2026 relationships to watch in 2027
Analysis specified November 7, 2026, after the paper's historical results had
already been seen. This is expressly in-sample research, not a new holdout test.

Use every one of the 13 existing rules and 14 existing candidate rules, with
their existing thresholds and availability treatment. Test both directions:
27 rules x 2 directions = 54 rows, including constant and sparse rules.
No new thresholds, horizons, combinations or production-model changes.

Rank on completed 2026 signal weeks in the frozen research data, January 2
through October 23, with outcomes January 9 through October 30. Do not silently
substitute the latest 52 weeks or include an unfinished outcome. This is the
available 2026 sample, not a full calendar year or the latest possible download.

For a duration view p in {-1,0,+1}, gross score = sum(-p * next_week_yield_bp).
+1 means add duration relative to a chosen benchmark, -1 reduce, 0 no tilt.
Timing score = gross score - mean(p) * sum(-next_week_yield_bp).
The second term is the hindsight score of a constant tilt with the same mean
position. It removes a constant directional bias from the descriptive ranking.
It is not an executable comparator, alpha estimate, portfolio return or net
score. Constant rules have zero timing score even if their gross score is large.

For each rule select the direction with higher 2026 timing score (tie: retain
the raw rule direction). Rank eligible rules by that score. For the watchlist,
require at least 20 active weeks, at least two states each observed >=8 weeks,
and >=4 changes of state. These are pragmatic descriptive filters, not tests
of significance; they were set here before this new ranking was calculated.
Retain all excluded rules and exclusion reasons. Show the first five eligible
rules, even if the older comparisons are poor; do not pick on the comparison.

Apply the 2026-selected direction unchanged to 2023, 2024 and 2025, and to the
January-June and July-October 2026 slices. Each slice removes its own mean tilt.
These backward comparisons are descriptive, not independent validation.
Report active/state counts, changes, gross score, timing score, hit rate,
turnover, sensitivity to removing the largest positive timing week, and each
annual comparison. No post-selection p-values or confidence claims.

Separately calculate same-week correlations of changes in five market series
(US equities, EM equities, spot WTI, the raw six-FX dollar basket and VIX) with
the 10-year yield change over exactly the same weekly intervals. Use unlagged
observation dates only in this explicitly ex-post panel. These are co-movements,
not forecasts or evidence of causation; never feed them into a Friday forecast.
Use log changes for equities and dollar, level changes for WTI and VIX.

Publish all results, this protocol and the script. Current-vintage market
limitations, intraday timing and the paper's publication-lag assumptions remain.
The 2027 watchlist is a list to monitor prospectively; the frozen live model
retains its original rules. No 2027 results exist at the time of this analysis.
