Long (K ≥ 250) PJME, Jena, Bitcoin