Second exploratory search, November 7, 2026
Written after observing the first 2,430-trial classification search's failure on
the later sample, and before running this search. Both rounds must be disclosed.

Hypothesis: direction-only classification ignores move size, although the reported
yield score depends on size. Test ridge regression of the next weekly yield fall
in basis points on the same numerical inputs. This remains a simple linear model.
Positive prediction means lower yields. Coefficients use only known past outcomes.

Keep the first protocol's data, date boundaries, five feature families, three
training windows, standardization, two refit frequencies, eligibility and primary
in-sample ranking. Replace C with ridge alpha = 1, 10, 100, 1000. Test both using
the regression alone and requiring it not to conflict with the checklist.
Minimum absolute forecast: 0, 1 or 2 yield basis points.
Position: fixed unit, or min(abs(predicted yield fall)/5,1) times its sign.
Total combinations: 5*3*4*2*2*3*2 = 1,440. No new features are searched.

Freeze overall and feature-family winners using only the early sample. Evaluate
all five on the same later dates. Do not combine this with another round after
seeing the result. Report the first and second rounds together: 3,870 settings.
Later results remain retrospective, and data availability limitations are unchanged.
