Improvement searches - November 7, 2026

I tested 3,870 settings in two disclosed rounds. Each round selected its winner
using only outcomes that ended before November 1, 2016. The later history had
already been inspected during prior research, so it remains a retrospective
evaluation rather than an untouched holdout.

The searches did not identify a replacement for the existing reference model.

                              Early gross   Later gross   Later gross t
                              yield bp/yr   yield bp/yr
Existing 13-rule model             18.3           22.6            1.20
Classification search winner       44.8            7.0            0.46
Expected-move search winner        23.4            2.6            0.17

The early sample here ends one signal week earlier than the original paper's
signal-date convention. It excludes the November 4 outcome. That explains the
baseline early score differing from Table 1 in the paper; the later sample is
identical. Exact values are in comparison.json and ridge_comparison.json.

ROUND ONE: 2,430 settings
Five input sets, three training windows, three regularization strengths, two
refit frequencies, three minimum checklist scores, three probability margins
and three position-sizing choices. Positions were fixed, probability-scaled or
checklist-scaled, always capped at one unit. Fractional positions were scored
at their actual sizes, not converted back into directions.

Ranked on early calendar-week SR after an illustrative 0.25 yield-bp charge per
position unit, subject to minimum activity and positive results in both halves
of the early sample. The winner used fifteen rules, a ten-year training window,
C=0.1, 52-week refits, a 45%/55% probability filter and fixed positions.
Its cost-adjusted score fell from 41.9 to 2.6 bp/year afterward. The later paired
gross difference versus the reference model was -15.6 bp/year, t=-1.22.
None of the five input-set winners improved the later average or cost-adjusted SR.

ROUND TWO: 1,440 settings
Specified after seeing the first round. The hypothesis was that predicting the
size of a weekly yield move might align better with the reported score than
predicting only its direction. I used simple ridge regression on the same five
input sets. The grid tested training windows, shrinkage, refit frequency, whether
the checklist could veto the forecast, minimum predicted moves and fixed versus
forecast-scaled positions.

The winner used fourteen inputs including breakeven inflation, five years of
training, alpha=1, a 52-week refit, checklist non-conflict, a minimum 2 bp predicted
move and fixed positions. Its later cost-adjusted score was -0.6 bp/year. The
paired gross difference versus the reference model was -20.0 bp/year, t=-1.12.
The market-only family had no setting meeting the early-sample eligibility rules;
all four other family winners failed to improve the later average or net SR.

WHAT THIS SAYS ABOUT SCALING
Neither overall winner used scaled positions. The existing classifier's later
Brier loss is 0.2514, versus 0.2500 for always forecasting 50%; lower is better.
That does not establish that more extreme probabilities should receive more
exposure. Brier loss combines several aspects of probability-forecast quality,
so it should not be treated as a pure calibration statistic.
See https://scikit-learn.org/stable/modules/calibration.html

WHAT CHANGED IN THE PAPER
The draft is now in the first person, with the market question and intuition up
front. Implementation details sit mainly in the appendices. The searches, all
input-set outcomes, scaling rules and selection history are explicit. No personal
earlier-work references are included. The current main model's forecasting rules
and position sizes are retained. Costs are deducted from the unrounded score
before reporting the final 17.6 bp/year figure. A subsequent arithmetic correction
makes an unchanged quoted curve neutral instead of accepting subtraction dust as
a directional move. Both searches have been rerun with that correction; their
overall winners are unchanged. All figures above include the correction.

These failures do not prove that every possible variation will fail. They show
that these specific in-sample winners did not generalize to the later history.
Further threshold searches on these same inputs would add selection risk. A next
study should test a new economic hypothesis or better historically available
information, and reserve a genuinely new evaluation.

Files:
PROTOCOL.txt and SECOND_PROTOCOL.txt: choices fixed before each search round.
all_in_sample_trials.csv and ridge_in_sample_trials.csv: every trial.
selection_before_later_evaluation.json and ridge_selection_before_later_evaluation.json:
  chosen settings saved before their later-period calculations.
comparison.json and ridge_comparison.json: all selected-model comparisons.
shortlist_weekly.csv and ridge_shortlist_weekly.csv: actual position sizes.
probability_bins.csv: descriptive probability bins for the baseline classifier.
search.py and expected_move_search.py: runnable code.
