winedarksea/AutoTS
Automated Time Series Forecasting
What it solves
AutoTS is designed to rapidly deploy high-accuracy time series forecasts at scale. It automates the process of selecting the best forecasting models, preprocessing steps, and ensembling techniques for a given dataset, removing the manual trial-and-error typically associated with time series analysis.
How it works
The library provides a wide array of forecasting models (naive, statistical, machine learning, and deep learning) and over 30 time series specific transforms. It uses a genetic algorithm to perform an AutoML feature search, automatically finding the optimal combination of models and preprocessing for the dataset. It supports multivariate outputs, probabilistic forecasts, and can scale to hundreds of thousands of input series.
Who it’s for
Data scientists and analysts who need to actually deploy high-accuracy forecasts for large-scale time series data without manually tuning every model.
Highlights
- Automated Model Selection: Uses genetic algorithms to find the best models and preprocessing steps.
- Scalability: Capable of handling tens to hundreds of thousands of input series.
- Flexible Data Formats: Works directly with Pandas Dataframes in both long and wide formats.
- Ensembling: Features horizontal and mosaic style ensembles to maximize accuracy while maintaining scalability.
- Probabilistic Forecasting: Supports upper and lower bound forecasts for uncertainty estimation.