v0.3.0

Highlights

corrupt() High-Level API

Accepts DataFrame, 2D array, or 1D array directly — handles standardization internally.

from tsrbench import CollectiveNoise
import pandas as pd

df = pd.read_csv("my_data.csv")
cn = CollectiveNoise(seed=2025)
results = cn.corrupt(df, noise_level=3, skip_first_col=True)
# results['shift'], results['spike'], results['impulse'],
# results['gaussian'], results['missing'], results['combined']

• skip_first_col=True — skip date/index column in DataFrames
• zero_clip=True — clip negatives to zero (e.g., electricity data)

6 Corruption Types

• Level Shift — sustained deviation over multiple time steps
• Exponential Spike — transient peak with exponential rise/fall
• Impulse — single-point EVT-calibrated spike
• Gaussian — additive N(0, σ) noise on every time step
• Missing — zero-valued dropout at independent positions
• Combined — union of Shift + Spike + Impulse

All anomaly positions are non-overlapping within a severity level (priority: Shift → Spike → Impulse → Missing).

Quickstart Notebook

examples/quickstart.ipynb — interactive examples for 1D, 2D, and DataFrame inputs with visualizations.

Install

pip install tsrbench==0.3.0