voter_edge.py

Sentinel/Lab/voter_edge.py


Family	Lab (Python)
Version	—
Size	89 lines
Scope	public — ships in `sentinel-suite`
Documented by	no doc tracks this artifact

Rendered from the published copy in sentinel-suite/src/, not the author’s private tree — so this page describes the file you actually have.

What the file says about itself

Model-free voter edge — is a voter CONFIRMING or CONTRARIAN, and on how much data?

    python voter_edge.py --inst GC --bartypes 212201v6x24 0v150x1 9v1x1

The ridge weights in train.py are confounded by heavy collinearity among the price-derived
voters, so a negative coefficient can be an artifact of another voter stealing the credit.
This looks at each voter UNIVARIATELY: fold x = vote*dir (+1 = agreed with the taken side),
then the uniqueness-weighted first-touch win rate when it AGREED vs DISAGREED. A real
confirming voter: wr(agree) > base > wr(disagree). Contrarian: the reverse. n_agree /
n_disagree is the data behind the claim -- a trigger that rarely fires can't be trusted, and
'(thin)' flags when either bucket is under 40. Read this BEFORE believing any learned weight.