# Cage-day classifier — pre-registration (frozen 2026-09-09 14:20, before the run)

## Question
Can factors known by 09:30 predict whether the RTH close lands inside the pre-open gamma walls (a "cage day", base rate ≈ 68 %), and does that prediction turn the walls into a tradable rule?

## Sample
Sessions with a reconstructed pre-open wall map (`fp-data/gex/history`, D−1 book at the 09:10 spot, as validated in the backfill study): 2023-10-11 → 2026-09-08, quad-witching excluded (~693).

## Target
`cage` = RTH 15:59 close inside [put wall, call wall] (walls scaled to NQ by 09:30 open ÷ spot).

## Features (all known before 09:30; no others)
1. `onr` = overnight range ÷ ATR20 (validated expansion predictor F7)
2. `tilt` = net/gross GEX (continuous) and regime dummies long/short
3. `gap` = (open − prior close) ÷ ATR20, signed, and |gap|
4. `vxn_shift` = VXN(D−1)/VXN(D−2) − 1 (B1)
5. `width` = (call wall − put wall) ÷ ATR20
6. `pos` = (open − put wall) ÷ (call wall − put wall), open's position inside the cage (clipped −0.5…1.5)
7. `opex` = opex week dummy

## Model and evaluation
Standardised logistic regression (C = 1), walk-forward: train on all sessions before year Y, test on year Y, for Y = 2025 and 2026 (2024 is training only). Baselines: (a) regime dummies only, (b) width only. Metrics on pooled out-of-sample (OOS) sessions: AUC; cage rate in the top and bottom tercile of predicted probability (terciles set on the training distribution each year).
**Classifier decision rule:** VALIDATED if pooled OOS AUC ≥ 0.65 AND top-tercile OOS cage rate ≥ 80 % AND bottom-tercile ≤ 55 %. Otherwise REFUTED (report AUC regardless).

## Conditional trade rules (evaluated on OOS sessions only, Part A fill rules: next 1m open after a 5m signal close, 0.25 slip, stop max(12, 2×ATR3 5m), target 1.5R, cap 60 min, trigger window 09:35–12:00)
- **C1 cage-day fade**: predicted top tercile → after a 5m close outside a wall, the next 5m close back inside → trade toward the cage (any regime; P8 mechanics).
- **C2 break-day break**: predicted bottom tercile → 5m close beyond a wall → trade with the break (P9 mechanics).
**Decision rule:** mean R > 0 with one-sided p < 0.025 (Bonferroni over 2), positive in both halves of the OOS period, n ≥ 50 triggers. C1/C2 are evaluated only if the classifier is VALIDATED; if it is refuted they are reported descriptively with no verdict.
Descriptive extras: univariate cage rate by feature quintile; the unconditional P8/P9 on the same OOS sessions for comparison. No variants, no threshold tuning.