Simulating Prediction INtervals And CHance -- placeholder for a tidymodels friendly package for a generalizable approach to simulating uncertainty in predictions. See {spin} for initial toy package that contains some initial notes on API: https://github.com/brshallo/spin
- end-goal would be to pass in a tidymodels workflow (that may include e.g. a parsnip model specification, a preprocessing recipe, a cross-validation resampling scheme) and would output an object that could be used in a
predict()method to provide uncertainty levels for new observations. - see field of conformal inference for broad range of approaches / "uncertainty specifications" that may look to set-up.
- a lot of the most successful approaches though seem to incorporate quantile regression, etc, so may not be able to stay agnostic to model fit procedure.