Welcome to sktime’s roadmap.

Contributors: @mloning, @fkiraly, @sveameyer13, @lovkush-a, @bilal-196, @GuzalBulatova, @chrisholder, @satya-pattnaik, @aiwalter

Created during the 2021 sktime dev days, 25/06/2021.

Project aims#

The aim of sktime is to:

  • Develop a unified framework for machine learning with time series in Python

  • Advance research on algorithm development and software design for machine learning toolboxes

  • Build a more connected community of researchers and domain experts who work with time series

  • Create and deliver educational material including documentation and user guides

Work streams#


  • Core documentation needs to be created “properly”

  • Improve tutorials, examples

  • Improve extension guidelines

  • For research algorithms, possibly pairing up researchers with ‘engineer’ to improve readability/documentation

Community building#

  • Integrate “off-line” contributors

  • For research algorithms, possibly pairing up researchers with “engineer” to improve readability/documentation

  • Establish regular technical and social meetings

Refactoring and extending existing modules#

  • Support for data input types and conversion (e.g. awkward-array)

  • Distance metrics

  • Reduction interface

  • Advanced pipelining

  • Forecasting
    • Prediction intervals and probabilistic forecasting

    • Streaming data interface, “update” capability of estimators

    • multivariate/vector forecasting

    • consistent handling of exogeneous variables

    • fitted parameter interface

  • Time series classification/regression/clustering
    • add support for unequal length time series

    • add data simulators for algorithm comparison and unit testing

  • Clustering
    • interface scikit-learn estimators

    • implement time series specific estimators (e.g. k-shapes)

  • Series annotation
    • implement more estimators for outlier anomaly/detection and segmentation

Adding new modules and algorithms#

  • Panel annotation

  • Probabilistic interface, event modelling(time-to-event modeling, survival analysis)

  • Panel & supervised forecasting

  • Time series regression

  • Sequence-similarity tasks

  • Uniform reduction interface between tasks

Software engineering & dev ops#

  • Improve dependency management

  • Create template repository for companion packages

  • Improve continuous integration & deployment
    • Refactoring unit tests

    • Extending unit tests

    • Speed up unit tests

    • Make unit tests for estimators importable from other packages