Last updated: September 2026
Welcome! This page shows what the core developers of loo are working on right now and what we have planned for the coming releases. It highlights a selection of our main goals, not a complete list of everything we work on.
We hope the roadmap is useful to you in a few ways:
A note on plans. This roadmap shows our strategic plans for the medium and long term. Alongside this work, we continue to respond to bug reports, questions, and requests as they come in. The roadmap describes our current intentions, not fixed commitments. Priorities can shift, and releases have no fixed dates. We update this page as our plans change.
Beyond 3.0.0
Your ideas are welcome
Status: In progress Planned Open for ideas
A smaller release with improvements to the current interface.
Planned
Do we want to make some further changes?
Our next major release is a big step for loo. Beyond the ELPD, you will be able to evaluate and compare models with many other predictive measures.
In progress
The new *_pred_measure() functions provide predictive
performance measures beyond the ELPD, including RMSE, MAE, MSE, \(R^2\), (C)RPS, the Brier score, and
classification accuracy. They work with in-sample, PSIS-LOO, K-fold, and
test-data predictions, and you can add your own measures.
In progress
model_compare()model_compare() succeeds loo_compare() and
lets you compare models on all built-in and custom measures. No worries:
loo_compare() will keep working, with a warning that points
you to the new function.
In progress
We are actively working on several tutorials that introduce all new features from three angles: the implementation, the workflow, and the formal, mathematical background.
Planned
We will make it easier to control warnings and messages at the package level, for example to silence them in simulations or to see more detail when debugging.
In progress
We will add more reliable Monte Carlo standard errors for PSIS-LOO, and diagnostics that show when K-fold and other refit-based results are unreliable. We also want to guide you step by step to a more robust computation when the diagnostics warn.
Planned
We will update how loo runs computations in parallel, so that it works better in settings such as computing clusters.
This part of the roadmap is still open, and we would like to shape it together with you.
Open for ideas
Is there a measure, a diagnostic, or a workflow you miss in loo? Or a topic you think deserves a higher priority? We would love to hear about it.
Contributions are welcome! Here are a few ways to join in:
remotes::install_github("stan-dev/loo"), and let us know
what works well and what does not.