Auction additions to the selected quarterly model - November 7, 2026

Decision: retain 13-week refitting WITHOUT auction additions.

This round followed inspection of the annual auction results and the decision
to use quarterly refitting. The same four flags and both vote signs were fixed
before the quarterly combinations were calculated. It is eight further trials,
not an untouched test. The first round's eight annual variants remain separate.

The early ranking selects the nominal 30-year auction flag with an add-duration
vote. Early score: 23.0 versus 15.4 bp/year for quarterly refitting alone. In the
521 later weeks, the selected extension scores 26.3 versus 30.6; after the assumed
charge, 20.9 versus 25.4. The paired gross difference is -4.3 bp/year, t=-0.72.
Hit rate is slightly higher, 52.8% versus 52.6%, but the yield score is lower.
The size of wins and losses matters; hit rate alone cannot determine usefulness.

Later gross scores (all calendar weeks, fixed unit positions):
                                    Reduce vote    Add vote
20-year nominal, full later period       25.0          21.9
30-year nominal                          25.0          26.3  [early winner]
30-year TIPS                             32.1          32.6
Any of the three                         25.5          23.6
Quarterly baseline                       30.6

Nominal 20-year has zero early events and was ineligible for selection. For a
matched May 2020-onward view, its reduce/add variants score 24.1/19.1 versus
32.7 for the quarterly baseline. The paper displays these matched-period rows.

Only TIPS improves later gross scores. Gains are 1.5-2.0 bp/year, paired t
0.99-1.27. Neither was the early winner; there are only 22 later TIPS event
weeks. These data do not establish an improvement. Choosing the better TIPS
sign now would be another selection informed by the later results.

The early-selected 30-year addition scores 70 bp over the final 52 weeks,
versus 52 for quarterly alone, and 70 versus 52 over completed 2026 weeks.
These recent gains do not erase the weaker full later-period result.

The first round's apparent improvement from the any-auction flag therefore
does not carry over to the selected quarterly baseline. Event-only results do
not depend on refit frequency and are the same study observations, not fresh
confirmation. Source timing, sparse TIPS history, 20-year composition changes,
current-vintage data limitations and multiplicity caveats remain unchanged.

Files: PROTOCOL.txt; selection_before_later_evaluation.json; summary.json;
all_early_trials.csv; all_results.csv; all_weekly_positions.csv; weekly_flags.csv;
eligible_events.csv. The parent folder holds the shared official source extract,
23-announcement sample audit, downloader, analysis code and timing tests.
Run: python auction_study/analyze.py --refit-weeks 13
