# PRE-REGISTRATION — Do macro events, surprises and administrations predict NQ? **Frozen: 2026-08-25 20:25 MDT, BEFORE any result was computed.** Study #3 today. Predecessors: `2026-08-25-rvol-orb-*` (REFUTED), `2026-08-25-cross-instrument-*` (REFUTED). Data: `/fp-data`, synced through 2026-08-25 23:59Z. --- ## 1. Scope, and an honest limit on the word "news" The request is to predict NQ from **news, events and administrations**. Only part of that is answerable with the data on this desk, and the boundary is stated before running: - **Events — fully answerable.** `fmp/econ-calendar-2016..2026.json` carries point-in-time consensus *and* first-print actuals (vintage-verified in `manifest.json`), so surprises are computable without lookahead. - **Administrations — answerable but confounded.** See §6. - **"News" in the broad sense — NOT answerable.** This desk has no archived, timestamped news-sentiment series (no RavenPack/Bloomberg feed; the Desk Read consumes news live and does not store it). So "news" here means **scheduled macro releases + FOMC decisions only**. Unscheduled news (geopolitics, headlines, single-name shocks) is out of scope and no claim is made about it. The library's grade-A news results (Tetlock; Boudoukh et al.) are therefore **not testable here** — recorded as a data gap, not as a null. ## 2. Prior desk work this must not duplicate `fp-terminal/scripts/backtest-surprise.mjs` already tested pre-open surprise → NQ at **1m / 5m / 15m** from the open (1,589 classified events, 2016–2026). This study is additive on three axes it never touched: the **full RTH session** as target, **event-day vs non-event-day** distributions, and **administration conditioning**. Its `FAMILIES` table and `econSign` convention are **reused verbatim** so numbers stay comparable. ## 3. Data and definitions | Item | Spec | |---|---| | Price | `glbx/NQ/1m/*.csv`, ET via `zoneinfo`, ≥330 RTH bars, quad-witching `.c.1` rule (as Studies #1–2) | | Calendar | `fmp/econ-calendar-{2016..2026}.json`; timestamps stored **UTC** (verified: FOMC 18:00/19:00Z = 14:00 ET) | | Surprise sample | 2016-01-01 → 2026-08-25 (calendar coverage) | | Event-day / administration sample | 2010-06-07 → 2026-08-25 (full price history; FOMC dates derivable) | **Families and signs — copied from the shipped desk script, not re-chosen:** core_cpi_mom −1, cpi_mom −1, core_pce_mom −1, pce_mom −1, core_ppi_mom −1, ppi_mom −1, retail_mom +1, durable_mom +1, claims −1, nfp 0, unemp 0. `econSign` = expected NQ direction for a *positive* surprise. Sign 0 = ambiguous (good-news-is-bad-news); reported standalone, **excluded from the composite**. GDP is excluded (FMP carries revisions as separate rows — vintage-contaminated), same as the desk script. **Surprise standardization — stricter than the existing desk script.** The shipped script divides by the *full-sample* family σ, which is a mild lookahead. Here `z` uses an **expanding, strictly backward** σ (minimum 20 prior observations in that family). The full-sample version is reported only as a robustness check for comparability with the existing numbers. **Composite**: `signed_z_d = Σ econSign · z` over non-ambiguous families released **before 09:30 ET** on date d, then standardized (expanding). Pre-open restriction guarantees the signal is knowable at the open. **Targets** (all in bps, `×1e4`): - **`ret_rth` = close(15:59) / open(09:30) − 1** ← primary - `ret_5m`, `ret_15m` (continuity with the existing desk study) - `ret_preopen` = open(09:30) / prior 18:00 ET − 1 (for the drift test) **Administrations** (inauguration boundary, 20 January): Obama-II → 2017-01-20 · Trump-I 2017-01-20 → 2021-01-20 · Biden 2021-01-20 → 2025-01-20 · Trump-II 2025-01-20 → present. ## 4. Primary test — ONE pre-specified cell > **Does the pre-open composite macro surprise predict NQ's RTH open-to-close return?** > OLS `ret_rth ~ signed_z`, full sample 2016–2026. > Statistic: **t on the slope**. **Pass threshold: |t| > 3.0** (Harvey–Liu–Zhu; the cumulative > pre-counted family across the three studies is now large — see §7). Sign is *not* pre-specified as positive: a reliable negative slope (good news sells off) would be equally interesting and equally tradable. Two-sided. ## 5. Secondary tests (reported, never confirmatory) 1. **Event-day effect (Savor–Wilson on NQ).** Mean `ret_rth` on tier-1 event days vs all other days, 2010–2026. Savor–Wilson's result is close-to-close *including* overnight; the RTH-only version is what a day trader can actually capture, so **both** are reported and the difference is the point. 2. **Pre-announcement drift.** `ret_preopen` on the night before a tier-1 release vs other nights (Lucca–Moench / Hu et al. analogue). Includes the pre-FOMC 24-hour window. 3. **Per-family slopes** — which single indicator, if any, moves NQ's session. 4. **Directional hit rate** of `sign(signed_z)` on `sign(ret_rth)`, and bps/trade after **1.0 pt** friction (the desk `COST_PTS`), i.e. is any of it tradable. 5. **Administration conditioning** — §6. 6. Year-by-year sign consistency; IS/OOS split at 2020. ## 6. The administration test, and why it is descriptive only Administration will be crossed with the primary and with the event-day effect. **It cannot support a causal claim, and none will be made.** Stated in advance: - **n = 4 regimes.** Any "administration effect" is one draw per regime, however many sessions it contains; the effective sample size for the *administration* claim is 4, not 4,000. - **Total confounding.** Administrations are collinear with the Fed cycle (ZIRP → hiking → cutting), the inflation regime, COVID, and the secular AI/mega-cap trend in the Nasdaq specifically. Trump-I contains the 2018 vol regime; Biden contains the 2022 bear market; Trump-II is ~19 months long. Nothing here can separate "administration" from "everything else happening at the time." - **Trump-II has ~400 sessions**, far too few for a stable estimate, and is reported with that caveat attached rather than as a finding. So the administration section answers a **descriptive** question — *did the macro→NQ relationship look different across these eras?* — which is legitimately useful for regime awareness, and explicitly not *does the party in power cause NQ behaviour.* Any apparent difference will be reported alongside the Fed-cycle and VIX-regime context that plausibly explains it. ## 7. Multiple-testing accounting — cumulative across today's three studies Studies #1–2 pre-counted 368 cells. This adds: 4 targets × 11 families × 4 administrations ≈ **176**, running family **≈ 544**, with 3 confirmatory cells total. The t > 3.0 bar is retained. With ~544 trials, the expected maximum |t| under the pure null is ≈ 3.3, so **even a t of 3 is not comfortable** and anything at t ≈ 2 is noise. Deflated-Sharpe reasoning per `statistical/classical-econometric-selection.md` §3.25. ## 8. Falsification criteria (decided in advance) **H1 (primary)** CONFIRMED only if |t| > 3.0 on the composite slope **and** the sign is stable across the IS/OOS split **and** the directional strategy is positive after 1.0 pt friction. Pass slope but fail cost/stability → **contested**, shadow-only. Fail → **REFUTED**: pre-open macro surprises do not predict NQ's session on this desk. **H2 (event-day)** CONFIRMED if event-day mean `ret_rth` exceeds non-event-day at t > 3.0. **H3 (administration)** — **no confirmatory bar is set, by design.** With n = 4 regimes any threshold would be theatre. Reported descriptively with confidence intervals and confounds. ## 9. Known limitations, stated before results - Calendar starts 2016 → the surprise tests get ~10.5 years, not 16. - FMP consensus is a single vendor's snapshot, not a whisper number; Study limitation, not a bug. - Pre-open restriction excludes 14:00 ET FOMC decisions from the *composite* (they cannot be known at 09:30). FOMC is therefore handled separately in the event-day and drift tests, not in the primary. - Intraday event reactions (the 08:30 release → 09:30 open path) are already the desk's existing study's territory and are not re-run. - Results are hypotheses. This study can remove candidates; it cannot certify an edge.