A Simple Weekly Signal
for the U.S. 10-Year Treasury Yield

Appendix C. Data, calculations and reproduction

Timing. ALFRED dates identify source, provider or archive availability; they are not a guarantee of each original intraday release. Core PCE growth uses the latest month and the same month a year earlier in one snapshot, including revisions already present there. The unemployment average excludes the current month and uses 60 distinct earlier available months. Before there is enough history, the rule is neutral. Foreign-yield rules are neutral before June 2013. The early model therefore has fewer than thirteen informative inputs.

Archive checks. I reject nonfinite values, duplicate or nonconsecutive months, observations ahead of their vintage, very old latest observations and backwards jumps in the latest month. One German snapshot dated August 1, 2013 ended in December 1969 and is excluded. No revised-data fallback fills the early foreign-yield gaps. Current market histories can still contain revisions or backfills. Treasury reference-yield quotations are around 3:30 p.m., while some other inputs close later.

Calendar rules. The supplied Fed dates are scheduled decisions, but historical announcement-time schedule revisions have not been fully audited. The jobs rule means that next Friday falls in the first seven days of a month. It is a calendar proxy and never reads future realized jobs-release dates. Neither event rule is in the thirteen-rule specification.

Training and scoring. Continuous inputs use an expanding mean and standard deviation, including the current observable value, with at least 52 observations; standardized values are clipped at four and missing values become zero. December and the selected auction flag stay 0/1. Four-week curve changes are rounded to ten decimal places in percentage points before taking a sign, so exact quoted ties stay neutral. At each refit, the regression uses only training weeks whose outcomes have ended by that Friday. Unchanged targets are classified as not down. HAC statistics use an intercept-only regression with Bartlett weights and four lags; eight-lag results are also shown. Volatility and SR use all calendar weeks and annualize with 52. Wilson hit-rate intervals assume independent observations and do not adjust for selection.

Sample boundaries. The original tables define the early period by signal date: the last early signal is October 28, 2016 and its outcome is November 4, before that Friday's next forecast. This convention also applies to the initial two-addition search. The two broader searches impose a stricter information cutoff: all ranking outcomes end before November 1, leaving 564 early signal weeks. The later evaluation has 521 completed weekly outcomes. Circular-shift diagnostics use offsets of at least 52 weeks and compare unrounded statistics; their reported fraction is (k+1)/(N+1).

Reproduction. The package contains the research model, downloader, input-sensitivity check, weekly script, frozen specification, Fed-date file, dependencies and manuscript builder. The improvement_study folder contains both protocols, both search scripts and every trial. The watchlist_2027 folder contains its dated protocol, all 54 direction trials, every rule's ranking and selected weekly scores. The economic_review folder contains interpretation notes for all 27 rules and the post-discussion reversal and curve checks. The auction_study folder contains the additional protocol, official calendar records, announcement audit, timing tests and all eight variants. The separate refit comparison is in improvement_study/refit_comparison.json. README.txt gives the commands. Downloads create a new snapshot and do not overwrite the archived local study inputs. A research run cannot overwrite the production specification. Raw market and macro files are excluded; their hashes identify the archived study inputs. A later download can change the historical results and need not reproduce the paper exactly.

Appendix D. Broader searches and the 2027 watchlist

The five input sets are: the main thirteen; those thirteen without the three foreign yields; seven inputs covering US and EM equities, oil, dollar, VIX, curve and December; thirteen plus breakeven inflation; and thirteen plus the jobs proxy and 300-day yield trend. The last two are named candidates carried into this search, not new discoveries independent of earlier results.

The classification grid uses C = 0.01, 0.1 or 1; all history, 520 weeks or 260 weeks of training; 52- or 13-week refits; minimum absolute checklist scores of 0, 1 or 3; probability distances from 0.5 of 0, 0.025 or 0.05; and three sizing rules. Probability sizing is min(abs(p-0.5)/0.10, 1); checklist sizing is abs(score)/number of rules. The existing combination supplies the direction. These choices give 2,430 settings.

The ridge grid uses the same five input sets, training windows and refit intervals; alpha = 1, 10, 100 or 1,000; with or without the checklist non-conflict condition; minimum absolute predicted moves of 0, 1 or 2 bp; and either a unit position or min(abs(predicted move)/5, 1). These choices give 1,440 settings. Positive forecasts mean lower yields.

For each grid, I rank the early calendar-week SR after a 0.25 bp charge per position unit. A candidate must be active in at least 25% of weeks, average at least 0.10 absolute exposure, and have positive cost-adjusted scores in both 2006-2010 and 2011-October 2016. The overall winner and the best setting within each input set are saved before their later results are calculated. The table reports every input-set winner, including failures. The second grid was specified after reviewing the first grid's result.

Detailed parameter choices, annual blocks, paired differences and full weekly positions are in the accompanying JSON and CSV files. The first grid's overall winner has a paired later-period difference of -15.6 bp/year versus the annual reference model (t = -1.22); the ridge winner's difference is -20.0 (t = -1.12). Neither this table nor a larger search can turn already-inspected history into a prospective test.

2026 timing attribution. With p denoting the duration view and dy the following week's yield change in bp, the total timing score is sum[-(p - mean(p)) dy]. Each displayed year or subperiod uses its own average tilt; subperiod timing scores need not add to the full-window timing score. The constant comparator is constructed after observing the window. No t-statistic is attached to this selected ranking. The 2023-2025 columns use the 2026-selected direction unchanged, with years assigned by signal date. Same-week comparisons use unlagged observation histories, log changes for equities and dollar, and level changes for WTI and VIX; they are never forecast inputs. Monthly foreign yields can repeat one reading for several Fridays, so week counts overstate independent information.

Appendix E. The economic reasoning behind the votes

Start with the horizon. A release's level, its surprise and the price response already behind us are different observations. A weekly forecast asks what happens next. The same information can motivate continuation if adjustment is incomplete or reversal if prices overshot. The sign alone cannot identify which occurred. News research typically measures the surprise relative to prior expectations (New York Fed); this checklist does not observe that surprise.

Curve compensation. The add-duration vote after steepening is motivated by the possibility that the extra yield farther out the curve attracts demand. Carry and roll are the intended intuition. But a change in 2s10s does not calculate either return, and a constant-maturity ten-year yield does not follow a bond as it ages. Steepening can occur with rising yields, falling yields, or a curve that remains inverted. It does not by itself imply policy easing. Curve positions and outright duration are distinct choices (CFA Institute).

Inflation and employment. Lower core inflation and unemployment above its recent norm are treated as conditions that may favor a lower future policy path. The fixed 2% core-PCE cutoff is a research convention. The formal Fed target concerns headline PCE; the numerical goal was announced in 2012, and maximum employment has no corresponding fixed numerical goal (Federal Reserve). The five-year unemployment average does not estimate the natural rate. Neither rule measures the next policy decision's surprise.

Yield compared with past inflation. I retain nominal yield minus trailing core inflation as a simple comparison with experienced inflation. The ten-year yield and past twelve-month inflation have different horizons. Its sign does not establish expected purchasing-power returns or whether a bond is cheap. The retained directional vote must stand on its empirical results.

Market prices and risk. Equity gains may signal easier conditions or stronger growth, motivating higher-yield votes. They do not demonstrate money leaving bonds. The dollar vote instead hypothesizes tighter conditions and less inflation pressure; a repricing toward higher US policy rates can lift both the dollar and yields. VIX is option-implied equity volatility (Cboe), not Treasury buying. Its add-duration vote assumes a demand-for-safety channel that need not dominate inflation or liquidity shocks.

Energy and international rates. Rising energy prices motivate a continuation hypothesis about inflation pressure, while the 2026 watchlist explores different subsequent associations. Foreign-yield votes assume common rate forces or spillovers can persist. Archived monthly foreign yields are lagged states, not contemporaneous market quotes, observed Treasury purchases, or a demonstrated lead.

Calendar and additional readings. December is a seasonal hypothesis; the calendar does not measure liquidity or imply a sell-off. Its later record contains 45 weeks from 10 Decembers. Fed and first-Friday flags identify events, not the signs of surprises. The selected below-300-day-average rule bets on a yield rebound, so its name should not be mistaken for trend following. Breakeven inflation includes risk and liquidity premia; total Fed assets include different balance-sheet operations. Volatility and price-pattern candidates similarly require explicit hypotheses. All 27 are discussed in the accompanying notes.

From a view to a portfolio. A higher probability of a yield fall need not imply a positive expected score if the less frequent rises are larger. For example, 60% odds of a 1 bp fall and 40% odds of a 3 bp rise give a long-direction expected score of -0.6 bp. The classifier's 50% cutoff is a direction rule, not an optimal duration allocation. A ten-year point forecast also does not measure a portfolio's other maturity exposures, credit spreads, carry, financing or execution costs.

Appendix F. Does next week's long-bond auction help?

An upcoming auction is information an investor can know before acting. The hypothesis is that investors may require a yield concession to absorb supply. That motivates a reduce-duration vote, but an auction can also remove uncertainty or attract demand. This weekly test covers both the period before the auction and the adjustment afterward. It cannot identify the mechanism or an auction-day trading return.

I specified four flags before calculating their results: a 20-year nominal auction, a 30-year nominal auction, a 30-year TIPS auction, or any of the three. Reopenings count by original maturity. A flag is one only when the auction falls after the signal Friday and no later than the following Friday, and its official announcement is dated strictly before the signal Friday. Same-Friday announcements are excluded. I use no auction result, bid-to-cover ratio, tail, eventual demand or backfilled tentative schedule. Each flag is tried as one add or reduce vote alongside the thirteen rules and as a 0/1 regression input. All eight combinations keep annual refits and unit positions.

The early ranking uses the highest calendar-week gross t-statistic, with at least 20 early event weeks. Settings and results are saved before the later comparisons. This is a disclosed November 2026 extension after seeing the earlier research, not an untouched holdout. The 20-year nominal bond returned in May 2020 and has no early events, so neither of its signs qualifies for early selection. The any-auction flag changes composition when that maturity enters; its early history tests 30-year nominal bonds and TIPS only.

The conditional sign is not stable: nominal 30-year auction weeks have falling yields on average in the early sample and rising yields later. The TIPS comparison has only 20 early and 22 later event weeks, including just two in 2026. It cannot support a strong standalone claim.

The selected any-auction, reduce-vote variant improves the early gross score from 18.3 to 29.4 bp/year. Later it scores 33.5 gross and 28.4 after the assumed charge. Its paired later gain is 11.0 bp/year (t = 1.81). In the last 52 weeks it scores 69 bp, versus 58; over the 43 completed 2026 weeks, 72 versus 58. The constant higher-yield benchmark scores 31.4 bp/year over the full later period. Against that benchmark, the selected variant's paired gain is 2.1 bp/year (t = 0.07). The close benchmark and the additional search temper the result.

Both vote signs can improve a combined model because the new input also changes the regression's coefficients and its views on non-auction weeks. That is not evidence for two opposite economic explanations.

Simply reducing duration in any-auction weeks has a larger later score than the combined model, but loses in the early sample. That reversal and the introduction of the 20-year bond prevent treating the later number as a stable edge. The accompanying files also report hit-rate ranges, eight-lag checks, turnover, all early trials and weekly positions.

Source and limits. I retrieved identifying and calendar fields from the Treasury auction database on November 7, 2026. An outcome-independent sample of 23 original Treasury announcements confirms their announcement dates, auction dates and CUSIPs. This checks a sample, not every announcement or historical calendar revision. The data retain that limitation, as well as the paper's market-history limitations. The 20-year restart is documented by the New York Fed. This annual any-auction variant is retained as a research comparison. Appendix G tests all additions against the subsequently selected quarterly baseline and identifies the adopted TIPS variant.

Appendix G. Auction additions with quarterly refitting

After inspecting the annual auction results and choosing quarterly refitting, I repeated the same eight additions with 13-week refits for both the baseline and augmented models. The dates, flags, votes, early ranking and minimum event count stay unchanged. The second-round protocol was recorded before these calculations. This adds eight trials to the first eight; it is informed by earlier research and is not an independent confirmation. Standalone event-only scores are identical to Appendix F.

The early winner adds a nominal 30-year auction flag with an add-duration vote. Its early score rises from 15.4 to 23.0 bp/year. Later it scores 26.3, against 30.6 for quarterly refitting alone: a paired difference of -4.3 bp/year (t = -0.72). Its hit rate is slightly higher, 52.8% versus 52.6%, illustrating why hit rate alone cannot determine usefulness.

The two TIPS variants score 32.1-32.6 bp/year, with paired t-statistics of 0.99-1.27. Neither is the early winner. After reviewing the later results, I adopt the add-vote variant in version 4: 32.6 bp/year, or 2.0 above the quarterly baseline (paired t = 1.27). There are only 22 later TIPS event weeks. The addition also changes fitted weights in non-event weeks, so its effect is not an auction-specific causal estimate.