metaforecast

metaforecast is a Python package for time series forecasting using meta-learning and data-centric techniques.

This package implements various techniques to improve forecasting accuracy based on dynamic model combination, data augmentation, algorithm selection, and adaptive learning, building upon Nixtla’s awesome ecosystem of state-of-the-art forecasting methods.

Features

metaforecast currently consists of five main modules:

1. Dynamic Ensembles: Combining multiple models with adaptive ensemble techniques, including online learning (exponential, polynomial, and related updates), sliding-window selection, and meta-learning-based weighting (ADE).
2. Synthetic Time Series Generation: Creating realistic synthetic time series data for robust model training and testing. Includes pure generators, semi-synthetic methods, transformation-based augmentation, and a callback for online data augmentation.
3. Long-Horizon Meta-Learning: Instance-based meta-learning for multi-step forecasting.
4. Algorithm Configuration and Selection (COSEAL): Meta-learning methods for selecting forecasting algorithms and their configurations, including MetaARIMA and ActiveTesting.
5. Evaluation: Series-wise cross-validation splitters and aspect-based accuracy analysis with ModelRadar (horizon, groups, anomalies, hard series, and ROPE comparisons).

Installation

You can install metaforecast using pip:

pip install metaforecast

[Optional] Installation from source

To install metaforecast from source, clone the repository and run the following command:

git clone https://github.com/vcerqueira/metaforecast
cd metaforecast
pip install -e .

Documentation

Check the documentation for the API reference and module descriptions. You can get started with a few tutorials.

⚠️ WARNING

metaforecast is in the early stages of development. The codebase may undergo significant changes. If you encounter any issues, please report them in GitHub Issues

License

metaforecast is dual-licensed.

• AGPL-3.0-or-later for open-source use. You may use, modify, and share the source. If you distribute a modified version or run one in production, you must make the corresponding source available.

See LICENSE for the full terms.