# PRE-REGISTRATION — the overnight drift under three layers: technical, event, behavioural

**Frozen 2026-09-16 09:45 ET, BEFORE any number in this design was computed.** Supersedes nothing; it extends `2026-09-16-overnight-event-nights-*` (events only) and subsumes the narrower `2026-09-16-attention-arm-prereg.md`, which stays on file and is answered by block B1 here.

## Why this study

The desk's tape study (`<data-root>/fp-data/studies/2026-09-15-why-the-tape-loses.md`) refuted the published level-breakout family and found one thing that replicates: **the overnight hold**, NQ +$168/contract net across ~1,660 nights, t 2.27. The event study that followed reproduced the pre-FOMC night (+$973 net, n=53, t=3.12) but found **no event arm explained the rest** — quiet nights still carried ~31% of NQ gross.

So the question is not "is there a drift" (there is) but **what conditions it**. This study asks that with all three layers the desk now has on disk, together, on the same nights:

| Layer | Source | What it encodes |
|---|---|---|
| **T — technical** | `<data-root>/fp-data` GLBX 1m + VIX daily | the tape's own state going into the night |
| **N — non-technical** | `<data-root>/fp-context` events | what was scheduled or had happened |
| **B — behavioural** | claims from `fp-library/Business Behavior Psychology` and `asset-selection` | how retail participants are documented to behave |

## Sample and dependent variable

- **Roots:** NQ primary. ES and GC secondary (reported, not used for the headline verdict).
- **Window:** nights from **2016-01-01** to the last complete night in the archive. 2016 is where the event layer starts; the attention series covers the same span.
- **Night:** prior session 16:00 ET close → next session 09:30 ET open, exactly as `run_overnight_event_nights.py` computes it.
- **Dependent variable:** net dollars per contract for a long held overnight, gap × per-point value − round-turn cost (NQ $14, ES $29, GC $24 — the harness's existing constants).
- **Exclusions:** quad-witching dates (the continuous-contract seam), nights with missing bars.

## Blocks and hypotheses (frozen)

### T — technical state

- **T1 (mirror).** Drift is larger after a **down** RTH session than after an up one. *Lou–Polk–Skouras overnight/intraday tug of war, graded A in the library.*
- **T2 (outsized move).** Drift differs after a prior RTH move larger than **1.5× the trailing 20-day standard deviation** of RTH returns.
- **T3 (volatility regime).** Drift differs across prior-close **VIX buckets: < 15, 15–25, > 25**.
- **T4 (weekend).** Drift differs on **Friday→Monday** nights versus weeknights.

### N — event state (the frozen arms already in `features/context_daily.jsonl`)

- **N1** pre-FOMC · **N2** mega-cap earnings after the close, heavy · **N3** the same, light · **N4** Fed speech after 16:00 ET · **N5** CPI/PCE/NFP the next morning · **N6** EIA/WASDE the next morning (GC/CL only) · **N7** policy or geopolitics headline · **N8** quiet.

### B — behavioural

- **B1 (attention-induced buying).** Nights after a **high-attention day** (`attn_nq_z ≥ 1.0`, Wikipedia pageviews, point-in-time D-1) carry a larger drift. *Barber & Odean's attention-driven buying: retail buys what it notices, and retail buys at the open.*
- **B2 (disposition × attention).** The interaction: high attention **after a down day**. *Disposition effect — holding losers, buying the dip — amplified by attention.*
- **B3 (sensation seeking / chasing).** Nights after a **large up day (> 1.5σ)** carry a larger drift. *Sensation-seeking and overconfidence raise trading activity after wins.*
- **B4 (loss aversion).** Nights after **two or more consecutive down days**. *Loss aversion predicts risk-taking to recover losses; the opposite sign is equally interesting.*

## Method

1. Per-cell mean net $/night, n, win rate, t-statistic, and a 10,000-resample bootstrap 95% interval.
2. One OLS on NQ with all blocks entered together (dummies; VIX as buckets), reporting coefficients with heteroskedasticity-robust standard errors, to separate a block's own contribution from overlap with the others.
3. **Train/confirm split: 2016–2021 and 2022–2026.** A sign flip between halves is reported as refuted, never averaged away.
4. **Multiple testing:** 15 frozen hypotheses on the primary root (T1–T4, N1–N8, B1–B4 minus the GC-only arm). Bonferroni threshold **p < 0.0033**.
5. **Placebo:** (a) 1,000 random-night resamples of equal size for every arm; (b) the attention series shuffled across dates, re-running B1.
6. Everything net of friction. No target, no stop: the night is held flat to flat, as in the tape study.

## Decision rules (frozen)

- **Validated** — significant at the corrected threshold in the full sample **and** the same sign in both halves **and** the placebo distribution does not cover the effect.
- **Suggestive** — passes in the full sample only. Suggestive means *not tradeable*; it goes on the ledger, not into the product.
- **Refuted** — otherwise, and written up as refuted, including the arms I expect to work.

## Frozen parameters

Thresholds are fixed here and may not be re-tuned after seeing results: attention z **1.0**; outsized move **1.5σ**; VIX cuts **15 / 25**; trailing window **20 sessions**; consecutive down days **2**. If a result appears only at another threshold, that is a finding about fragility, not a signal.

## Prior expectation

Low for every behavioural block. The desk has already refuted the event-day premium, COT positioning, net liquidity, gamma-conditioned late-day momentum and the entire level-breakout family. The most likely outcome is that the drift is a **beta-like risk premium that is not conditionable** with what is on this drive — which would itself be worth knowing, because it says: hold the night, stop trying to time it.