itsalissonsilva/probabilistic-timeseries-dgps

Probabilistic Time-Series DGPs

A deterministic benchmark of 20,000 forecasting tasks generated from ten known data-generating processes. Each row contains a 192-step history, a 64-step future, the noise-free latent signal, the future noise scale, and the exact simulation parameters.

The dataset is intended for probabilistic forecasting, calibration, distribution-shift testing, and faithful reproduction of TCN/ModernTCN/transformer baselines. It is not generated by an LLM and does not contain scraped or personal data.

Families

1. Linear trend
2. Seasonal
3. Trend plus seasonality
4. Multiple seasonalities
5. Structural changepoint
6. Markov regime switching
7. Intermittent demand
8. Count process with time-varying Poisson rate
9. Heteroskedastic trend
10. Damped oscillation

Every family contributes 2,000 rows: 1,600 train, 200 validation and 200 test. The second half of each family’s test partition uses extrapolated parameter ranges and is marked with is_distribution_shift=true.

Row schema

Field	Type	Meaning
series_id	string	Stable unique identifier
group_id	string	Leakage-control group identifier
split	string	train, validation, or test
family	string	Data-generating-process family
difficulty	string	Noise-to-signal heuristic: low, medium, or high
is_distribution_shift	bool	Whether parameters come from the shifted test range
seed	int	Per-row deterministic seed
history	list[float32]	192 observed context values
future	list[float32]	64 observed target values
latent_history	list[float32]	Noise-free context signal or conditional mean
latent_future	list[float32]	Noise-free future signal or conditional mean
future_noise_scale	list[float32]	Known conditional noise standard deviation or proxy
parameters_json	string	Exact DGP parameters, serialized deterministically
primary_period	int	Main period, or -1 when not applicable
generator_version	string	Generator semantic version

Usage

from datasets import load_dataset

dataset = load_dataset("itsalissonsilva/probabilistic-timeseries-dgps")
row = dataset["train"][0]
history = row["history"]
target = row["future"]

The latent fields and generator parameters are provided for diagnostic evaluation. Do not pass them to a forecasting model when reporting ordinary history-to-future performance.

Evaluation

Report results separately for:

• validation versus test;
• in-distribution versus shifted test rows;
• each DGP family;
• observed targets versus latent conditional means.

At minimum report MAE or RMSE for point forecasts and pinball loss/CRPS plus empirical coverage for probabilistic forecasts. baselines.json contains last-value and history-mean point baselines by split and family.

Reproducibility and validation

Run:

python generate.py --per-family 2000

The generator validates row counts, identifiers, split/family coverage, array lengths, nulls and finite values. release_manifest.json records the generator version, global seed, split counts, file sizes and SHA-256 checksums.

Limitations

• These simulations are deliberately simpler than real operational time series.
• DGP parameters are sampled independently across rows and do not reproduce a real panel hierarchy.
• future_noise_scale is exact for the Gaussian/count families and a documented variance proxy for intermittent demand.
• Strong performance here does not establish performance on finance, medicine, web traffic or any other real domain.

License and citation

Dataset contents are released under CC BY 4.0. The generator code is released under Apache-2.0. See CITATION.cff and LICENSE.md.