Exploratory improvement search - November 7, 2026
Written before running this additional search. The 2016-2026 history has already
been inspected during prior development. This is disclosed in-sample data mining,
not preregistration of an untouched test or a new prospective result.

Question: can an explainable variation improve the calendar-week yield score
without relying on larger maximum exposure? The frozen production model and its
source files are not changed by this experiment.

Data: the existing archived local research panel, with the corrected timing and
archive rules. The current-vintage market-history limitations still apply.

Selection sample: forecasts from January 6, 2006 whose following-Friday outcomes
occur BEFORE November 1, 2016. Thus the final selection signal is October 21 and
its outcome is October 28. No November 4 outcome enters this search's ranking.
Evaluation: November 4, 2016 through October 23, 2026 signals, outcomes through
October 30, 2026. It is a later historical evaluation, already seen by the author.

Grid fixed for this run (2,430 combinations):
5 feature sets: main 13; omit all three foreign-yield inputs (10); market-move
inputs plus December (US equities, EM equities, WTI, USD, VIX, curve, December: 7);
main 13 plus breakeven inflation (14); and main 13 plus the previously studied
jobs calendar proxy and 300-day yield trend (15). The last two are named candidates
in this new search, not represented as discoveries independent of prior results.
Candidate directions use only outcomes ending before November 2016.
3 training histories: all available history; latest 520 weeks; latest 260 weeks.
3 regularization strengths C: 0.01, 0.1, 1.0.
2 refit intervals: 52 or 13 weeks, anchored on January 6, 2006.
3 minimum absolute checklist scores: 0, 1, 3.
3 minimum distances of probability from 0.5: 0, 0.025, 0.05.
3 position sizes: fixed unit; probability strength min(abs(p-0.5)/0.10, 1);
checklist strength abs(score)/number_of_rules. Every position is capped at one.
Direction always follows the existing non-disagreement combination. A zero
checklist permits the classifier only when the minimum score is zero. Checklist
scaling makes a zero checklist flat. Standardization stays expanding and causal.

Ranking: highest in-sample calendar-week Sharpe-like ratio AFTER an illustrative
charge of 0.25 yield bp per position unit, including opening and closing. A flip
counts as two units. This is a score sensitivity, not an executable return or a
validated transaction-cost estimate. Require at least 25% active weeks, mean
absolute exposure at least 0.10, and positive net scores in both 2006-2010 and
2011-October 2016. Ties resolve in the deterministic grid order.

Freeze the overall winner and the winner within each of the five feature sets
to a selection file BEFORE calculating their later-period performance. Report
all five feature-set winners, the baseline and the always-higher-yield benchmark.
Do not replace the overall in-sample winner using later-period rankings.
Record all 2,430 in-sample trials, including failed eligibility checks.

Evaluation: annual gross and cost-adjusted score, 4-lag HAC t, calendar-week
Sharpe-like ratio, active-week hit rate, mean absolute exposure, turnover, score
drawdown and paired differences to baseline. Also show 2016-2019, 2020-2022,
2023-2026 and the final 52 weeks. Fractional positions must be scored at their
actual sizes; model.evaluate() takes signs and therefore cannot score them.
Check probability quality with Brier loss versus 0.5 and a past-frequency
benchmark, and show fixed probability bins. Do not call these calibration tests
proof that probability magnitude is an appropriate position size.

No formal search-adjusted significance claim will be made. All reported t-values
are descriptive and do not account for the 2,430 trials or previous human choices.
Even a later-period gain does not resolve vintage risk, live data timing or
selection history. No source code for the frozen weekly signal will be replaced.
