# METHOD · Umar Punjabi LIVE record audit + rules backtest (DRAFT, 30 Oct 2026) Two separate questions, never mixed: **(A) does his published list match his own live streams?** and **(B) do the rules he states on stream make money on history?** Neither tests the trades he takes with his own "key level" indicator, which is not public. ## Sources (his first, ours second) 1. **His live streams** ("Umar Punjabi Live", 140 numbered public streams, #1 14 Jan → #140 29 Oct 2026, plus 4 renamed videos). Audio supplied by our download agent; every quote carries stream number, date and time. Readings: YouTube Hindi captions (one reading; the English caption track is a machine translation of it and is never used as his words) and our Whisper large-v3-turbo transcripts in Hindi and English (the second reading). A quote reaches the screen only with two agreeing readings or an owner listen. 2. **His pages**, read 29 Oct 2026: umarslivetradingrecord.manus.space (61 rows) and umarsviptrackrecord.manus.space (124 rows; two Aug 25 BE rows carry realizedR 0.5); the rows are hard-coded in the pages' JavaScript, exported here verbatim (`data/his_record/`). Wayback copies of 14 Oct 2026 exist; the 14 Oct list (56 rows) is a strict subset of today's. 116 live descriptions (24 Mar → 23 Oct 2026) and his VIP-access video (`_QOO33JjC7A`, 29 Oct 2026) for the money trail. 3. **Price data:** Dukascopy 1-minute BID + ASK. Gold (XAUUSD) 1 Jan 2015 23:00 → 28 Oct 2026 23:59 UTC (4,189,024 minutes); EURUSD 1 Jan → 28 Oct 2026. HistData (BID only) as a cross-check on the record replays (agrees on the first event in 39 of 40). ## (A) The list vs his streams - **Stream layer:** every public stream read for trades; one row per trade (`trades_called.csv`): side, kind (LIVE, PRE_STREAM, VIP_REVIEW, REVIEW, OPTIONAL, NOT_TRIGGERED, HYPOTHETICAL), entry / stop / target as spoken, partials, break-even moves, his stated outcome, his words. - **Time:** stream time → UTC from the stream's start timestamp; checked against three spoken order prices (within about a minute). One gold pip = $0.10 as he uses it. - **Price layer** (`trades_checked.csv`): a buy fills at the ASK and exits on the BID, a sell the reverse; pending orders fill when price crosses the spoken level; time entries fill at the next minute's open and are re-run at −2…+2 minutes ("robust" = same outcome in all five). Stop counted first when stop and target share a minute. - **Label standard (one rule for every row, computed, not typed):** `reconcile.py` derives each label from the per-row evidence table `data/record_audit/record_v2_evidence.csv` (what he said, when, which reading). His words on THAT day's public stream decide (or his first review next stream if the trade was open at stream end); price never decides a label for his own trade. **MATCH** = he took that side and his final stated outcome equals the list (RR within his own partial convention) · **LISTED DIFFERENTLY** = on stream, listed differently (bigger result / other side / other outcome / against himself) · **NOT ON THAT DAY'S PUBLIC STREAM** (no stream that day / stream not public: #64, #124, the 17 Jun FOMC live / setup left for viewers, not taken / "unofficial", "won't add" / he says no trade) · **PENDING FRAME** (needs his chart, or he never states the RR / exit). Never "false" or "invented". Direction = "in the page's own data" (the public page shows no Buy/Sell column). Readings: W+C (Whisper + captions agree) or C (captions only, Whisper pending); rules in RECORD_AUDIT.md §1. - **Fair side, always next to the count:** wins he took live that are not on the list either, rows listed against himself, levels he left for viewers that were reached later. - **Sources for the money trail and the page** (live descriptions, VIP rules, "Verified" wording, FortressFX's published specs, the stream index): RECORD_AUDIT.md §7, so every number in this pack traces to a file inside it. - **His list's arithmetic** as published: SL = −1R, BE = 0, TP = the listed RR. 133.5R over about 9.23 months = 14.46R a month; × 2% = 28.9% a month, simple, before costs. No single "real R per month" is published: it depends on the standard (signal vs his own exit), 5.7–12.7 R/month. ## (B) His stated rules, coded - **Pre-registered** before the first run: `data/backtest/prereg_2026-10-29_S1_S2.md`, git commit `4d4b85c` (29 Oct 2026); code frozen in `a8f8082` before the first full-history run. No parameter changed after results were seen. - **S1 Asia sweep** (his #136 strategy) and **S2 London sweep → BOS → retest, New York** ("his checklist without his key level"): his words next to every rule in `backtest/strategies/sweep.py`; ambiguities are named variants, all published. - **His session and time zone** (`backtest/structure.py`, zoneinfo per day): Asia box 20:00–02:00 New York, London box 08:00 London → 08:00 New York, NY open 08:00 New York, flat 17:00 New York; UK / NYSE holidays; FOMC days and 30 minutes before NFP / CPI skipped (S2). The calendars are in `data/backtest/calendar/` (federalreserve.gov, bls.gov via the Internet Archive, gov.uk). - **Costs:** the real Dukascopy spread at each fill + $0.05 slippage on market/stop fills + commission $0.07/oz round trip ($7/lot); EURUSD 0.1-pip slippage, 0.7-pip commission. Inside FortressFX's own published account specifications (read 30 Oct 2026; no gold-specific spread stated), so worded "costs like a raw-spread account", never "his broker's exact costs". Zero-cost and optimistic-fill lines are published too. - **Money:** fixed $10,000, 2% = $200 a trade, no compounding (the compounded 2%-of-balance path is shown separately). - **Proof:** 38 hand-built tests (one rule each) + a second, separately written implementation (`backtest/strategies/sweep_simple.py`, imports nothing from the engine) that gives identical trades on every run (both presets × 3 cost cases, all variants; EURUSD 25 runs). Same author, so this shows the code = the written rules, "two separately written programs agree on every trade", not independent verification. - **What could not be tested:** his key level (his own indicator, not public), his discretionary bias overrides, the London continuation sold in his course, news trades. Said on screen. ## Re-run (from the realshyt reel-engine repo) ``` .venv/bin/python tools/umar_live_calendar.py # calendars .venv/bin/python tools/umar_live_backtest.py # gold: results JSON, trades/, tables.md (~10 s) .venv/bin/python tools/umar_live_eurusd.py # EURUSD S1 2026 .venv/bin/python tools/proofrun/umar_live_run.py # the on-screen run; exit 1 if any number differs from the JSON .venv/bin/python -m pytest tests/test_structure.py tests/test_sweep.py .venv/bin/python projects/.research/umar-live/tools/reconcile.py # the record audit tables (research tree) ``` Our test, our opinion. Historical tests depend on rules, data and costs and do not guarantee future results. Not financial advice.