09-14, 13:30–15:00 (Europe/Amsterdam), Hello, World! (Tutorials)
Probabilistic predictions are predictions that include some statements about uncertainty of the prediction, e.g., prediction intervals that make statements about a likely range of values that a prediction can take.
This workshop gives an introduction on making probabilistic predictions with the sktime and skpro python packages, for forecasting and supervised regression. Both packages are sklearn-compatible, built using skbase, with composable and modular interfaces.
The presentation includes a practical primer of different types of probabilistic predictions, algorithms and estimators, and evaluation workflows, with python code examples.
Probabilistic predictions make statements about the uncertainty or likely variation of the forecast, e.g., intervals at nominal coverage or conditional distributions.
They appear in probabilistic forecasting as well as in probabilistic supervised (tabular) regression.
This tutorial presents probabilistic forecasting capability in the skpro and sktime packages, combined with a methodological overview.
sktime is a widely used package for time series, skpro covers probabilistic (tabular) regression. Both are based on skbase, and designed for interoperability with each other and sklearn.
This tutorial presents the joint designs for probabilistic predictions and modular estimator interfaces. It also gives an overview of pipelines, tuning using probabilistic metrics, and compositors that can be used to turning any point forecaster into probabilistic forecasters, such as conformal or empirical interval estimators.
The presentation will showcase skpro and sktime, for tabular and time series tasks:
- probabilistic prediction interfaces
- metrics, e.g., quantile loss, or CRPS, log-loss for distribution forecasts
- tuning using probabilistic metrics
- conformal probabilistic intervals for any pipeline
- compositors to make any point prediction estimator probabilistic
- for time series: hierarchical and global probabilistic forecasts, reduction to regression
From a methodological perspective, we will cover:
• interval forecasts: producing intervals with a nominal probability of the observation to be contained in the interval
• quantile forecasts: specifying one or multiple quantiles of a predictive forecast distribution
• fully probabilistic forecasts: producing a symbolic representation of a predictive forecast distribution
• simulators or samplers from probabilistic forecasting models
As research on software interfaces and mathematical conceptualization in this area is still an ongoing endeavour, challenges will also be discussed, with invitations to contribute.
sktime and skpro are developed by an open community, with aims of ecosystem integration in a neutral, charitable space. We welcome contributions and seek to provides opportunity for anyone worldwide.
No previous knowledge expected
(1 speaker will attend if accepted, exact speaker tbd)