snapback-scan

Buy when nobody dares, but with a date, a size, and an invalidation price. The violent post-capitulation snapback is unpredictable by day, yet identifiable by setup, and the setup only becomes tradable when a scheduled catalyst gives it a date.

Origin, 2026-07-30: the night before the semis epicenter ripped +15–23%, mean-reversion-scan's top-3 were SNDK / MU / AMD (scores 79/79/78, MU a 60-day-first listing, the Score ≥ 70 profile mean-reversion-scan now paper-tracks). The bell rang at the bottom; the missing piece was the join against the one thing that made the NEXT day the day: MSFT's capex verdict, on the earnings calendar for weeks. Powder keg × spark calendar = this skill.

Five components, with their receipts, including the ones that failed:

1. Powder keg: mean-reversion-scan's latest run (violent decline, structure intact; its 200DMA gates enforce this). A crash above a rising 200DMA is a coiled spring; below it, a falling knife.
2. Freshness: Score ≥ 40 AND listed ≤ 2 runs keeps the list on fresh panics; a name camped on the oversold list for a week is a downtrend. It was +1.83%/signal in mean-reversion-scan's 2026-05→07 sample, and that scan's 2021→2026 replay put it at SPY's return over the same days (−0.14%/signal against SPY), so it is a shortlist, not an edge. Kegs scoring ≥ 70 carry 📝, the one MR stratum that beat SPY over six years (+1.06%/signal) and lost 2.7%/signal in 2022. Deep RSI(2) earns no bonus in either test, and quiet-tape oversold stays a warning (the old pocket on quiet days ran −0.58% against SPY over six years).
3. Spark calendar: the scheduled events that can flip the narrative. The keg's own earnings (a coin flip, never a verdict; the earnings-gap rule), same-sector megacap verdict prints (MSFT's report owned the 07-30 rip, not any chip name's own), and FOMC/CPI-class macro.
4. Seller mechanism: down-gaps, streaks, elevated volume mark price-insensitive selling. Forced flows stop all at once; that vacuum (the forced seller finishing while buyers are still scared off) is the fuel.
5. Survival math: last cycle's base rate killed 5 of 6 bounce attempts, and over six years an oversold bounce earned the index's own return. The protocol exists to make being early survivable; any edge has to come from the spark, which this skill's own ledger hasn't tested yet. +16% days are the right tail. The goal is cheap, systematic exposure to that tail, not bottom-ticking.

<SKILL_DIR> is the directory containing this file. Dependencies auto-fetch via uv (yfinance>=1.3,<2, pandas>=2). Reuses sister caches (best run after the post-close daily-market-scan; the packet says if they're stale).

Run

uv run --with 'yfinance>=1.3,<2' --with 'pandas>=2' \
  python <SKILL_DIR>/scripts/build_snapback.py            # JSON (default)
... --format table        # human-scannable table
... --window-days 5       # widen the spark window (default 3 trading days)
... --min-score 30        # loosen the keg gate when nothing qualifies (default 40)
... --max-age 3           # widen the freshness gate (default 2 = "listed ≤ 2
                          #   runs"; the six-year replay found no return edge
                          #   in the cutoff either way)
... --top-n 30            # cap on kegs pulled from the MR list (default 20)
... --no-save             # don't write state/runs/<date>.json (exploration)

The spark window is the next N trading days PLUS tonight's AMC prints when the run happens before ~20:00 ET (they're the nearest spark of all); that's the (AMC only) entry in spark_window.

The packet saves to state/runs/<date>.json. Per-keg fields that matter:

• armed: has ≥ 1 spark scoped to it (own print / sector verdict / macro; a marketwide megacap print is context and cannot arm a keg)
• paper_track 📝: score ≥ 70, mean-reversion-scan's paper-tracked stratum
• sparks[]: dated, with reporter symbol and slot
• ignited: already moved ≥ +7% since signal (chase-guard)
• quiet_warning 😴: no panic in the 5d tape (knife risk)
• signal_day_low / mr_stop: invalidation candidates; mr_target
• next_own_earnings: for unarmed kegs, the next known date beyond the window
• down_streak, down_gaps_5d, vol_ratio_5d_20d: seller-mechanism reads
• tape_freshness (packet level): which session latest_close, the setup reads and signal_day_low come from. Since 2026-09-02 Yahoo has often not published the just-closed session's daily bar by the evening run; the tape (and the prior-run grading prices) rebuild it from 30-minute intraday bars, and asof older than expected means the rebuild failed and every invalidation sits under the previous session's low. The setup reads end on the MR run's data_asof when it recorded one.
• prior_run_review: full-sample grade of the previous flag list (n, win rate, avg/median, best AND worst tails, armed subset), sourced from this skill's own prior packet when one exists, else the prior MR run. A scanner nobody grades is a horoscope, and one that only shows its winners is worse

Synthesize: the if-then brief

Write for a reader with no finance background, in the conversation's language, translating each term in place: a "keg" is a stock that just got hammered but whose long-term structure is intact; a "spark" is a scheduled, dated event that can flip its story; "tranche 1: 1/4 size" is "first buy, a quarter of what you'd eventually hold"; the invalidation price is "the number that means you were wrong: exit, never average down past it".

Open with the packet strip, readable in ten seconds:

**Packet**: ⭐️<n> armed · 👀<n> watch · 🔥<n> ignited · 😴<n> quiet · regime <🟢/🟡/🔴> → <sizing consequence>
**Base rate**: most kegs don't explode upward on schedule; last cycle killed 5 of 6 bounces. The protocol makes the five cheap and the one caught.
**Prior run**: <one line from prior_run_review: did the bell ring true or false last time?>

Then one brief per armed keg; unarmed kegs get one line each in a "no spark in window" watch list. Phrase it that way verbatim, never "no catalyst": the window is 3 days, not the world, and next_own_earnings gives the next known date beyond it (a real print two weeks out is information, not absence). Never a bare "buy X". Each brief:

### <TICKER> · score <s> · day <age> on list  <flags: 📝 paper-track / 🔥 ignited / 😴 quiet / ⚠️ crowded>
KEG    why it crashed + the seller-mechanism read (streak/gaps/volume: forced or drift?)
SPARK  <date + event>: what verdict it delivers, branched BOTH ways
PLAN   tranche 1: 1/4 size at ~<latest_close>, invalidation <signal_day_low>
       (hard exit, no averaging down below it)
       add: only AFTER the spark confirms (<specific observable>, e.g. "the
       verdict name holds its gap into the close"); pay up for confirmation
       target 1: <mr_target> (the bounce-back level); beyond it, trail
       abort: spark fizzles or breaks against → flat, the thesis expires with its date

Rules that make this honest. Apply every one:

• The protocol trio is all-or-nothing. Every brief carries a concrete tranche fraction, a concrete invalidation PRICE (a number, not a rule like "the earnings-day low"), and a concrete observable add-trigger. A name you can't write all three numbers for goes to the watch list; a half-protocol is worse than none, because the missing half gets improvised at 10am. This binds the 5th name on the list as tightly as the 1st.
• The regime gate sizes everything. RISK-ON → protocol as written. CAUTION → halve tranche 1 and gate every entry on spark confirmation. RISK-OFF → watch-only briefs, no entries; say so outright. The strip states the consequence once; each brief still carries its own sized numbers, so no plan reads right without the regime baked in.
• 🔥 Ignited kegs get a retest plan, never an entry. Since-signal ≥ +7% means the snapback already fired; in the sample so far, the morning after ignition is the worst entry of the cycle (AMD 07-27, TSM 07-17). Write the retest level and what would invalidate the whole episode.
• Demote 😴 quiet kegs to the watch list, with the reason: both backtests lean the same way; quiet-day oversold is a knife catch.
• Say where the edge would come from. Mean-reversion-scan's six-year replay found oversold bounces earn what SPY does over the same days, so a keg without a spark is the market's drift with a stop attached. The spark is the bet, and it stays untested until the outcome ledger's banner clears. A 📝 keg gets its flag named in the brief, never a bigger size.
• Own-earnings sparks are coin flips: branch the brief BOTH ways (earnings-gap rule: zero directional trust in the print). A sector-verdict spark is the better structure: someone ELSE re-prices the narrative and you act on the confirmed side.
• Never cut the strip's base-rate line. It is the sentence that keeps the whole packet honest.
• Grade the prior run out loud: the strip's third line comes from prior_run_review. Calibration compounds.
• State what's missing. Stale MR cache (stale_days > 1), a stale tape (tape_freshness.asof < expected, or the table's ⚠️ stale tape line), empty spark calendar, errors: name them instead of papering over. No qualifying kegs is a valid, useful output: say so and stop; never pad.

Interaction with the sisters

• mean-reversion-scan supplies the kegs; never rescan here. If its cache is stale, say so and suggest running it.
• regime-scan supplies the sizing gate.
• premarket-brief is the morning-of companion: on a spark day its packet grades whether the verdict confirmed (gap-keep, hour-one survival); the "add" trigger in these briefs tends to be observable there.

Outcome ledger

prior_run_review grades one packet against the next; it cannot see whether the rules work. The packets are already the history, so scripts/backtest_outcomes.py replays state/runs/*.json into state/outcomes.csv. The trade convention is this skill's protocol, not the MR scan's: entry at latest_close, hard exit at signal_day_low, target at mr_target. The entry bar is tape_freshness.asof when the packet has it; packets from 2026-09-02 to 10-02 predate it and mostly read the prior session's tape, so the replay picks between the packet-date bar and the one before it by which reproduces the keg's ret_5d_pct.

uv run --with 'yfinance>=1.3,<2' --with 'pandas>=2' \
  python <SKILL_DIR>/scripts/backtest_outcomes.py
... --window 5          # trading days to resolve (3/5/10 sensitivity always printed)
... --entry next-open   # honest execution: fill at the first open after the packet was built, skip what gapped away
... --refresh-prices    # bypass the price cache

It exists to settle the three claims stated above and never tested: does a dated spark pay (armed vs unarmed), is a sector verdict really the better structure than the keg's own coin-flip print, and what does the 🔥 chase actually cost. Two reading rules:

• T% / S% / R:R before any Win%. An ignited keg is bought after it ran, so mr_target sits ~1.5% overhead while the invalidation sits ~14% below. That geometry manufactures a high win rate out of a bad payoff — an ignited row beating the coiled row is not evidence the chase-guard is wrong.
• Exp% is a floor, not the return. The replay closes at target 1; the protocol trails past it. NoExit% is the untruncated read and only fills in once a signal's full window has elapsed.

Until the header banner clears (50+ resolved signals) the tables are a plumbing check, not evidence — the script prints how many more run days it needs. Never quote a stratum row into this file before then; rows with n < 10 are marked ⚠︎ for the same reason.

Cadence follows that banner: every couple of weeks while it's still up (the packets accrue daily, so the gap closes fast and the ledger is cheap to re-run), then quarterly alongside the sisters' backtest_outcomes.py once the sample is readable.

Tests

cd <SKILL_DIR>/scripts && uv run --with 'yfinance>=1.3,<2' --with 'pandas>=2' \
  --with pytest pytest -q

Pure-logic tests (no network; every network call site is stubbed) cover the keg age/score filters, arming rules (marketwide prints never arm; own earnings and macro do), signal-day slicing + the ignited chase-guard, the prior-run review's full-sample stats, and trading-day arithmetic across NYSE-observed holidays. The ledger's own suite covers the protocol's win/loss/expiry resolution with gap-aware fills, dividend re-anchoring (and the refusal to anchor before the signal day, which would grade a setup on its own decline), and the exclusion of missing attributes from every stratum.