Heidrich, Kiraly, & Ray - sktime - python toolbox for time series | PyData Global 2023

sktime is an open-source library for time series learning that integrates multiple time series packages and provides unified interfaces similar to scikit-learn

The library supports multiple time series tasks including:

• Forecasting
• Classification
• Regression
• Clustering
• Annotation

Key features introduced in 2022-2023:

• Graphical pipelines for complex non-sequential workflows
• Improved parallelization for multivariate and hierarchical data
• Probabilistic forecasting with distribution objects
• Benchmarking capabilities
• Marketplace and deployment features

Forecasting interfaces follow scikit-learn patterns with fit/predict methods and support:

• Exogenous variables
• Multiple seasonality patterns
• Automatic parameter tuning
• Probabilistic predictions with confidence intervals

The library provides adaptors to many popular time series packages including:

• ARIMA
• Prophet
• TSFresh
• PDM
• TBATS

Pipeline capabilities include:

• Sequential pipelines for simple workflows
• Graphical pipelines for complex cases with parallel steps
• Transformation pipelines for preprocessing
• Composable interfaces across different tasks

Strong focus on community and open governance:

• Permissive license
• Mentoring program for new contributors
• Active Discord community
• Regular developer sprints

Extensive support for different data formats:

• Univariate and multivariate series
• Hierarchical data
• Panel data
• Various time index types

Built-in tools for:

• Model evaluation and validation
• Parameter tuning
• Cross-validation
• Performance metrics

Integration capabilities with:

• MLflow
• scikit-learn
• pandas