← Back to Berg Signals
/home/agent-jay/claudeCode/jarvis/plans/berg-atomic-indicators.md

Berg Atomic Indicators — Plan

title: Milton Berg atomic indicators + combo engine

status: PHASE 3 SHIPPED — 4 live combos (C2/C4/C6/C10) in production, signals table + daily cron + dashboard panel. Post-2010 calibration only (G2=50%).

source: reference_milton_berg_indicator_catalog.md (Shapiro/Berg interview 2026-05-18 + X/Twitter expansion 2026-05-25)

owner: Jarvis (finance lab)

created: 2026-05-25

updated: 2026-05-25

Phase 3 (2026-05-25) — production pipeline for 4 live combos

SHIPPED. Commit: see feat(finance) below.

What shipped

Caveat (prominent, repeated)

Post-2010 calibration only. Gate G2: 8/16 = 50% on full 1957-2026 precedent set. Atoms were tuned to 2010-2026 regime; pre-2003 dates fail due to market microstructure differences. Use live signals with care.

Deferred

Phase 0 (2026-05-26) — backfill + re-run G2

Backfill landed. Via python -m finance.scripts.fetch_ndx_universe --start 1957-01-01 --tickers '^SPX,^NDX,^IXIC,^RUT,^VIX,^SOX,IWM,SOXX,SMH,SPY,QQQ':

TickerHistoryNotes
SPX1957-01-02 → 2026-05-2217,465 rows; full Berg precedent range
NDX1985-10-01 → 2026-05-2210,240 rows
IXIC1971-02-05 → 2026-05-2213,941 rows; older NDX substitute
RUT1987-09-10 → 2026-05-229,749 rows
VIX1990-01-02 → 2026-05-259,166 rows; extended from 2010
SOX1994-05-04 → 2026-05-228,068 rows; ^SOX index (SOXX ETF only from 2001)
IWM/SOXX/SMH/SPY/QQQextendedbenchmarks/ETFs back to inception

G2 PROXY_MAP switched to indices (SPX/NDX/RUT/SOX/VIX) — atoms scale-invariant for rolling/relative measures.

Result: 8/16 = 50% (FAIL @ 80% threshold). Absolute matches went UP (7 → 8), but 9 newly-testable precedents yielded only 1 new MATCH (C6 1998-10-21). Diagnosis:

Per hard exit condition: G2 has failed 1 of 2 allowed calibration passes. Next: calibration pass OR re-plan with Rob.

Goal

Reproduce Milton Berg's histogram-tail extreme-detection framework in finance/indicators/. End state: a daily scanner that fires when 3-4 atomic extremes align (combo C1-C11 + X-search extensions), with backtested forward-return distributions and Berg-style stops, surfaced on the dashboard.

Non-goal: clone all 30,000 of Berg's indicators or all 2,000 combos. We build the ~24 atoms and 11+ combos Berg explicitly named, then evaluate ROI before extending.

CURRENT STATE (2026-05-25)

Phase 1 status: core atoms shipped to finance/indicators/berg_atoms.py (24 functions, full AtomSpec registry, evaluator). Tests at finance/tests/test_berg_atoms.py — 33/33 green.

Confidence-weighted atom table (shipped registry)

Ordered HIGH → LOW. Rationale: build highest-confidence atoms first; each entry maps to one AtomSpec in BERG_ATOM_REGISTRY.

IDConfLevelCategorySourceNotes
T10.99HIGHtrendShapiro videoclose < SMA(250) — pure regime filter
P80.98HIGHmomentumShapiro videoSingle-day gain ≥ +9%
S10.98HIGHstreakShapiro video≥10 consecutive up closes
S20.98HIGHstreakShapiro video≥10 consecutive down closes
V10.95HIGHvolumeShapiro video5d volume = 200d high
V50.95HIGHvolumeShapiro videoUp day + volume +20%
P10.95HIGHmomentumShapiro video3d ROC ≤ −7%
P20.95HIGHmomentumShapiro video1d gain = 250d max
P30.95HIGHmomentumShapiro video10d gain = 180d max
P50.95HIGHmomentumShapiro video10d ROC ≥ +20% (SOX)
S50.95HIGHstreakShapiro videoUp 8 of 9 (Russell)
S60.95HIGHstreakShapiro videoUp 16 of 19 (Nasdaq)
B60.95HIGHbreadthShapiro videoNew 60d low
B70.95HIGHbreadthShapiro videoNew 30d closing high
P_ROC740.90HIGHmomentumForward Guidance Feb 20245d ROC ≥ +7.4% — every bull since 1928
V20.90HIGHvolumeShapiro video5d-avg vol = 375d high within 20d
P60.90HIGHmomentumShapiro videoDrawdown from 252d high (param)
B_DA_SP600_50_10.85HIGHbreadthBerg X 2026D/A 50:1 — 18 cases since 1995, 100% up 120d
T_RIN_HI0.85HIGHbreadthBerg X / FGNYSE TRIN ≥12.5 panic / ≥15.5 rare
B_AD_5D_0870.75MEDIUMbreadthBerg X5d A/D ≤0.87 at ATH — top precondition
VX_VIX_5_300.70MEDIUMvolatilityBerg XVIX 5d/30d mean ratio — compression/expansion
VX10.70MEDIUMvolatilityShapiro videoVIX intraday range 45-60
P70.70MEDIUMmomentumShapiro videoRecovery thrust into new high
T20.70MEDIUMtrendShapiro videoDays holding swing low
P40.65MEDIUMmomentumShapiro video10d ROC at 180d max (P3 companion, alt instrument)
VX20.60MEDIUMvolatilityShapiro videoVIX 5d/20d return-stdev ratio <0.87

Bold rows are X-search additions (2026-05-25) — not in original Shapiro video transcript.

Atoms NOT YET implemented (deferred to Phase 1.5)

IDReason deferredRequired data
V3needs NASDAQ upvol/dnvol feedPolygon I:UPVOL.NQ / I:DNVOL.NQ
V4needs NYSE up-volume compositePolygon I:UPVOL.NY
V6duplicate of V5 with looser threshold—
P9needs NYSE up/down volume ratioPolygon
S3, S4duplicate of S1/S2 logic on S&Ptrivial, add post-G2
B1, B2need A/D ratio feedsPolygon NYSE/NASDAQ breadth
B3reverse breadth thrust — requires V4depends on V4
B45d new highs − new lows on NasdaqPolygon breadth
B5Berg's marquee innovationneeds daily shares-out for NYSE universe — Phase 0
VX3trivial (vix.close > 45)inline at combo level
SE1-SE6sentiment data sources missingmanual scrape, Phase 0

Success criteria (gates)

GateTestPass
G1 — atoms computeEach atom returns NaN-free series on SPY 2010→presentAll 15 atoms produce series
G2 — Berg precedent matchRe-fire combos C1-C11 on the historical dates Berg cited≥80% of Berg's dates flagged by our combo logic
G3 — forward-return profile1-year median forward return on historical signals≥+10%, drawdown ≤8% on aggregate
G4 — retail rule backtestLong-or-flat S&P, buy 1st signal, trail −8%CAGR ≥12% over 2010-2026 (vs Berg's claimed 18.5% over 1957-2026)

G2 is the critical gate. If we can't reproduce Berg's dates, our atoms are misdefined and the rest is theatre.

G2 result — 2026-05-25 (Phase 2 ship)

OVERALL: 7/7 = 100% on testable precedents — PASS (threshold 80%).

Live combos (4 of 11):

ComboMatchesNotes
C2 (NDX panic + VIX 45-60)2/22011-08-08 ✓, 2025-04-04 ✓
C4 (April 9 thrust day)1/12025-04-09 ✓
C6 (held-low-9 + 4-way thrust)2/22011-10-14 ✓, 2026-04-13 ✓
C10 (late entry, 16-of-19)2/22023-11-22 ✓, 2026-04-27 ✓

Blocked combos (7 of 11) — pending Phase 0 data backfill:

ComboBlocking atom(s)Source needed
C1V4 + B3NYSE up/dn vol + A/D thrust
C3SE4 + B4NDR osc + NASDAQ 5d new-lows
C5B1NDR Multi-Cap A/D
C7RUT historyyfinance ^RUT to 1987
C8V3 + B2Nasdaq up/dn vol + 10d A/D
C9SOX historyyfinance SOXX/SMH
C11SE1IPO calendar

Bug found and fixed during G2

P6_HIST atom added (p6_drawdown_reached_within). Original P6 required drawdown to STILL be ≥10% on the bar — but Berg's verbal usage ("the market DECLINED 10%, held its low for 19 days") is a historical regime condition, not same-bar. C10 was 0/2 with current-bar P6; switched to P6_HIST (DD reached -10% anytime in trailing 60d) → 2/2.

Catalog date correction

C6/C7/C10 cite 2026-04-13 and 2026-04-27 as recent firings (not 2025 — interview was 2026-05-18 discussing the just-prior month's signals). My initial spec wrote 2025 dates; fixed to 2026 per catalog verbatim. Both dates now fire correctly.

Sample-size caveat

Only 7 of 33 named precedents are testable: research.db starts 2010 (SPY/VIX) / 2003 (QQQ) but Berg's named dates go back to 1957. Phase 0 backfill is needed to test the 1957-2003 precedents (1957-style turn, 1979 thrust, 1987 SOX, 1998 panic, 2008 capitulation).

G2 verdict authorizes proceeding to Phase 3 (dashboard + cron) for the 4 live combos, with Phase 0 (data backfill) in parallel to unlock blocked combos.

Data inventory (current state of finance/research.db)

Have ✓

SymbolHistorySourceNotes
SPY2010-01 → todayyfinanceOHLCV daily
QQQ2003-01 → todayyfinanceOHLCV daily
VIX2010-01 → todayyfinanceOHLC daily — gives us VX1 (intraday H/L range)
NDX-100 components2003 → todayyfinancefor cap-weighted breadth
MAG7 names2003-2010 → todayyfinance
svix_compression.spx_close2025-05 → todaysvix pipelineSPX close, no OHLC

Recent additions (backfill needed)

SymbolCurrent rowsAction
SPX5Run python -m finance.scripts.fetch_ndx_universe full backfill — yfinance ^SPX from 1957 if available, 1990 minimum
IWM5Same — yfinance to inception (2000)
RUT5Same — yfinance to 1987
^NDX5Same — yfinance to 1985

Gaps (Phase 0 to fill)

NeedWhySource candidates
SOXX or SMHCombo C9 (SOX 20% / 10d)yfinance SOXX since 2001, SMH since 2000
NYSE composite up-volume / down-volumeV4, V3 (volume thrust, panic)Polygon I:UPVOL.NY / I:DNVOL.NY if subscribed; else Stooq ^NYUPV/^NYDNV
NYSE composite new highs / new lowsB4, B5Polygon or Stooq; manual scrape from WSJ
Daily shares outstanding for NYSE universeB5 (mkt-cap-weighted new lows)yfinance Ticker(x).info["sharesOutstanding"] weekly; SEC EDGAR quarterly
AAII / Investors Intelligence bull-bearSubstitute for Market Vane (SE2)Stooq, manual CSV weekly

Substitutes (Berg → ours)

BergOur substituteRationale
NDR Multi-Cap A/DRussell 3000 A/D from StooqSame broad-market spirit, free
NDR overbought/oversold oscillatorWilliams %R or Stochastic on SPXPublic formula
Market VaneAAII bullish %Both retail sentiment, free weekly
Berg's market-cap-weighted new lowsNYSE universe via prices_daily join with shares-outWe rebuild from scratch

Phase 0 — data prep (½ day)

  1. Run python -m finance.scripts.fetch_ndx_universe in full-backfill mode for SPX, IWM, RUT, NDX. Verify ≥15 years of history.
  2. Add SOXX, SMH to the ingest universe in fetch_ndx_universe.py. Re-run.
  3. New script finance/scripts/fetch_nyse_breadth.py — pull NYSE A/D, up-vol/down-vol, new highs/lows from Stooq. Store in new table nyse_breadth_daily(date, up_issues, down_issues, up_vol, down_vol, new_highs_52w, new_lows_52w).
  4. New script finance/scripts/fetch_shares_outstanding.py — weekly yfinance info.sharesOutstanding for all NYSE+Nasdaq members in prices_daily. Store in new table shares_outstanding(ticker, date, shares). Forward-fill within week.
  5. Verify gate G1 prerequisite: every atom has ≥10 years of input data available.

Exit criteria: all 15 atoms can compute on at least SPY 2010→present with no NaN gaps.

Phase 1 — atomic indicators (1 day)

New file: finance/indicators/berg_atoms.py

Pure pandas functions, one per atom, signature: def atom_v1(prices: pd.DataFrame) -> pd.Series: — input is OHLCV DataFrame indexed by date, output is boolean series (True = extreme condition met).

Volume atoms

def v1_ndx_5d_vol_200d_high(qqq: pd.DataFrame) -> pd.Series:
    """V1: 5-day rolling sum of volume is at its 200-day high."""
    v5 = qqq["volume"].rolling(5).sum()
    return v5 == v5.rolling(200).max()

def v2_nyse_5d_vol_375d_high_in_20d(spy: pd.DataFrame) -> pd.Series:
    """V2: SPY 5-day avg volume hit 375-day high anytime in past 20 days."""
    v5avg = spy["volume"].rolling(5).mean()
    is_high = v5avg == v5avg.rolling(375).max()
    return is_high.rolling(20).max().astype(bool)

def v3_ndx_10d_volume_thrust(qqq_up_vol, qqq_dn_vol) -> pd.Series:
    """V3: NASDAQ 10-day upside/downside volume ratio ≥ 1.89."""
    return (qqq_up_vol.rolling(10).sum() / qqq_dn_vol.rolling(10).sum()) >= 1.89

def v4_nyse_upvol_pct_below_45(up_vol, total_vol) -> pd.Series:
    """V4: NYSE 5-day upside vol as % of total 5-day vol < 45%."""
    return (up_vol.rolling(5).sum() / total_vol.rolling(5).sum()) < 0.45

def v5_signal_day_volume_up_20pct(prices: pd.DataFrame) -> pd.Series:
    """V5: Today's volume ≥ +20% vs yesterday AND close > open."""
    up_day = prices["close"] > prices["open"]
    vol_up = prices["volume"] >= 1.20 * prices["volume"].shift(1)
    return up_day & vol_up

Price-momentum atoms

def p1_nasdaq_3d_decline_7pct(prices) -> pd.Series:
    """P1: 3-day rate-of-change ≤ −7%."""
    return prices["close"].pct_change(3) <= -0.07

def p2_spx_1d_gain_250d_high(prices) -> pd.Series:
    """P2: today's 1-day pct change is the highest in 250 trading days."""
    chg = prices["close"].pct_change()
    return chg == chg.rolling(250).max()

def p3_nasdaq_10d_gain_180d_high(prices) -> pd.Series:
    """P3: 10-day pct change ranks #1 in trailing 180 days."""
    chg10 = prices["close"].pct_change(10)
    return chg10 == chg10.rolling(180).max()

# P4 identical to P3 but on S&P
# P5 = SOXX 10d return ≥ +20%
# P6 = drawdown_from_252d_high <= -0.05 (or -0.13)
# P7 = recovery thrust: (close - low_since_prior_-10pct) / low ≥ 6% AND close == max-since-low
# P8 = single-day gain ≥ +9% (S&P or Nasdaq)
# P9 = NYSE up-vol / total-vol on the day ≥ 100/101

Streak atoms

def s_run_length(series_bool: pd.Series) -> pd.Series:
    """Length of consecutive True-runs ending at each row."""
    grp = (series_bool != series_bool.shift()).cumsum()
    return series_bool.groupby(grp).cumsum()

def s1_vix_up10(vix) -> pd.Series:
    return s_run_length(vix["close"] > vix["close"].shift()) >= 10

def s5_russell_8of9(rut) -> pd.Series:
    up = (rut["close"] > rut["close"].shift()).rolling(9).sum()
    return up >= 8

def s6_nasdaq_16of19(qqq) -> pd.Series:
    up = (qqq["close"] > qqq["close"].shift()).rolling(19).sum()
    return up >= 16

Breadth atoms

def b1_ndr_multicap_ad_7d_ratio(ad_data) -> pd.Series:
    """B1: 7-day A/D ratio ≥ 2.10. Substitute = Russell 3000 advances/declines."""
    return (ad_data["adv"].rolling(7).sum() / ad_data["dec"].rolling(7).sum()) >= 2.10

def b4_nasdaq_5d_new_lows_minus_highs_minus_16(nh, nl) -> pd.Series:
    """B4: 5-day cumulative (new highs − new lows) ≤ −16."""
    return (nh.rolling(5).sum() - nl.rolling(5).sum()) <= -16

def b5_marketcap_weighted_new_lows(prices_panel, shares_panel) -> pd.Series:
    """B5: NYSE 52w-low mkt-cap as % of total NYSE mkt-cap.

    THE KEY BERG INNOVATION.

    For each date:
      1. Identify which NYSE-listed tickers are at 52-week low close.
      2. Sum their market caps (close * shares).
      3. Divide by sum of market caps of NYSE universe.
      4. Return % as scalar.

    Berg's table:
      1979 = 4.38%, 1980 = 4.15%, 1983 = 3.90% — none crashed.
      2026-05 = 3.12% — not extreme.

    Implementation note: NYSE universe must be filtered (not Nasdaq).
    Source: SEC EDGAR exchange membership.
    """
    # 252-day rolling min of close per ticker
    lo52w = prices_panel.groupby("ticker")["close"].transform(lambda x: x.rolling(252).min())
    is_low = prices_panel["close"] <= lo52w
    mcap = prices_panel["close"] * shares_panel  # joined
    daily = mcap.groupby("date").agg(
        total=("mcap", "sum"),
        at_low=("mcap", lambda s: s[is_low].sum()),
    )
    return (daily["at_low"] / daily["total"]) * 100  # percent

VIX atoms

def vx1_vix_range_45_60(vix) -> pd.Series:
    """VX1: VIX traded between 45 and 60 intraday (low ≥ 45 AND high ≤ 60 OR straddles)."""
    return (vix["high"] >= 45) & (vix["low"] <= 60) & (vix["high"] >= 45) & (vix["low"] < 60)

def vx2_vix_5d_20d_stdev_ratio_below_087(vix) -> pd.Series:
    """VX2: VIX 5-day stdev / 20-day stdev < 0.87 (compression)."""
    s5 = vix["close"].pct_change().rolling(5).std()
    s20 = vix["close"].pct_change().rolling(20).std()
    return (s5 / s20) < 0.87

Sentiment atoms

def se1_ipo_4y_high(ipo_dollars) -> pd.Series:
    return ipo_dollars == ipo_dollars.rolling(252 * 4).max()

def se4_obos_below_10(obos) -> pd.Series:
    """SE4: NDR-style overbought/oversold oscillator < 10. Substitute = Williams %R(14)."""
    return obos < 10

Trend filter

def t1_below_250dma(prices) -> pd.Series:
    return prices["close"] < prices["close"].rolling(250).mean()

def t2_days_holding_low(prices) -> pd.Series:
    """Return Series of int = days the swing low has held without violation."""
    # Identify swing low: most recent rolling-min close
    # Count consecutive days the low has not been broken
    rolling_min = prices["close"].rolling(60).min()
    is_low_day = prices["close"] == rolling_min
    days_since = (~is_low_day).groupby(is_low_day.cumsum()).cumsum()
    return days_since

Tests (in same commit)

finance/tests/test_berg_atoms.py:

Phase 2 — combo engine (½ day)

New file: finance/indicators/berg_combos.py

from . import berg_atoms as a

COMBOS = {
    "C1": [a.v2_nyse_5d_vol_375d_high_in_20d, a.b6_spx_60d_new_low, a.p6_drawdown_10pct,
           a.v4_nyse_upvol_pct_below_45, a.b3_reverse_breadth_thrust],
    "C2": [a.v1_ndx_5d_vol_200d_high, a.p1_nasdaq_3d_decline_7pct, a.vx1_vix_range_45_60],
    "C3": [a.v2_nyse_5d_vol_375d_high_in_20d, a.se4_obos_below_10, a.b4_nasdaq_5d_nh_nl_minus_16],
    "C4": [a.t1_below_250dma, a.p7_recovery_thrust_6pct, a.p2_spx_1d_gain_250d_high, a.v5_signal_day_volume_up_20pct],
    "C5": [a.t2_held_low_7d, a.p7_recovery_9pct, a.b1_ndr_multicap_ad_7d_ratio],
    "C6": [a.t2_held_low_9d, a.b7_30d_new_high, a.p3_nasdaq_10d_gain_180d_high, a.p4_spx_10d_roc_180d_high],
    "C7": [a.t2_held_low_9d, a.b7_30d_new_high, a.p3_nasdaq_10d_gain_180d_high, a.p4_spx_10d_roc_180d_high, a.s5_russell_8of9],
    "C8": [a.t2_held_low_9d, a.p6_drawdown_13pct, a.v3_ndx_10d_volume_thrust, a.b2_nasdaq_10d_ad_130],
    "C9": [a.p5_sox_10d_20pct, a.b7_spx_2y_closing_high],
    "C10": [a.p6_drawdown_10pct, a.t2_held_low_19d, a.s6_nasdaq_16of19],
    "C11": [a.se1_ipo_4y_high, a.vx2_vix_compression],
}

def evaluate_combos(date_range, prices_dict) -> pd.DataFrame:
    """For each date, evaluate all combos. Return long-form DataFrame:
       date, combo_id, fired, atom_results (dict-of-bool).
    """
    ...

Backtest each combo

For every historical date a combo fires:

Store in berg_signals(signal_date, combo_id, atoms_met_json, fwd_5d, fwd_10d, ..., fwd_252d, max_dd_60d, next_-10pct_days).

Validation against Berg's cited dates (gate G2)

Berg explicitly named these precedent dates per combo:

Gate G2: For each combo, check that our atoms fire on at least 80% of Berg's named dates ±2 trading days. Where they don't, debug atom thresholds — Berg's exact thresholds may differ from his rounded verbal descriptions.

Phase 3 — daily scanner + dashboard (½ day)

  1. Sentinel cron berg-signals-daily at 5:30 PM ET weekdays. Runs after fetch-prices-daily.
  2. Output: row in berg_signals table per combo that fired, plus Telegram alert if ≥3 combos fire same day.
  3. Dashboard panel /finance/berg-signals:

Phase 4 — retail rule backtest (½ day)

finance/scripts/backtest_berg_retail.py:

state = "FLAT"
peak = -inf
for each day d:
    if state == "FLAT":
        if any combo fired on d:
            buy SPY at close(d)
            state = "LONG"; entry = close(d); peak = close(d)
    else:  # LONG
        peak = max(peak, close(d))
        if close(d) <= peak * 0.92:
            sell SPY at close(d)
            state = "FLAT"

Compare:

Gate G4: if CAGR ≥ 12% (vs SPY's ~10% over the same window) and max DD < SPY's, ship to dashboard.

Premortem — what kills this project?

Imagine it's 2026-08. The Berg replicator is built. What went wrong?

  1. Atoms don't fire on Berg's dates (Gate G2 fails). Berg's verbal descriptions are rounded; his real thresholds are tighter. Result: combo signals look nothing like his. Mitigation: relax thresholds ±10% in a grid search, then re-tighten to the values that catch ≥80% of his dates. If still failing, contact Milton's retail product ($10/mo) and reverse-engineer from his published reports.
  1. B5 (cap-weighted new lows) data is broken. Daily shares-outstanding from yfinance is unreliable for delisted historicals. Mitigation: CRSP would be authoritative but expensive. Quarterly EDGAR + forward-fill is OK. Validate B5 against Berg's published 1979/1980/1983 figures.
  1. Combos fire too often. Histograms over our 2010-2026 sample have different tail thicknesses than Berg's 1957-2026 universe. Atoms calibrated to "rare" may be "common" in our window. Mitigation: require absolute thresholds, not percentile ranks. Validate on the 1990s if data permits.
  1. Combos fire too rarely. Same data window problem the other way. Most combos signal 0-2 times. Mitigation: OK for backtest if we have Berg's named precedents. Live monitoring still useful.
  1. Forward-return cherry-picking. Berg's claim of "120 days later, 100% up" across 5 historical instances is small-sample. With our combos we may see 5-10 instances → median forward-return is noise. Mitigation: report bootstrap CIs on forward returns. Don't claim signal until CI lower bound > 0.
  1. Retail rule underperforms SPY (Gate G4 fails). Berg's 18.5% CAGR may rely on combos we can't reproduce. Mitigation: still ship the scanner; drop the retail-rule claim. The scanner has standalone value.
  1. Data ingest fragility. NYSE breadth from Stooq is unofficial and may break. Mitigation: fall back to Polygon (paid) if Stooq breaks. Keep last 30 days cached in nyse_breadth_daily.
  1. Drift over time. Berg's combos worked because he selected them post-hoc on the full history. Adding combos as we find them is fine; forward-validating against ≥1 new signal occurrence before trusting it is essential.

Most likely failure: premortem #1 (threshold mismatch). Plan for it upfront: keep atoms parameterized, calibrate to Berg's named dates as ground truth.

Open questions for Rob

  1. Scope. Build all 11 combos, or start with the 5 highest-conviction (C1, C2, C4, C6, C10)? I'd vote start with 5 — faster G2 validation.
  2. B5 priority. Ship B5 in Phase 1 (slower, needs shares-out data) or defer to Phase 1.5? I'd vote ship in Phase 1 — it's the marquee insight.
  3. Data subscription. Pay for Polygon breadth feed (~$30/mo) for V3/V4/B4/B5, or commit to Stooq + scraping? I'd vote Polygon — we already use it.
  4. Retail-rule backtest as gate. If our retail rule clears Gate G4 (≥12% CAGR), is it eligible for live paper-trading in the strategy lab? Default yes unless you object.
  5. AAII vs Market Vane substitute. Acceptable? AAII is free, weekly, ~similar.
  6. Time budget. 2.5 days of focused work end-to-end (Phase 0–4). OK to commit, or want to scope down to Phase 0–1 first and reassess?

Estimated effort

PhaseWorkWall timeRisk
0Data backfill + new feeds½ daylow
1berg_atoms.py + tests1 daylow
2berg_combos.py + backtest½ dayhigh (G2 gate)
3Cron + dashboard panel½ daylow
4Retail-rule backtest½ daymedium
Total~3 days

Hard exit conditions

Stop and re-plan with Rob if: