sktime for composable time-series learning

How it works

The library provides forecasting horizons, temporal splitters and transformers that respect ordered data. Pipelines combine time-series preprocessing with estimators.

What you need

Decide the forecasting horizon and which external variables are available at each timestamp. Build the baseline with the same information that production predictions will receive.

What to try first

Backtest a seasonal baseline and one learned model on rolling windows. Compare errors by forecast horizon rather than relying on an overall average.

Backtest a period with missing observations and one with a seasonal change. The forecast must remain aligned with the timestamps supplied to the application.

Where to look for failures

Feature availability changes with the forecast timestamp. Ensure regressors and scaling use only information available at prediction time.

Related Franklin pages

A random train-test split can hide time-dependent failures. Use rolling or sequential evaluation and match the feature information available at each prediction time.

• Build predictive pipelines with consistent preprocessing, validation and estimator APIs. scikit-learn for reproducible machine-learning baselines
• Update models and metrics one observation at a time in streaming applications. River for learning from data streams

For a primer on learning from data and evaluating predictions, see Franklin’s glossary. Machine learning

Project documentation

• Current project README
• README at the reviewed revision