# Process memo — How Financial Markets Respond to Economic Data Surprises Web commission `how-financial-markets-respond-to-economi-2026-08-12-50ae` · venue BPEA · built 2026-08-12/13 by Paperwright (Claude Fable 5), resumed after the mode-3 emailed round. Companion to the delivered PDF; the paper repo (`paper-how-financial-markets-respond-to-economic-data-s`) is the canonical record. ## What was done, in order 1. **Setup + research (first run, before the mode-3 park):** repo scaffolded; four data sources acquired and profiled (Nasdaq calendar surprises 2013–2026; WRDS fed funds futures; FRED/Treasury yields and breakevens; CRSP sector ETFs); 30-PDF literature review with a Literature Map; the emailed question round drafted from what the research showed actually mattered. 2. **Design (resume):** both question rounds closed with zero submissions, so all six design calls fell to the engine under the reconciliation rule — recorded with rationale in `Brief/Design Decisions.md` (reversion-test spine · daily windows · honest counts + a consensus-free variance co-spine · four regimes with the acute window excluded · GICS-11 sector headline · nominal-headline rates presentation with one decomposition section). 3. **Analysis:** cleaning re-anchored every release to FRED's published calendars, disambiguated MoM/YoY twins against revised series, and standardized surprises excluding the acute window; estimates on the full grid (10 releases × 10 asset outcomes × regimes), the decomposition, the variance spine back to 1999, rolling windows, joint same-morning regressions, asymmetry. The auditor reproduced the pipeline byte-identically before anything was drafted. 4. **Writing:** plan → layered outline (a referee memo on the outline forced the power-honesty, ZLB-confound, and same-morning-collinearity treatments before drafting) → full-density draft → quality pass → two voice rounds against the mined BPEA Furman profile. 5. **Review:** three blind referees + a case-for-acceptance reader + a full source-batched citation audit (round 1), all findings dispositioned (`WIP Docs/Referee Reports/Round 1 dispositions.md`); a second round of three fresh blind referees on the revised draft (round 2 — see the convergence note below); finishing gates (integration, fresh-context claim check, citation diff, anti-AI, spelling, number-auditor, final integrity audit). The build was interrupted once by an account usage limit during the pre-send gates and resumed cleanly; the number-auditor and copy-critic passes were re-run from scratch on resume, yielding Draft v1.5 (ten mechanical must-fixes and a register-rounding alignment, no substantive changes). ## The decisions that shaped the paper, and why - **Unanswered rounds.** Every design question was presented twice (instant + emailed) and answered never; the engine chose and documented. If any call reads wrong, a revision round can rerun the affected leg — the pipeline is one command end to end. - **Hand-repair of August 10 and November 10, 2022.** The Nasdaq feed is missing the cycle's two biggest soft-CPI mornings — a non-random gap all three round-1 referees flagged as capable of manufacturing the equity-asymmetry result. Their consensus/actual values are documented in contemporaneous press (Dow Jones and Bloomberg survey medians; sources shipped in `data/raw/manual_consensus_repairs.jsonl`) and were injected as ordinary observations. **The repair changed a finding:** the "relief rallies were negligible" reading of the unrepaired sample was an artifact; the paper now reports real soft-print rallies (+0.56 percent per σ) against larger hot-print losses (−1.35), and says exactly why. - **Beta-adjusted sector cross-section.** The raw sector gradient is mostly market beta (round-1 methods referee); the paper now reports market-beta-adjusted responses and reframes the mechanism claim — the one significant abnormal response is financials (+0.18, the rate-margin channel). - **Small-sample inference.** All headline p-values are wild-bootstrap; leave-one-out ranges bound single-day influence; the persistence claim leads on the consensus-free leg with the ZLB-benchmark caveat stated at first use. - **What the paper claims and does not.** Full reversion rejected on both legs; zero-vs-partial reversion not distinguishable at these cell sizes; perceived-rule story argued as the parsimonious reading, with attention-tracking-the-mandate and signal-weighting conceded as observationally close on this sample and the separating tests named. ## What the referees converged on (both rounds) Round 1 consensus: the missing-days repair; benchmark honesty on the reversion test; small-sample inference; the sector beta adjustment; the 2024–26 institutional record (2025 framework revision, BLS stress, shutdown-delayed releases); a conditional Implication 1. All implemented. Round 2 (three fresh blind referees) converged on one big thing: the persistence headline was a core-CPI-only fact, with headline CPI, core PCE, and PPI showing large partial reversion in the paper's own tables. **The paper's headline claim was restated accordingly** — a new composite inflation-news response shows the complex returned roughly three-fifths of its repricing (wild p 0.019) while core CPI alone shows no decline — together with the corrected consensus-free contrast (rejection of reversion to the 2016–19 multiple at p 0.027), the within-normalization split (4.07 in 2024 → 1.53 in 2025–26), a full missingness audit with population denominators (post-COVID coverage ≈ one-third; covered mornings ~2x the uncovered ones), the TIPS-indexation mechanical floor (which strengthens the anchoring conclusion), and family-wide interaction tests. Details: Referee Reports/Round 2 dispositions.md. ## Honest limitations - Consensus cells are small (core CPI: 14 in the fight, 10 in the normalization) and the pre-COVID core cell is ZLB-tail only; every count is printed and the variance leg carries the years the survey cannot. - Daily windows: pre-announcement drift is intercept-absorbed, not excluded; intraday identification is the natural upgrade. - The sector panel ends December 2024 (CRSP subscription), so the sector-level reversion test is not runnable here. - The feed's consensus as-of-ness cannot be audited internally; external checks are Cline-Jia agreement and the two hand-collected mornings. - FOMC days remain in the variance-spine baseline (bias direction stated; a full 2013–26 meeting-date history was not available to this build) — leave-one-out ranges bound the one known overlap day. ## Where more or better data would change the answer 1. **A licensed consensus history (Bloomberg or Action Economics).** Fills 2016–Feb 2020 core CPI outright, retires the ZLB-benchmark caveat, and doubles most post-COVID cells. The single highest- value acquisition; requires an institutional license the build did not have. 2. **TAQ intraday windows** (available on the same WRDS subscription; heavy pull): sharpens every estimate, extends the sector panel past 2024, and enables the sector-level reversion test. 3. **An FOMC meeting-date history** (public; an afternoon's collection): removes FOMC days from the variance baseline and supports the days-to-meeting interaction test. ## Specific asks a co-author could fulfill - Confirm the Furman (2022) op-ed characterization against the full WSJ text (the audit verified the accessible excerpt; the draft was softened to the verified wording). - The round-1 discussant's best idea, recorded here as the priority extension: chain Bauer-Pflueger- Sunderam's perceived-rule estimates through a signal-extraction step to a **rule-implied CPI beta** by year — the residual against the measured beta is a unit-consistent estimate of what attention contributed at the 2022 peak. This would convert Section VIII's adjudication-by-parsimony into a decomposition. - Dating the 2024–26 payroll-beta re-emergence against the FOMC communication record (Implication 2 now deliberately makes no timing claim). ## Verification state at delivery Reproducibility passport: 181 claims, `passport.py check` clean at delivery; `run_all.sh` reproduces `output/` end to end from the raw pulls (the data auditor re-ran it independently, byte-identical); environment and package pins in `replication/`; typography gate PASS on the delivered PDF (BPEA manuscript format, 54 pages); anti-AI linter and American-spelling gate clean; final integrity audit APPROVE (2026-08-13). Delivered: the manuscript PDF and this memo.