TBATS for Time Series Forecasting

1.1.3 · active · verified Sun Apr 12

tbats is a Python library implementing BATS and TBATS models for time series forecasting, known for handling complex seasonality and Box-Cox transformations. The current version is 1.1.3, and it receives updates primarily for bug fixes, dependency upgrades, and occasional feature enhancements, with an irregular release cadence.

Warnings

gotcha Prior to version 1.1.3, numeric precision issues could cause Box-Cox transformations to fail or produce incorrect results in certain scenarios.
Fix: Upgrade to tbats version 1.1.3 or higher to resolve Box-Cox precision issues.
breaking In version 1.1.0, the default method for spawning new processes for parallel execution (`n_jobs > 1`) was changed from the system default to 'spawn'. This may alter behavior, particularly on Windows or macOS where the default 'fork' method has different implications.
Fix: If unexpected behavior related to multiprocessing occurs, explicitly set `multiprocessing_start_method` in the `TBATS` constructor (e.g., `TBATS(..., multiprocessing_start_method='fork')`) to match previous behavior or to 'fork' if that is preferred and supported by your OS.
gotcha Versions prior to 1.0.6 (and specifically 1.0.5 had fixes) suffered from resource leaks where multiprocessing pools were not reliably closed, potentially leading to 'too many open files' errors or system instability.
Fix: Upgrade to tbats version 1.0.6 or higher. If using older versions, ensure your application handles process pool management explicitly or restricts `n_jobs` to 1.
gotcha When `n_jobs=1` was specified in versions prior to 1.0.9, subprocesses might still have been triggered, leading to unnecessary overhead or potential issues with environments not configured for multiprocessing.
Fix: Upgrade to tbats version 1.0.9 or higher to ensure that `n_jobs=1` correctly disables all subprocess spawning.

Install

pip install tbats Install tbats

Imports

TBATS
```
from tbats import TBATS
```

Quickstart

This quickstart demonstrates how to initialize, fit, and forecast with a TBATS model. It generates a sample time series with trend and seasonality, then applies TBATS with Box-Cox transformation and ARMA errors to predict future values. `n_jobs=1` is used to prevent parallel processing in simple examples.

import numpy as np
from tbats import TBATS

# Generate some example time series data with seasonality
np.random.seed(42)
n_points = 100
seasonal_period = 24 # Daily seasonality
t = np.arange(n_points)
y = 50 + 2 * t + 10 * np.sin(2 * np.pi * t / seasonal_period) + np.random.normal(0, 5, n_points)

# Create and fit the TBATS model
estimator = TBATS(seasonal_periods=[seasonal_period],
                  use_box_cox=True,
                  use_trend=True,
                  use_damped_trend=False,
                  use_arma_errors=True,
                  n_jobs=1)
model = estimator.fit(y)

# Make a forecast
forecast = model.forecast(steps=10)

print("Original series (last 5 points):", y[-5:])
print("Forecasted values (next 10 steps):", forecast)

view raw JSON →