Files
momentum-system/docs/IGNITION_SETUP_SCOPE.md
T
momentum devandClaude Opus 4.8 1c8b4f4309 Ignition P0: pure detector + tests, and a decisive real-data finding
detect_momentum_ignition (strategy/evaluation/ignition.py): standalone, pure,
binary-rule detector for a vertical squeeze (blue-sky fresh-HOD + price velocity
+ ABSOLUTE volume burst, NOT rvol/float). Stop = last established higher low (per
operator). NOT wired into classify_setup (deliberately). 8 unit tests on
synthetic verticals: fires with the right trigger/stop, and each gate blocks
(not-blue-sky / not-new-HOD / velocity / volume / insufficient-bars / no-ATH).

REALITY CHECK (verify, don't assume): ran it over the REAL PLSM 2026-06-24 bars ->
0 fires, CORRECTLY. PLSM's high was at the 9:35 ET open (it ran +400% PRE-MARKET,
which we don't ingest), and in RTH it was HALTED 228 of ~390 minutes — a
halt-resumption GAP chop (+3.6/-12.5/-17.3/-14.4% jumps), not a smooth vertical.
A consecutive-bar detector structurally cannot see that. So the ignition
opportunity is a PRE-MARKET + halt-resumption phenomenon -> capturing it needs
pre-market bar ingestion + halt-aware gap detection + pre-market/halt trading
(into the FCUV halt-down risk), a far bigger lift than an RTH detector. Documented
front-and-center in docs/IGNITION_SETUP_SCOPE.md as a pre-P1 operator decision.
The P0 detector + tests are a correct foundation regardless.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-25 06:15:12 -04:00

101 lines
8.6 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Momentum-ignition setup — shadow scope
_Scope produced 2026-06-24 by a read-only 4-agent mapping workflow (wf_55582b6d). No code changed.
A SECOND, shadow-only lane to catch PLSM-class vertical squeezes the pullback-ORB scorer is blind to._
## Why
The live setup is `impulse → pullback → breakout` (`strategy/evaluation/structure.py` `classify_setup`).
A true vertical squeeze gives **no pullback and no formed opening range**, so `classify_setup` returns
NONE and the scorer blocks it — verified: PLSM was evaluated **1,895× on 2026-06-24 and blocked ~30%
every time**. This lane is the missing discipline: ride the vertical, not the base.
## What it detects
An intraday vertical ignition with no consolidation: a fresh **high-of-day above the prior all-time
high (blue sky)**, **accelerating price velocity** (consecutive green bars + an X% move over K minutes),
and an **absolute volume burst**. It deliberately catches the pattern that makes `classify_setup`
return NONE.
## Detection rules — all computable TODAY (binary, no fitting at this n)
| signal | rule | source |
|---|---|---|
| **Blue-sky fresh-HOD** | `current.high > prior_ATH` (max `daily_bars.high` before today) AND this bar is the running session-high HOD. *Don't* reuse `detect_hod_break` (needs consolidation a vertical never gives). | `daily_bars.high` + `minute_bars.high` |
| **Price velocity** | `pct_move(close[t]/close[t-N]-1) ≥ 0.08` over N=5–10 RTH min, AND ≥4 of last 5 bars green, AND higher-highs. (PLSM 12:02–12:16: 10/14 green, +13.9% in 14 min.) | `minute_bars` OHLC, pure arithmetic |
| **Absolute volume burst** | `last_bar_vol / mean(prior K=5 bars) ≥ 3.0` AND cumulative session volume ≥ 100k shares (or a $-volume floor). **Absolute floors, NOT RVOL** — `scan_gappers` RVOL zeros out on thin IPOs (adv=0), and `rolling_avg_volume_20d` is NULL. | `minute_bars.volume` |
| catalyst *(tag only)* | attach max-conviction bullish row (`sentiment>0`, `is_dilutive=False`); never a gate (Ollama gated off / can be down). PLSM "Soaring" → 0.9 conviction. | `news_catalyst_cache` |
| VWAP *(tag only)* | log `above_vwap` + `dist_from_vwap` as grading context (the one validated ~1.5× signal). | `compute_key_levels.vwap` |
| float *(placeholder)* | **DO NOT GATE** — `symbols` table is empty, `float_shares` NULL. Log `None` so a later float backfill can re-grade historical shadows. | `symbols.float_shares` (empty) |
## Hook point
A new zero-arg closure **`step_shadow_ignition()`** in `run_live_paper.py` beside `step_watch` (~L528),
registered via `scheduler.add('shadow_ignition', …, enabled=_flag('SHADOW_IGNITION_ENABLED','0'))` —
**ships dark**. Pure-read against `research_con` (single psycopg2 conn — not thread-safe), emit-only,
never references `rt['execution']`. Reuses the watcher provider's `get_candidates`/`get_bars` (same
symbols/bars) but with its OWN wider caps (`SHADOW_PRICE_MAX≈100`, no gap ceiling) so it validates the
very names live config excludes. Detection is a new pure `detect_momentum_ignition(bars)` in
`strategy/evaluation/ignition.py` — **NOT** wired into `classify_setup`/`evaluate_setup` (the pullback
criteria would reject it and `STRATEGY_SETUPS=opening_range_break` would filter it).
## Shadow mechanism — structurally cannot trade
New `EventType.SHADOW_SIGNAL` + `ShadowSignalEvent` (symbol, trigger, would-be stop, signal_values
dict, session_date). It is **invisible to execution**: `query_ready_signals_snapshot` filters strictly
on `event_type='signal_ready'`, the only thing `TradingExecutionService.tick` reads — a `shadow_signal`
is never seen, and the lane never calls `submit_breakout_now`/`request_approvals`. (Precedent: the VWAP
gate already shadow-logs via `RiskRuleTriggeredEvent action_taken=shadow_logged`.) Safety: wrap the
step in try/except (a shadow failure can't kill the loop), once-per-session per-symbol debounce (so one
squeeze logs once, not 1,895×), never emit `signal_ready`, never set a symbol "ready".
## Measurement — reuse existing forward-outcome machinery (zero new metric code)
1. **Labeler**: add `build_ignition` to `research/labeler.py` (`setup_name=momentum_ignition`,
`setup_version=ign_v1`) reusing the identical label block (`reached_1r/2r_before_minus_1r`,
`max_upside_5/15/60m`, `time_to_max`, `failed_breakout_flag`=the fizzle trap, `held_vwap`). `report`
/`lift` filtered by version give runner-vs-trap separation per signal.
2. **eod_replay**: add `score_shadow_signals` — after the close, read the day's actual `shadow_signal`
events, walk `minute_bars` strictly after each, classify **runner (+1R) / fizzle / halt-down-trap**
(the FCUV mode: a forward bar gapping far below the prior low, or R ≤ −1 within minutes). Run nightly.
## Promotion bar — PRE-REGISTER before shadowing (so the bar can't move to fit noise)
- `reached_1r` rate **materially above** the `failed_breakout` trap rate, AND
- halt-down-trap rate **below a ceiling**, AND
- **distinct runner instances** above the thin-n floor (≥30, or rate×count ≥8) — counted in distinct
runners, NOT calendar days (the ~17–19 days are autocorrelated, one regime).
- Decision hangs on the **eod_replay as-fired** numbers (what a live lever would actually capture).
## Build phases
| phase | blast | what |
|---|---|---|
| **P0** | none | `detect_momentum_ignition(bars)` pure + unit-tested (synthetic bars + the PLSM 12:02–12:16 window). No wiring. |
| **P1** | low | `SHADOW_SIGNAL`/`ShadowSignalEvent` + `step_shadow_ignition` (default OFF, own caps, debounce, try/except). Test asserts it NEVER emits `signal_ready` / never calls execution. |
| **P2** | low | `build_ignition` (labeler) + `score_shadow_signals` (eod_replay), nightly. DB-unreachable tolerant; tests use `:memory:`/`MOMENTUM_PG_SCHEMA`. |
| **P3** | none | Pre-register the promotion thresholds; shadow N weeks; accumulate. Zero capital, zero trading-path edits. |
| **P4** | high | *Only after P3 separation is proven* — a tiny live lever behind `require_ignition_confirmed` (default OFF, mirrors `require_above_vwap`), minimally sized, bounded by ALL shipped catastrophe controls (enforced stops, anti-chase, BE@+5%, halt guard). The ORB book + anti-chase guards untouched — strictly additive. |
## Open questions for you
- **Stop definition** for the R denominator: last-higher-low vs bar-low vs −8%? Changes every reached_1r/trap rate — pick ONE and pre-register. (Maps lean last-higher-low.)
- **Shadow gap ceiling**: drop entirely or cap ~80%? Dropping surfaces the most extreme/trap-prone names (fine for shadow).
- **Catalyst tag** needs Ollama running that day — force-enable it for the shadow grading pass so news-backed vs not can be measured?
- **Distinct-runner-n ≥30** may be unrealistic given runner sparsity — a Bayesian interval at low n instead of a point rate?
## ⚠️ P0 reality-check (2026-06-25) — READ BEFORE P1
The P0 detector is built + unit-tested (fires correctly on a clean synthetic vertical). But running
it over the **real** PLSM 2026-06-24 bars fired **0 times — correctly** — and the *why* reshapes this
whole effort:
- PLSM's session HIGH (12.85) was made at the **9:35 ET open** (it gapped from ~$4 **pre-market**); RTH
was then a day-long **fade**. The "+400% to $20" the trader cites happened **pre-market**, which the
bot does not ingest (RTH-only `minute_bars`) or trade.
- PLSM was **halted for 228 of ~390 RTH minutes** — its RTH move is **halt-resumption GAPS**
(+3.6%, −12.5%, −17.3%, −14.4% …), NOT a smooth consecutive-bar vertical. A detector keyed on
consecutive green bars + smooth velocity **structurally cannot** see a halt-gapped move.
**Implication:** the momentum-ignition opportunity is largely a **pre-market + halt-resumption**
phenomenon, not an RTH smooth-vertical one. Capturing it would need (a) **pre-market/extended-hours bar
ingestion**, (b) **halt-aware, gap-up detection** (resumption gaps, not consecutive bars), and (c) the
bot to **trade pre-market / halt resumptions** — which runs straight into the **halt-down blowup risk**
(FCUV) the anti-chase guards exist to prevent. That is a far bigger lift than an RTH detector, and an
operator decision before P1: invest in pre-market+halt data/logic, or shelve the ignition lane and bank
the ORB/float work instead. The P0 detector + tests remain a correct foundation either way.
## Hard constraints (non-negotiable)
Shadow-only until forward data proves it · never touches the ORB book or anti-chase guards · ships dark
(`SHADOW_IGNITION_ENABLED=0`) · binary rules, no fitting · do NOT gate on RVOL or float (both broken for
exactly these names) · tests use `:memory:`/`MOMENTUM_PG_SCHEMA` (never the live DB).