Skip to content

KB/signal

GEX Flip Risk

Last verified

Auto-generated. This article is rebuilt from app/signals/config/signal_definitions.json by scripts/build_signals_kb.py. Edit the registry entry and re-run the script — do not edit this file directly.

GEX Flip Risk

What it is

GEX Flip Risk — registry key gex_flip_risk.

Evening gex sign-flip risk STATE (SIGPRED #87 G1) — a deterministic, hand-thresholded 0/1/2 state computed from the flip study’s published conditioning stats, never a fitted model or probability output. State 2 (primed) = |gex| compressed inside 0.5B of zero AND the change vs the prior distinct vendor print points across zero — the configuration the gex flip study (2026-06-11) conditions a 31% next-day sign-flip rate on, vs the 17% unconditional base. Not scored / not aligned, and no surface acts on it: the gex_flip_primed standing tripwire wire was DEACTIVATED 2026-07-17 when the prereg’s kill bar fired live (forward AUC 0.6957 < the 0.70 bar, n=3,857; the state-2 precision multiple held at 6.1x — recorded, not a bar). The state keeps accruing and the standing nowcast-grading gex_flip carrier keeps grading it forward (capture-as-hedge); it re-earns a surface only via that standing verdict. Complement to the gex_sign transition surface: this state describes the evening BEFORE a possible flip, the transition records the flip AFTER it happens. Prereg: research/planning/prereg_gex_flip_risk_2026-06-12.md.

Data-pyramid tier

T1 — Signal. This metric sits at the signal layer of the platform’s six-tier data pyramid (T0 raw inputs → T5 actionable read). The tier reflects how far the value is from a raw measurement — not how strongly it is validated. Abstraction and validation are separate axes: a higher tier is not a claim of stronger evidence.

Horizon & validation

Horizon: 1d — the window this metric’s measured skill speaks to. A horizon stamp is an evidence claim, not a trading signal.

Evidence: Evening same-day STATE conditioning next-day flip odds (not a forecast). Distance-from-zero conditioning from the gex flip study 2026-06-11 Q3: |GEX|<0.5B -> 31% next-day sign flip vs 17% base; walk-forward flip model OOS AUC 0.858, 2.67x precision over base (SCREEN-grade evidence). FORWARD GRADE (standing nowcast-grading verdict, read 2026-07-17): the prereg kill bar FIRED — live AUC 0.6957 < the 0.70 bar on n=3,857 matured primed events; the state-2 precision multiple held at 6.1x (recorded honestly, not the bar). Kill executed per prereg research/planning/prereg_gex_flip_risk_2026-06-12.md: the gex_flip_primed standing tripwire was deactivated; the column keeps accruing and grading (capture-as-hedge) and re-earns a surface only via the standing verdict.

Source

Source module: computed
Data source: back_fillable

Fetched directly from app/sources/computed.py; see source code for the upstream API call and any provider-specific handling.

How it’s computed

Deterministic three-state machine over the same-day gex level and its change vs the prior DISTINCT vendor print (capture._gex_flip_state; compression cutoff capture._GEX_FLIP_COMPRESSION_B = 0.5 $B). State 2 (primed): |gex| < 0.5 AND the delta points across zero — gex > 0 with delta < 0, or gex < 0 with delta > 0. State 1 (compressed): |gex| < 0.5 but the delta is unavailable, zero, or pointing away; gex == 0.0 exactly lands here (crossing direction undefined). State 0 (none): |gex| >= 0.5. NULL: gex absent or uncastable (honest abstain, never a fake 0). delta = current gex minus the prior distinct print: walk daily rows strictly before trade_date newest-first, take the first non-NULL gex whose vendor obs date differs from the current row’s (the vendor reprints the same observation on weekends/holidays — a same-obs delta of 0.0 would mask a real pointing-across move); when either obs date is NULL, fall back to the first prior non-NULL gex on a different trade_date. Thresholds are hand-set constants from the gex flip study (2026-06-11, Q3): |GEX| < 0.5B conditions a 31% next-day sign-flip rate vs the 17% unconditional base (the net_gex < 0.5B variant reads 49.6% — research context only, net_gex is not captured live); the walk-forward flip model behind the study (OOS AUC 0.858, 2.67x precision over base) is the evidence for the threshold, never the deployable — this is a risk STATE with research-grounded base rates, not a prediction. Forward grading: research/planning/prereg_gex_flip_risk_2026-06-12.md.

Where it surfaces

Health-score / alignment role

Data carrier — no implication, no health-score contribution.

Persisted for downstream consumers (sparklines, base-rate matcher, calibration substrate) but does not classify into BULLISH / NEUTRAL / BEARISH and does not contribute to the 0-100 health score.

Release cadence

Derived at capture from the dark-pool analytics vendor’s daily gex print (shares the DIX/GEX EOD publish; in practice the print lands in the evening, ~21:45 ET observed). The state is meaningful for the NEXT session once the same-day gex row is in — before that it reflects the prior session.

See also