Skip to content

Add risk-controlling prediction sets - #30

Open
sgreenbury wants to merge 4 commits into
mainfrom
rcps
Open

sgreenbury wants to merge 4 commits into
mainfrom
rcps

Conversation

@sgreenbury

Copy link
Copy Markdown
Collaborator

Summary

  • Add risk-controlling prediction-set calibration with coverage loss and a Hoeffding upper confidence bound.
  • Support asymmetric scaled intervals and raw ensemble forecasts using quantile or standard-deviation interval families.
  • Configure alpha and delta grids before calibration and return their Cartesian product on explicit output axes.
  • Document the relationship to conformal ensemble calibration, per-pair guarantees, and extension contracts for other losses and bounds.

Testing

  • uv run pytest --maxfail=3
  • uv run prek run --hook-stage manual --all-files

Closes #22

Introduce bounded-loss RCPS calibration with a Hoeffding upper
confidence bound and per-example coverage loss. Support asymmetric
scaled intervals and raw ensemble forecasts through quantile or
standard-deviation interval families.

Configure alpha and delta grids before calibration so fitted scales and
prediction axes are explicit. Keep structural loss and bound contracts
open to future extensions, and document per-pair guarantees alongside
the existing conformal approaches.
Use semantic tensor aliases in the risk-loss interface and document the
ensemble CP and RCPS APIs side by side. Add a regression test for the
expected relationship between confidence level and fitted scale.
@sgreenbury
sgreenbury marked this pull request as ready for review August 11, 2026 14:12
Keep the additive Hoeffding term on the same display-math line so
Markdown renderers do not interpret its leading plus sign as markup.
Clarify that scaled-interval inputs contain a centre and non-negative
distances from it, rather than interval endpoints.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

RCPS (bounded loss, Hoeffding UCB)

1 participant