Implementation:Sktime Pytorch forecasting Generate Ar Data

Knowledge Sources	pytorch-forecasting
Domains	Time_Series, Data_Engineering, Experimentation
Last Updated	2026-02-08 07:00 GMT

Overview

Concrete tool for generating synthetic autoregressive time series data provided by the pytorch-forecasting library.

Description

The generate_ar_data function creates a DataFrame of synthetic multivariate time series without covariates. Each series is generated from a combination of linear trend, quadratic trend, sinusoidal seasonality, and Gaussian noise with controllable parameters. The output DataFrame has three columns: series (group identifier), time_idx (integer time index), and value (target). This is the canonical data source for DeepAR and N-BEATS examples.

Usage

Import this function when you need synthetic univariate time series data for testing or demonstrating DeepAR or N-BEATS models. The function provides reproducible data via a seed parameter. Canonical usage for DeepAR: generate_ar_data(seasonality=10.0, timesteps=400, n_series=100).

Code Reference

Source Location

• Repository: pytorch-forecasting
• File: pytorch_forecasting/data/examples.py
• Lines: L53-111

Signature

def generate_ar_data(
    n_series: int = 10,
    timesteps: int = 400,
    seasonality: float = 3.0,
    trend: float = 3.0,
    noise: float = 0.1,
    level: float = 1.0,
    exp: bool = False,
    seed: int = 213,
) -> pd.DataFrame:
    """
    Generate multivariate data without covariates.

    Each timeseries is generated from seasonality and trend.

    Args:
        n_series: Number of series. Defaults to 10.
        timesteps: Number of timesteps. Defaults to 400.
        seasonality: Normalized frequency. Defaults to 3.0.
        trend: Trend multiplier. Defaults to 3.0.
        noise: Level of gaussian noise. Defaults to 0.1.
        level: Level multiplier. Defaults to 1.0.
        exp: If to return exponential of values. Defaults to False.
        seed: Random seed. Defaults to 213.

    Returns:
        pd.DataFrame: data with columns series, time_idx, value
    """

Import

from pytorch_forecasting.data.examples import generate_ar_data

I/O Contract

Inputs

Name	Type	Required	Description
n_series	int	No	Number of independent series to generate (default: 10)
timesteps	int	No	Length of each series (default: 400)
seasonality	float	No	Normalized frequency of seasonal component (default: 3.0)
trend	float	No	Trend multiplier (default: 3.0)
noise	float	No	Gaussian noise level (default: 0.1)
level	float	No	Level multiplier for series amplitude (default: 1.0)
exp	bool	No	Whether to exponentiate values (default: False)
seed	int	No	Random seed for reproducibility (default: 213)

Outputs

Name	Type	Description
return	pd.DataFrame	DataFrame with columns: series (int group id), time_idx (int time index), value (float target)

Usage Examples

DeepAR Data Generation

from pytorch_forecasting.data.examples import generate_ar_data

# Generate 100 series with strong seasonality for DeepAR
data = generate_ar_data(seasonality=10.0, timesteps=400, n_series=100)

print(data.head())
#    series  time_idx     value
# 0       0         0  0.123456
# 1       0         1  0.234567
# ...

# Split into training and validation
training_cutoff = data["time_idx"].max() - 20
training = data[lambda x: x.time_idx <= training_cutoff]
validation = data

N-BEATS Data Generation

from pytorch_forecasting.data.examples import generate_ar_data

# Generate data for N-BEATS univariate forecasting
data = generate_ar_data(seasonality=10.0, timesteps=400, n_series=100, seed=42)

# Add static grouping column (N-BEATS example pattern)
data["static"] = 0

# Define training cutoff
context_length = 150
prediction_length = 20
training_cutoff = data["time_idx"].max() - prediction_length

Related Pages

Implements Principle

• Principle:Sktime_Pytorch_forecasting_Synthetic_Data_Generation

Requires Environment

• Environment:Sktime_Pytorch_forecasting_Core_Python_Dependencies