Skip to content

feat: publish calibrated scoring probability as stat - #152

Merged
nathan-probert merged 2 commits into
mainfrom
feat/calibrated-probability
Nov 9, 2026
Merged

nathan-probert merged 2 commits into
mainfrom
feat/calibrated-probability

Conversation

@nathan-probert

Copy link
Copy Markdown
Owner

Implements calibrated probability for the ML picks stat column while keeping pick ranking unchanged.

  • ensemble_scores now computes prob = mean of member predict_proba outputs (actual scoring probability), with optional bundle calibrator applied (monotonic, so ranking stays on the ensemble percentile).
  • stat_value now returns a 0-1 probability (no /100 scaling); stat_update_picks_db and pick_rows_to_picks_db publish prob instead of the ensemble percentile.
  • train_export fits an isotonic calibrator (monotonic) onto the mean member probability vs the scored label and stores it on the artifact.
  • normalise_bundle propagates the calibrator (absent on older artifacts -> raw probability fallback).
  • Updated affected unit tests.

@nathan-probert
nathan-probert merged commit 4049713 into main Nov 9, 2026
6 checks passed
@nathan-probert
nathan-probert deleted the feat/calibrated-probability branch November 9, 2026 20:09
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.

1 participant