sktime/skpro

What it solves

Many standard machine learning regressors only provide a single point estimate (the mean) for a prediction. skpro solves this by enabling supervised probabilistic prediction, allowing users to predict intervals, quantiles, and full probability distributions instead of just a single value.

How it works

skpro provides a set of scikit-learn compatible interfaces and tools to transform standard regressors into probabilistic ones. It uses techniques such as reductions (e.g., bootstrap or conformal prediction) to turn existing scikit-learn regressors into probabilistic models. It also provides dedicated modules for tabular probabilistic regression, time-to-event (survival) prediction, and symbolic probability distributions that integrate with pandas DataFrames.

Who it’s for

Data scientists and ML engineers who need to quantify uncertainty in their predictions, specifically those working with tabular data or survival analysis, and those who prefer the scikit-learn ecosystem.

Highlights

• Probabilistic Regression: Predicts intervals, quantiles, and full distributions.
• Survival Analysis: Supports probabilistic time-to-event and survival prediction.
• Evaluation Metrics: Includes specialized metrics for probabilistic forecasts, such as pinball loss, CRPS, and empirical coverage.
• Interoperability: Fully compatible with scikit-learn and sktime, and interfaces with third-party libraries like MAPIE, ngboost, and cyclic-boosting.