UPCOMING LONG-BOND AUCTIONS - EXPLORATORY EXTENSION
Specified November 7, 2026, after the user's suggestion and before calculating
auction-conditioned yield results. Earlier paper results were already known.

Question: does an announced long-bond auction in the following week help time
the direction of the US ten-year yield, alone or alongside the main model?

Inputs: official Treasury auction records. Only announcement date, auction date,
original term, inflation-index flag and identifying fields may define features.
Never use an auction's yield, tail, bid-to-cover, allocation, or eventual demand.
Do not infer historical expectations from the current tentative calendar.

Four fixed candidates:
1. 20-year NOMINAL bond auction.
2. 30-year NOMINAL bond auction.
3. 30-year TIPS auction.
4. Any of the above (a single combined flag, not a count or three votes).
Include reopenings using original term, not remaining term. Exclude 20-year TIPS
from the 20-year nominal candidate. No size, threshold, volatility interaction,
calendar interaction, issuance-netting, or other tenor search in this round.

For Friday t, the flag is one only if t < auction date <= t+7 days and the
announcement date is STRICTLY BEFORE t. This one-calendar-day gate avoids
assuming a historical Friday publication time. Otherwise the flag is zero.
The database's date fields remain a current retrieval of dated records; audit
archived announcements and report discrepancies. Do not call this a complete
point-in-time reconstruction or ignore unresolved gaps.

The prespecified supply-pressure hypothesis is reduce duration on an auction
week, flat otherwise. Also disclose the opposite sign. The weekly target spans
both pre-auction concession and post-auction adjustment; it does not isolate
either mechanism and does not measure auction-day execution returns.

Tests:
A. For all four flags, report event and non-event counts and conditional mean
   following-week yield changes. Score both event-only directions with zero
   exposure on non-event weeks. Use calendar-week HAC t statistics (4 and 8
   lags), hit rates and simple Wilson ranges. Report all trials.
B. Add ONE auction vote at a time to the unchanged thirteen-rule checklist and
   one binary auction input to its expanding logistic regression. Use C=0.1,
   the original annual refit anchor and fixed positions. Both vote directions
   are tried; the classifier input remains the same 0/1 observable. Report
   all eight variants and paired weekly differences versus the reference.
C. Choose one exploratory combined-model candidate using the highest EARLY
   calendar-week gross t statistic, subject to at least 20 early event weeks.
   Break exact ties by candidate name then sign. Save the selection before
   calculating later-period comparisons. Also save each eligible flag's chosen
   sign. No choices use later outcomes. This is still retrospective research.

The early ranking uses signal weeks from January 6, 2006 whose outcomes ended
before November 1, 2016 (last signal October 21, outcome October 28). Later
evaluation starts November 4, 2016; the frozen outcome endpoint is October 30,
2026. Report the last 52 weeks and completed 2026 weeks separately, with counts.

The reintroduced 20-year nominal bond has no early events. It cannot be selected
or validated on the original early sample. Report both signs as exploratory
2020-onward diagnostics with a matched-period baseline; do not disguise the
structural zero period as an early test. Report 30-year TIPS sparsity explicitly.

Costs: retain the same illustrative 0.25 yield-bp per position-unit turnover,
including opening/closing exposure. It is an assumption, not an observed cost.
Show gross and illustrative net paired comparisons. No nominal p-value is a
search-adjusted probability. All eight trials, eligibility and period boundaries
must accompany any winner. The earlier placebo tests do not cover this search.

Decision: keep the frozen production specification and previous historical
results intact. Any improvement is a candidate for prospective evaluation,
not permission to replace the live rules silently.
