sktime — Unified Framework for Time Series Machine Learning

Introduction

sktime is an open-source Python library that extends the scikit-learn API to time series data. It provides a unified interface for multiple time series learning tasks, making it straightforward to apply, compare, and compose algorithms across forecasting, classification, regression, clustering, and annotation.

What sktime Does

• Provides scikit-learn-compatible estimators for time series forecasting, classification, and clustering
• Supports pipeline composition with transformers, feature extractors, and reducers
• Implements model selection tools including cross-validation strategies for temporal data
• Offers adapters to integrate algorithms from statsmodels, Prophet, and other libraries
• Includes benchmark datasets and evaluation utilities for reproducible experiments

Architecture Overview

sktime is built around a base class hierarchy that mirrors scikit-learn's estimator pattern. Each task type (forecaster, classifier, transformer) implements fit, predict, and transform methods with type-checked inputs. Pipelines chain transformers and estimators, while reduction strategies convert between task types (e.g., forecasting via regression). The framework uses a tagged object system for capability discovery.

Self-Hosting & Configuration

• Install via pip install sktime or conda install -c conda-forge sktime
• Use extras for optional dependencies: pip install sktime[all_extras]
• Configure forecasters with fh (forecasting horizon) as integer, list, or ForecastingHorizon object
• Set up cross-validation with SlidingWindowSplitter or ExpandingWindowSplitter
• Integrate with MLflow or other tracking tools using standard scikit-learn patterns

Key Features

• Consistent API across all time series tasks following scikit-learn conventions
• Composable pipelines with time series-specific transformers and reducers
• Extensive algorithm library covering classical, ML, and deep learning methods
• Built-in temporal cross-validation and performance metrics
• Active community with regular releases and comprehensive documentation

Comparison with Similar Tools

• Darts — Focused on forecasting with deep learning; sktime covers more task types
• tsfresh — Specializes in feature extraction; sktime provides a full ML pipeline
• statsmodels — Statistical models only; sktime wraps statsmodels and adds ML workflows
• scikit-learn — General ML; sktime adds time series-specific data types and methods

FAQ

Q: Is sktime compatible with scikit-learn? A: Yes. sktime follows scikit-learn's estimator interface, so pipelines, grid search, and cross-validation patterns work similarly.

Q: Can I use deep learning models with sktime? A: Yes. sktime integrates with PyTorch and TensorFlow through adapter classes and supports neural network forecasters.

Q: How does sktime handle panel (multi-instance) data? A: sktime uses a multi-index pandas DataFrame format for panel data, with instance and time indices.

Q: What license does sktime use? A: sktime is released under the BSD 3-Clause license.

Sources

• https://github.com/sktime/sktime
• https://www.sktime.net/en/stable/