AutoGluon-TimeSeries

1.5.0 · active · verified Sun Apr 12

AutoGluon-TimeSeries is a module within the AutoGluon automated machine learning library, specifically designed for robust and accurate time series forecasting. It enables users to train state-of-the-art forecasting models with minimal code, leveraging various algorithms and ensemble methods. The library is actively maintained with frequent major and minor releases, providing continuous improvements in performance and features. The current version is 1.5.0.

Warnings

breaking Python version support changes significantly across major/minor releases. AutoGluon 1.2.0 dropped support for Python 3.8 and added 3.12. The current version (1.5.0) requires Python >=3.10 and <3.14.
Fix: Ensure your Python environment meets the specific `requires_python` range for your AutoGluon version before upgrading or deploying.
gotcha Models trained with one version of AutoGluon are often not compatible with other versions (especially across major/minor releases). Attempting to load models saved with a different AutoGluon version can lead to errors or unexpected behavior.
Fix: Always load previously trained models using the *exact same version* of AutoGluon that they were originally trained on. Consider pinning AutoGluon versions in production environments.
gotcha AutoGluon's `fit` method, especially with default `presets` or large datasets, can be very resource-intensive (CPU, RAM, GPU if enabled). Users may experience long training times or out-of-memory errors.
Fix: Start with `presets='fast_training'` or `presets='medium_quality'` for initial experimentation. Monitor resource usage and adjust `hyperparameters` or `time_limit` for larger datasets or production deployments.

Install

pip install autogluon.timeseries Install AutoGluon TimeSeries

Imports

TimeSeriesPredictor

from autogluon.timeseries import TimeSeriesPredictor

TimeSeriesDataFrame

from autogluon.timeseries import TimeSeriesDataFrame

Quickstart

This quickstart demonstrates how to prepare time series data using `TimeSeriesDataFrame`, initialize `TimeSeriesPredictor`, train a model with a specified preset, and generate forecasts for future time steps.

import pandas as pd
from autogluon.timeseries import TimeSeriesPredictor, TimeSeriesDataFrame

# Create dummy data for demonstration
data = {
    "item_id": ["A", "A", "A", "B", "B", "B"],
    "timestamp": pd.to_datetime(["2023-01-01", "2023-01-02", "2023-01-03",
                                "2023-01-01", "2023-01-02", "2023-01-03"]),
    "target": [10, 12, 15, 20, 22, 25],
}
df = pd.DataFrame(data)

# Convert to TimeSeriesDataFrame format
train_data = TimeSeriesDataFrame(df)

# Initialize and train the predictor
predictor = TimeSeriesPredictor(
    prediction_length=2, # Forecast 2 future time steps
    target="target",     # Column to forecast
    eval_metric="MASE"   # Evaluation metric
)
predictor.fit(train_data, presets="fast_training") # Use a fast preset for quick demo

# Make predictions
predictions = predictor.predict(train_data)
print(predictions.head())

view raw JSON →