useR! 2024

Ensemble models are increasingly popular tools for capturing heterogeneous information and improving predictive performance. We will present the `gamstackr` R package, which provides tools for aggregating or `stacking` the probabilistic forecasts produced by different models or `experts`. In particular, the package implements a versatile, easy-to-use framework for probabilistic stacking that allows to control the experts’ weights via additive models containing fixed, random or smooth effects. It also provides statistical and computational scalability in the number of experts by exploiting context-specific relationships between them.

We will illustrate the typical workflow of the `gamstackr` package, that is how to: create a heterogeneous set of experts, build and fit several types of stacking models and visualise the ensemble weights and their relationship with the covariates. The package is currently available at https://github.com/eenticott/gamstackr.