Bayway/italian-open-timeseries

Italian Open Time Series

46 public Italian time series in 17 configs (2,260,034 rows), cleaned and ready for zero-shot forecasting: smog in Milan, bikes in Bologna (with the city's weather, air quality and car traffic), tourism, energy prices and demand, earthquakes, deaths, prices and industry, and Italian Wikipedia. Every series comes from an open source (CC0 or CC BY 4.0), keeps its gaps as NaN, and is updated weekly.

Italian data is almost absent from the corpora that time-series foundation models are trained and evaluated on (GIFT-Eval, LOTSA, Chronos, fev-bench). This dataset fills that gap and comes with fev task definitions (tasks.yaml) and results for seven foundation models, including Mistral's Timestral family, measured on the past and on data published after the models were released.

Versione italiana in fondo.

Configs

Config	What	Series	Frequency	From	To	Unit	Licence
milano_no2_hourly	NO2 at four Milan stations (ARPA Lombardia)	4	hourly	1980-09-18	2026-10-21 14:00	µg/m³	CC0 1.0
milano_pm10_daily	PM10 at three Milan stations (ARPA Lombardia)	3	daily	2001-11-01	2026-10-20	µg/m³	CC0 1.0
milano_temperature_hourly	Air temperature, Milano Lambrate (ARPA Lombardia)	1	hourly	2001-06-29	2026-10-21 16:00	°C	CC0 1.0
bologna_bikes_hourly	Bike counters in Bologna, monthly batches (Comune di Bologna)	3	hourly	2018-12-31	2026-09-30 21:00	bicycles/hour	CC BY 4.0
bologna_bikes_live_hourly	Bike counters in Bologna, updated daily (Comune di Bologna)	2	hourly	2024-11-26	2026-10-20 21:00	bicycles/hour	CC BY 4.0
italy_tourism_monthly	Tourist nights and arrivals in Italy (ISTAT)	2	monthly	1990-01-01	2026-08-01	arrivals, nights	CC BY 4.0
italy_tourism_regions_monthly	Tourist nights in ten Italian regions (ISTAT)	10	monthly	2016-01-01	2025-12-01	nights	CC BY 4.0
italy_energy_prices_monthly	Consumer prices of electricity and gas in Italy, HICP (Eurostat)	3	monthly	1996-01-01	2026-09-01	index (2025=100)	CC BY 4.0
italy_energy_demand_monthly	Electricity supplied and gas consumed in Italy (Eurostat)	2	monthly	2008-01-01	2026-09-01	GWh, TJ (GCV)	CC BY 4.0
italy_macro_monthly	Consumer prices (HICP) and industrial production, Italy (ISTAT)	2	monthly	1990-01-01	2026-09-01	index, 2021=100, index, 2025=100	CC BY 4.0
italy_deaths_weekly	Weekly deaths in Italy (Eurostat)	1	weekly	2011-01-03	2026-06-29	deaths/week	CC BY 4.0
italy_earthquakes_daily	Daily earthquakes of magnitude 2 or more in Italy (INGV)	1	daily	2012-01-01	2026-10-20	events/day	CC BY 4.0
itwiki_pageviews_daily	Daily user pageviews of Italian Wikipedia (Wikimedia)	1	daily	2015-07-01	2026-10-20	pageviews/day	CC0 1.0
bologna_weather_hourly ¹	Air temperature and rainfall, Bologna idrografico station (Arpae)	2	hourly	2018-12-31	2026-10-23 13:00	mm, °C	CC BY 4.0
bologna_no2_hourly ¹	NO2 at three Bologna stations (Comune di Bologna / Arpae)	3	hourly	2016-12-31	2026-10-22 21:00	µg/m³	CC BY 4.0
bologna_pm10_daily ¹	PM10 at three Bologna stations (Comune di Bologna / Arpae)	3	daily	2017-01-01	2026-10-22	µg/m³	CC BY 4.0
bologna_traffic_hourly ¹	Cars per hour on three Bologna streets with bike counters (Comune di Bologna)	3	hourly	2018-12-31	2026-09-30 21:00	vehicles/hour	CC BY 4.0

¹ Bologna city data (weather, air quality, car traffic), published as covariates for the Bologna bike counters. They have no fev tasks in v1.

Each parquet file is in long format: id (string), timestamp (datetime, tz-naive), target (float, NaN = missing). Every series is on a regular grid from its first to its last observed value; nothing is interpolated.

• Timestamps. Hourly series are UTC, labelled at the start of the hour. Daily series use the Italian calendar date (Wikipedia: the UTC day). Weekly series use the Monday of the ISO week, monthly series the first day of the month.
• Sources and cleaning. Per-series source datasets, units, cleaning steps and update lags are in catalog.json. Known issues are listed below and in DATA_QUALITY.md.

from datasets import load_dataset

no2 = load_dataset("Bayway/italian-open-timeseries", "milano_no2_hourly", split="train").to_pandas()
no2.pivot(index="timestamp", columns="id", values="target").tail()

Evaluation protocol

tasks.yaml defines 19 fev tasks: 13 backtests and 6 live tasks.

• Backtest. Rolling windows that all end on 14 October 2026, the day Timestral was released, or at the config's last published value if earlier. A model may have seen some of this history during training.
• Live. Windows that start on 15 October 2026 (Italian time) and are scored on values published after every model in the comparison was released. They grow week by week as new data is published. Monthly and weekly series join once post-release months are published.
• Metrics. Scaled quantile loss (SQL, main metric), MASE, WQL and WAPE over the deciles 0.1–0.9, as in fev-bench. Leaderboards report the skill score against seasonal naive: 100 × (1 − geometric mean of the error ratio). 0 means as good as seasonal naive, 100 means perfect.

Task	Horizon	Windows	Step	First cut-off	Seasonality	Max context
milano_no2_hourly/backtest	24	20	168	2026-06-02 21:00:00	24	8192
milano_no2_hourly/live	144	1	144	2026-10-14 21:00:00	24	8192
milano_pm10_daily/backtest	7	20	14	2026-01-14	7	all
milano_pm10_daily/live	5	1	5	2026-10-14	7	all
milano_temperature_hourly/backtest	24	20	168	2026-06-02 21:00:00	24	8192
milano_temperature_hourly/live	144	1	144	2026-10-14 21:00:00	24	8192
bologna_bikes_hourly/backtest	24	20	168	2026-05-19 21:00:00	24	8192
bologna_bikes_live_hourly/backtest	24	20	168	2026-06-02 21:00:00	24	8192
bologna_bikes_live_hourly/live	144	1	144	2026-10-14 21:00:00	24	8192
italy_tourism_monthly/backtest	12	12	1	2024-09-01	12	all
italy_tourism_regions_monthly/backtest	12	12	1	2024-01-01	12	all
italy_energy_prices_monthly/backtest	12	12	1	2024-10-01	12	all
italy_energy_demand_monthly/backtest	12	12	1	2024-09-01	12	all
italy_macro_monthly/backtest	12	12	1	2024-09-01	12	all
italy_deaths_weekly/backtest	8	10	4	2025-08-25	52	all
italy_earthquakes_daily/backtest	14	12	14	2026-04-29	7	all
italy_earthquakes_daily/live	5	1	5	2026-10-14	7	all
itwiki_pageviews_daily/backtest	28	12	14	2026-04-15	7	all
itwiki_pageviews_daily/live	5	1	5	2026-10-14	7	all

import fev

benchmark = fev.Benchmark.from_yaml(
    "https://huggingface.co/datasets/Bayway/italian-open-timeseries/resolve/main/tasks.yaml"
)
for task in benchmark.tasks:
    predictions = []
    for window in task.iter_windows():
        past, future_known = window.get_input_data()
        predictions.append(my_model(past, task.horizon))  # "predictions" + one column per quantile
    print(task.evaluation_summary(predictions, model_name="my-model"))

Results with time-series foundation models

We ran seven zero-shot time-series foundation models and two classical baselines on every task:

Model	By	Parameters	Revision	Weights licence
Timestral-5M	Mistral AI	5M	bdd20cf	Modified MIT (<$20M/month revenue)
Timestral-28M	Mistral AI	28M	49b2640	Modified MIT (<$20M/month revenue)
Timestral-123M	Mistral AI	123M	5a14728	Modified MIT (<$20M/month revenue)
Timestral-209M	Mistral AI	209M	9874d25	Modified MIT (<$20M/month revenue)
Chronos-2	Amazon	120M	29ec376	Apache-2.0
TimesFM 3.0	Google	~330M	43046b8	TimesFM Non-Commercial v1.0
Toto 2.0 (313M)	Datadog	313M	a7bab28	Apache-2.0

Plus seasonal naive and AutoETS (statsforecast). No model was fine-tuned or tuned. Every model sees the same context, capped at 8,192 steps for hourly series. Timestral and Chronos-2 receive the gaps as NaN; for the other models gaps are forward-filled. Everything ran on an Apple M2 Pro (CPU and MPS). Model weights are not redistributed here; each keeps its own licence.

Backtest (13 tasks; skill against seasonal naive, higher is better):

Model	SQL skill	SQL win rate	MASE skill	MASE win rate
Toto 2.0 (313M)	28.4	0.69	21.8	0.66
Timestral-209M	28.2	0.71	21.4	0.68
Timestral-123M	27.9	0.65	21.3	0.64
Timestral-28M	27.0	0.58	20.1	0.56
Chronos-2	26.6	0.55	19.7	0.56
TimesFM 3.0	26.2	0.61	18.5	0.57
Timestral-5M	23.2	0.40	14.6	0.42
Seasonal naive	0.0	0.12	0.0	0.24
AutoETS	-7.5	0.19	-14.3	0.16

Live, on data published after 14 October 2026 (6 tasks, one window of 5–6 days each):

Model	SQL skill	SQL win rate	MASE skill	MASE win rate
Timestral-209M	32.2	0.88	28.9	0.83
Timestral-123M	30.9	0.85	27.4	0.71
TimesFM 3.0	30.3	0.69	26.6	0.73
Toto 2.0 (313M)	27.7	0.52	24.6	0.50
Timestral-28M	25.3	0.56	21.4	0.58
Chronos-2	23.9	0.40	19.8	0.40
Timestral-5M	18.5	0.31	14.4	0.31
Seasonal naive	0.0	0.08	0.0	0.12
AutoETS	-90.4	0.21	-16.8	0.31

The live round is small: a single window per task, scored on less than a week of data. Treat it as a first signal; it is re-run every week as new data arrives.

SQL per task (lower is better; best in bold):

Task	Timestral-5M	Timestral-28M	Timestral-123M	Timestral-209M	Chronos-2	TimesFM 3.0	Toto 2.0 (313M)	Seasonal naive	AutoETS
bologna_bikes_hourly/backtest	0.521	0.469	0.463	0.457	0.446	0.452	0.451	0.718	0.802
bologna_bikes_live_hourly/backtest	0.621	0.579	0.567	0.560	0.566	0.548	0.554	0.740	0.748
bologna_bikes_live_hourly/live	0.532	0.499	0.463	0.463	0.471	0.470	0.489	1.311	0.939
italy_deaths_weekly/backtest	0.503	0.449	0.450	0.453	0.491	0.494	0.417	0.542	0.903
italy_earthquakes_daily/backtest	0.264	0.263	0.264	0.263	0.265	0.264	0.263	0.443	0.321
italy_earthquakes_daily/live	0.451	0.448	0.447	0.449	0.454	0.451	0.452	0.501	0.456
italy_energy_demand_monthly/backtest	0.489	0.498	0.493	0.506	0.473	0.479	0.466	0.634	0.460
italy_energy_prices_monthly/backtest	0.746	0.771	0.805	0.775	0.712	0.767	0.824	0.793	0.696
italy_macro_monthly/backtest	0.331	0.284	0.274	0.275	0.295	0.310	0.241	0.599	0.423
italy_tourism_monthly/backtest	1.503	1.405	1.434	1.403	1.420	1.302	1.275	1.613	2.048
italy_tourism_regions_monthly/backtest	0.818	0.711	0.654	0.665	0.679	0.846	0.759	0.736	0.858
itwiki_pageviews_daily/backtest	0.452	0.439	0.429	0.420	0.498	0.435	0.477	0.732	0.690
itwiki_pageviews_daily/live	0.574	0.455	0.353	0.363	0.486	0.405	0.377	0.614	0.306
milano_no2_hourly/backtest	0.447	0.441	0.438	0.435	0.455	0.442	0.450	0.598	1.382
milano_no2_hourly/live	0.919	0.855	0.871	0.861	0.928	0.877	0.888	1.174	77.234
milano_pm10_daily/backtest	0.305	0.305	0.301	0.302	0.305	0.304	0.294	0.495	0.415
milano_pm10_daily/live	1.051	1.060	1.006	0.974	1.027	0.992	1.038	1.063	1.066
milano_temperature_hourly/backtest	0.752	0.736	0.719	0.722	0.758	0.746	0.766	0.923	1.416
milano_temperature_hourly/live	0.867	0.740	0.670	0.602	0.774	0.605	0.731	0.781	1.735

A Bologna-focused page with these data and forecasts: Due Torri, Due Ruote.

What stands out in this first round:

• All seven foundation models beat seasonal naive clearly on the backtest (skill 23–28). AutoETS does worse than seasonal naive overall.
• Toto 2.0 and Timestral-209M are level on the backtest (28.4 and 28.2). On the first live week, Timestral-209M leads (32.3), ahead of Timestral-123M (30.9) and TimesFM 3.0 (30.3).
• Size matters for Timestral: the 5M model is clearly behind the 123M and 209M models.
• Tourism is hard for everyone. Italian tourist nights jump by about 15% from late 2024 (see known issues), which no model can anticipate.
• Daily PM10 is the hardest live task: the best model is only about 8% better than seasonal naive. Daily smog is driven by weather that these models do not see.
• On monthly consumer prices of electricity and gas, classical AutoETS beats every foundation model.

An extra experiment (Timestral-209M, same backtest windows): what happens if the model also gets the Italian calendar (day of week, and day type with national holidays, the city's patron saint and ponti) and the weather, either as observed up to the forecast time (past) or as observed over the forecast horizon too (perfect forecast: an upper bound for what a weather forecast could add). Change in SQL against the model alone (negative is better):

Series	+ calendar	+ weather (past)	+ weather (perfect forecast)	+ calendar + weather (perfect forecast)
Bikes, via Sabotino (Bologna); weather = temperature and rain in Bologna	-3.8%	-2.1%	-17.3%	-22.3%
NO2, via Marche (Milan); weather = temperature at Milano Lambrate	-0.1%	-0.7%	-9.3%	-9.0%

Known issues

• Tourism break. ISTAT tourist nights grow 10–25% year on year from late 2024 through 2025 (+15% nationally, +39% Sardinia, +35% Sicily, +1% Veneto). This looks like a change in coverage, not demand, but ISTAT's methodology pages could not be checked at build time.
• Regional tourism is history only (2016–2025): ISTAT publishes it once a year.
• Provisional current-year ARPA data. Values for the current year are not yet validated by ARPA (validation ends by 30 March of the following year) and may be revised. The 2026 temperature at Lambrate has a different daily shape from validated years, either because it is unvalidated or because of this year's weather.
• Cleaning beyond the source flags. NO2 runs of 3 or more zero hours (stuck analysers), PM10 zero days, bike outages of 24 or more zero hours, and two −42 °C temperature readings are set to missing. See catalog.json for the counts.
• Bike counters. Via Sabotino, Ercolani and San Donato arrive in monthly batches; via Murri and Mazzini are updated daily and carry the live evaluation.
• Earthquakes are counts of events with magnitude ≥ 2 in INGV's catalogue (latitude 35–48 N, longitude 6–19 E), from 2012, when catalogue completeness becomes stable.

Leakage

The released models do not document their real training data. Timestral was pre-trained on synthetic data and post-trained on undisclosed real and synthetic data. Italian series are almost absent from the public corpora we checked. Exceptions are national macro data, which is mirrored widely, and Italian Wikipedia pageviews before December 2023, which are in TimesFM's training data. Only the live tasks are guaranteed unseen. Per-series notes are in catalog.json.

Sources, licences and attribution

The dataset as a whole is distributed under CC BY 4.0. Series from CC0 sources (ARPA Lombardia, Wikimedia) can be reused without conditions.

• ARPA Lombardia (CC0 1.0): Dati ARPA Lombardia, dati.lombardia.it (CC0 1.0). https://www.dati.lombardia.it
• Comune di Bologna (CC BY 4.0): Fonte: Comune di Bologna, opendata.comune.bologna.it, dataset colonnine-conta-bici (CC BY 4.0). https://opendata.comune.bologna.it
• ISTAT (CC BY 4.0): Source: Istat, Italian National Institute of Statistics, esploradati.istat.it (CC BY 4.0). https://esploradati.istat.it
• Eurostat (CC BY 4.0): Source: Eurostat (CC BY 4.0). Eurostat bears no responsibility for the modifications. https://ec.europa.eu/eurostat
• INGV (CC BY 4.0): INGV, Istituto Nazionale di Geofisica e Vulcanologia, earthquake catalogue (CC BY 4.0). https://terremoti.ingv.it
• Wikimedia (CC0 1.0): Wikimedia Foundation, Pageviews API (CC0 1.0). https://wikimedia.org/api/rest_v1/
• Arpae (CC BY 4.0): Fonte: Arpae Emilia-Romagna, https://www.arpae.it/ (CC BY 4.0, modificato). Air quality via Comune di Bologna open data. https://www.arpae.it

Changes made: reshaped to long format; timestamps converted to UTC (hourly) or calendar dates; values flagged invalid by the source and the sentinel values removed; the cleaning steps listed under known issues; bike directions summed; earthquakes aggregated to daily counts. No values were interpolated or adjusted.

The Italian wholesale electricity price (GME) and Terna's load data are not included: their terms do not allow redistribution.

Updates and versions

The data is rebuilt weekly from the sources, and the live tasks are re-scored. Each weekly release is a commit of this repository; the first build is tagged v1.0 (data as of 21 October 2026). The code that builds the data and runs the evaluation will be published alongside.

Citation

@misc{italian_open_timeseries_2026,
  title  = {Italian Open Time Series},
  author = {Fiori, Massimiliano},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Bayway/italian-open-timeseries}
}

Built with the help of Claude Code. Thanks to ARPA Lombardia, Comune di Bologna, ISTAT, Eurostat, INGV and Wikimedia for publishing open data.

---

In italiano

Serie temporali italiane aperte: 46 serie pubbliche italiane, pulite e pronte per la previsione. Ci sono lo smog a Milano (NO2 e PM10, ARPA Lombardia), le bici a Bologna (colonnine conta-bici del Comune, con meteo, qualità dell'aria e traffico delle auto), il turismo (ISTAT, nazionale e per regione), i prezzi e i consumi di luce e gas (Eurostat), i terremoti (INGV), i decessi settimanali, l'inflazione, la produzione industriale e le visite a Wikipedia in italiano. Tutti i dati vengono da fonti aperte (CC0 o CC BY 4.0) e sono aggiornati ogni settimana.

I modelli di AI per le serie temporali non hanno quasi mai visto dati italiani. Qui li mettiamo alla prova: Timestral di Mistral (quattro taglie), Chronos-2 di Amazon, TimesFM 3.0 di Google e Toto 2.0 di Datadog, sia sul passato sia sui dati pubblicati dopo la loro uscita (dal 15 ottobre 2026), che nessun modello può aver visto. Le classifiche sopra si aggiornano ogni settimana.

Il prezzo all'ingrosso dell'energia (GME) e i consumi di Terna non sono inclusi: le loro condizioni d'uso non permettono di ripubblicarli.