characteristics

name	characteristics
description	Extract TFB-style statistical characteristics (trend strength, seasonality strength, stationarity) from a dataset. Use when the user wants to profile, describe, or quantify the trend/seasonality/stationarity of a time-series dataset before benchmarking.

When to use

Profile a dataset's statistical properties — how trended, how seasonal, how stationary — before choosing models, or to report dataset-level stats. Works with any dataset config the project supports (single-file, custom, presplit, pre-processed, traffic bundles).

Command

uv run python tool/dataset_characteristics.py \
    --config configs/datasets/etth1.toml --split train --per-channel

Outputs trend_strength and seasonality_strength (STL-style, 0–1) plus stationarity (ADF p-value via statsmodels) to work_dirs/<dataset>/characteristics_<split>.csv. --period N overrides the FFT-estimated seasonal period; omit --per-channel for dataset-level rows only.

Reference

Full flags and metric definitions: docs/en/dataset-characteristics.md.