Deviations from the frozen design (docs/DESIGN.md, git tag design-frozen, 2026-11-09)
Timing convention: "before results" = decided before any H2–H5 output was inspected; "after H1" = decided after the H1 fidelity numbers (which the design allowed to be inspected first) but before H2–H5; "after results" = decided after seeing H2–H5.
| Date | Change | Rationale | Decided |
|---|---|---|---|
| 2026-11-09 | Design review by an econometric-referee agent inserted between data construction and the event study (docs/DESIGN_REVIEW.md). |
User request for an integrated review. | before results |
| 2026-11-09 | Inclusion rule relaxed to ≥ 100 minutes on ≥ 85 % of window days (sets now ZT ZF ZN US → +UB 2013 → +TN 2019); before the fix the rolling weights used only ZF/ZN/US (2011Q1–2024Q1) and ZT/ZF/ZN/US (2024Q2–), because of the ≥ 300 minutes on ≥ 95 % of training days rule; TN and UB never qualify. Not a decision but a consequence of the frozen rule discovered at documentation time. | Rule as implemented in 04_mimicking.py. |
n/a (documented after results) |
| 2026-11-09 | Added GSW 1–30y and CMT ≥ 1y targets as robustness (04b). GSW was pre-specified; CMT ≥ 1y was not. |
Isolate the bill-segment contribution to the slope shortfall. | after H1 |
| 2026-11-09 | Added SR3 (3-month SOFR future) as a seventh instrument from 2021Q1 (01b, 04b). Extension, clearly labelled. |
Futures ≥ 2y cannot span the short end; review Threat 1 fix (b). | after H1 |
| 2026-11-09 | Tested CMT 2–30y target (04c) following review item 1; rejected as ill-conditioned (slope/curvature SDs 15.5/28 bp, corr −0.75, OOS slope R² 0.53). Primary target unchanged. |
Review recommendation checked, not followed. | after H1 |
| 2026-11-09 | Added cash-yield spanning ceiling (2/5/10/30y) under the same OOS protocol (04c). |
Review Threat 1 diagnostic. | after H1 |
| 2026-11-09 | Added leave-one-contract-out PCA spanning test at 1/5/30 min (04c). |
Review Threat 2 (genuine intraday validation). | after H1, before H2–H5 inspected in detail |
| 2026-11-09 | Added independent ETF-based factor construction (duration-based, 06b). First version (daily-regression weights on regular-hours ETF returns) was discarded because ETF regular-hours returns do not match the 15:30→15:30 yield day; replaced by the structural construction. |
Review Threat 2 (ii). | after H1; final version after first H2/H4 outputs existed |
| 2026-11-09 | Added tick-noise-implied factor correlations (06c). |
Review Threat 3 / H4 guard. | after H2/H4 first outputs |
| 2026-11-09 | 5-minute sampling used as main frequency for H4 and as the HAR target in H5; 1-minute kept as robustness. | Review Threat 3. | before H5 was run; H4 1-min values had been seen |
| 2026-11-09 | Event study: placebo windows at the same clock time on matched non-event days; event-level bootstrap instead of daily blocks; mean and median shares; extra windows [−5, +15] and [−5, +90]. | Review §2.10. The placebo and bootstrap were in the first run of 07; the extra windows were added in a second run after the first event results had been printed. |
placebo/bootstrap before results; extra windows after results |
| 2026-11-09 | Short rate y(0) = L + S and implied 3m/2y/5y/10y/30y changes added to the event outputs. | Review §2.2 (level/slope not separately identified in the ZLB). | with the first run of 07 (before results) |
| 2026-11-09 | Holm correction on the Diebold–Mariano family in the paper tables. | Review §2.9. | after H5 results |
| 2026-11-09 | H5 re-estimation is monthly on an expanding window (08), as the design's H5 text said; the design's §3.2 "quarterly" refers to the mimicking weights only. Clarification, not a change. |
Resolve the internal inconsistency noted by the review. | before H5 |
| 2026-11-09 | Positioning changed to a measurement paper (spanning bounds, validated series, variance accounting by clock time and event); event-window language descriptive. | Review §3. | after results |
| 2026-11-09 (evening) | Return panel rebuilt from individual contracts (01d, new 03): most-traded contract per CME session, within-session returns only, no roll splice; replaces the vendor continuous series whose bars are sparse before expiries (a third of 08:30 releases in roll months were lost since 2018). Event counts 126/187/190. |
Referee report E1 (econometric). | after results and after referee reports |
| 2026-11-09 (evening) | Fidelity of the cumulated minute series over all 3 963 days (06 → fidelity_cumulated_intraday.json: 0.93 / 0.80 / 0.75) added next to the daily-regression fidelity; eq. (3) reworded. |
Referee E1: H1 did not test the object used. | after results and after referee reports |
| 2026-11-09 (evening) | Implied-yield fidelity (3m 0.09, 1y 0.64, 2y 0.86, 5y 0.92, 10y 0.94, 30y 0.92; L+S vs 1-month bill −0.05) computed in 12_text_macros.py; short rate dropped from headline results and from the event tables' interpretation; both readings of the spanning gap (bill segment for the primary target, futures–cash basis for a 2–30y target) reported. |
Referee E2. | after results and after referee reports |
| 2026-11-09 (evening) | Dynamic analyses added (13, 14): local projections at 25 min–10 days on the 2-year window move; five-minute VAR with generalized IRF/FEVD and Diebold–Yilmaz connectedness by regime; HAR-VAR volatility connectedness (full and 250-day rolling). Not pre-registered. |
User request ("IRF et tout le kit") and referee suggestions. | after results |
| 2026-11-09 (evening) | H5 reported under both pre-registered losses: QLIKE (7/9 Holm-significant at 5 %) and MSE (1/9); conclusion "supported under QLIKE only". | Referee E3 / integrity 2. | after results |
| 2026-11-09 (evening) | Robustness claims that had no metric (boundary shift, expanding window, 5-minute denominators, release-time check) computed in 04d and quoted; unsupported sentences removed; citations corrected; Pesaran–Shin 1998 and Diebold–Yilmaz 2012 added (Crossref-verified). |
Integrity and econometric reports. | after results |
| 2026-11-09 (evening) | Exhibits: Okabe–Ito palette, bootstrap bands, heatmap and cartography figures, table layout fixes, "Appendix Table" wording. Numbers unchanged by these edits. | Exhibits report and figure mandate. | after results |
Unchanged from the frozen design: 11-maturity CMT target with λ = 0.7308 as primary; 15:30→15:30 yield day; quarterly 3-year rolling weights with ≥ 500 days; no intercept at minute level; H1–H5 statements and thresholds; sub-periods; two-sided 5 % tests; §7 falsification logic (not triggered: level and slope OOS R² ≥ 0.6).