BeTiSe — Benchmark Time Series Generator

A Python library for generating synthetic time series datasets with configurable statistical properties and rich metadata.

Installation

pip install betise

Quick Start

from betise import generate_dataframe, load_config

cfg = load_config(dataset={"base_series": "arma", "num_series": 5, "length_range": [300, 500]})
df, ctx = generate_dataframe(cfg)
print(df[["series_id", "time", "data", "primary_category"]].head())

Save to parquet and add feature overlays:

from betise import run, load_config

cfg = load_config(dataset={
    "base_series":  "ar",
    "num_series":   100,
    "length_range": [300, 700],
    "output_dir":   "output",
    "output_name":  "ar_trend.parquet",
    "features": {
        "linear_trend":       {"enabled": True, "direction": "upward"},
        "single_seasonality": {"enabled": True},
        "point_anomaly":      {"enabled": True, "is_spike": True},
    },
})
run(cfg)

Series Types

Feature Overlays

Multiple features can be stacked on top of any base type:

Published Dataset

A large-scale benchmark dataset (120,000 series, 23.8 GB) generated with BeTiSe is available on Zenodo:

• DOI: 10.5281/zenodo.18513505
• Conference: Submitted to ITISE 2026

Documentation & Examples

Full usage guide, config reference, and ready-to-run examples are on GitHub:
github.com/ismailguzel/betise

The paper appendix (metadata schema, generation templates, LLM prompts, and example figures) is available in APPENDIX.md.

Citation

@dataset{betise2026,
  author    = {Kerem R Gür, P. Cemre Yazıcı, Pelin Erkaya, Yağmur Türkmen, Berke Baytak, İsmail Güzel, Pınar Karagöz and Ceylan Yozgatlıgil},
  title     = {{BeTiSe: A Benchmark Time Series Dataset for Stationarity
                and Structural Analysis}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.18513505},
  url       = {https://doi.org/10.5281/zenodo.18513505}
}

Contact

İsmail Güzel — ismailgzel@gmail.com

License

MIT — see LICENSE.