Upcoming long-bond auctions - exploratory extension, November 7, 2026

These indicators were not in the earlier feature search. This is a new research
round suggested after the previous paper results were known. This file describes
the first round, using the ANNUAL reference. The later quarterly-model round is
in quarterly/FINDINGS.txt. No auction rule is in the selected production model.
Read PROTOCOL.txt for the choices recorded before the first calculations.

INFORMATION TIMING
The auction must occur after signal Friday and by the following Friday, and its
announcement must be dated strictly before signal Friday. This conservative
gate excludes same-Friday announcements without guessing their publication time.
Reopenings use original tenor; 20-year TIPS are not nominal 20-year bonds.
Older 30-year issues with an initial additional three or six months remain in
the 30-year family. No auction-result fields define any predictor.

The source is the official Treasury calendar database. All 23 original
announcements in an outcome-independent first/last-per-calendar-block sample
confirm announcement date, auction date and CUSIP. This is a sample audit, not
a complete point-in-time database or verification of every calendar revision.
The paper's existing market-history limitations also remain.

MAIN FINDING
The highest early-sample gross calendar-week t-statistic selects one combined
auction flag, with a reduce-duration checklist vote, added to the unchanged
13-rule model and its regression. It uses 118 early event weeks. Its early
score is 29.4 yield bp/year versus 18.3 for the reference. The ranking excludes
all outcomes ending on or after November 1, 2016. Eight variants were tried,
four flags times two vote directions, with at least 20 early events required.

Later comparison, 521 completed weeks:
                              Gross bp/yr   Assumed net bp/yr
Main 13-rule model                  22.6          17.6
Selected combined-auction model     33.5          28.4
Always higher yields                31.4

The selected model's paired gain over the main model is 11.0 bp/year,
t=1.81. This is descriptive, without adjustment for the new search or earlier
research decisions. It does not establish a reliable improvement.
The paired gain over always higher yields is only 2.1 bp/year, t=0.07.
Its last-52-week score is 69 bp versus 58; its completed-2026 total is 72 versus 58.
Those short windows are not annualized. Scores are not portfolio returns.

QUALIFICATIONS THAT MATTER
1. The nominal 20-year bond returned in May 2020. It has zero early events and
   cannot qualify for the original early-sample selection. Both signs are
   disclosed as exploratory, with a matched May 2020-onward comparison.
   The union flag changes composition when 20-year issuance enters.
2. There are only 20 early and 22 later 30-year TIPS event weeks. Two fall in
   2026. This is too sparse for a strong standalone or annual conclusion.
3. Nominal 30-year auction-week average changes reverse from -1.76 bp early
   to +1.83 later. The supply story is a hypothesis, not an identified cause.
4. Reducing duration only in any-auction weeks scores 46.5 bp/year later
   (t=2.78), but -13.5 early. Its opposite sign reverses those gross scores.
   This larger later number does not establish stability across regimes.
5. Both checklist vote signs can improve an augmented model because the new
   regression input also changes fitted coefficients and non-auction weeks.
   This does not establish two opposite economic mechanisms.
6. The outcome covers concession before and adjustment after the auction.
   It does not isolate either, or simulate auction-day execution.

All eight combined variants and both standalone signs are in all_results.csv
and summary.json, including 2026, last 52 weeks, and matched 20-year periods.
The summary includes paired comparisons against the main model and always
reduce, hit-rate Wilson ranges, four/eight-lag HAC checks and assumed costs.
The earlier placebo exercises do not adjust this search.

NEXT USE
Retain as a prospective research candidate. Do not retrofit the live model's
specification, its historical headline, or its future archive without a clearly
dated new model version and a separate forward record.

Related quarterly-refit check: improvement_study/refit_comparison.json.
Thirteen weeks means refitting every quarter with expanding history, not
training on thirteen weeks. It wins in the later decade (30.6 versus 22.6)
but loses early (15.4 versus 18.3); the later paired t is only 0.71. The last
52 weeks also favor the annual reference, 58 versus 52 bp. Subsequently the
quarterly version was selected for prospective use, explicitly informed by the
later historical comparison. Version v3 discloses that retrospective choice.
