PO10 Extreme — full receipts

The NQ 10-minute Phase Oscillator strategy from the August reveal thread, in the form it trades today: enter when the 10m PO crosses beyond ±100 in the first 30 minutes, hold to the close, and step aside only when a 1-minute PO extreme prints against the position. Every trade since 2008, the costs, the years that paid and the years that did not, the parameter neighbourhood, the resample envelope, and the paper and live record so far.
On 1 NQ contract, 1.0 point round-trip cost, 2008-03 → 2026-10: 511 trades, 53.4% win rate, 2.55 profit factor, +13,044 pts ≈ $260,872. 27.5 trades a year. The 2019→ half carries most of it: PF 2.96, ≈ $29,204 a year per contract. 2026 year-to-date: 23 trades, PF 11.1, +3,504 pts.

📦 Full reproduction package

01 backtest engine (same oscillator code the live trader runs), 02 forward/live collector, 03 page builder, the raw trade lists for both signal windows, the 5×5 parameter grid, every statistic on this page as JSON, the paper-trade log, and a README with the rule and every formula.

⬇ download full package.zip
Against chance
Recent window (what matters most)
1 · The rule, in words

Where it comes from. The indicator is Saty Mahajan's Phase Oscillator (@SatyMahajan). The entry came out of our July 2026 sweep of PO trend rules on ES and NQ: crossing the ±23.6 line had no timing edge, crossing ±61.8 a little, and crossing ±100 the most. That ±100 cross on the NQ 10-minute chart, signals until 11:00 ET, was the "mystery strategy" posted on X on 2026-08-25 and revealed on 08-26. Two upgrades followed. @etbent1 rebuilt the strategy from the Pine source, replicated it on USTEC, and showed that stopping new signals at 10:00 ET roughly halves the drawdown; we confirmed it on NQ futures and switched the live sleeves on 2026-08-29. The 1-minute adverse-extreme exit is ours, from the 2026-07-22 sweep that produced the frozen spec. The DefinedEdge TradingView script "Phase Oscillator Momentum Strategy" uses the same entry without the 1-minute exit.

Indicator. Computed on the full 24-hour tape ("ETH calc"), bars aligned to midnight ET. Pivot = EMA(21) of close. ATR(14) = Wilder-style exponential true range (alpha 1/14). Raw = (close − pivot) / (3 × ATR14) × 100. PO = EMA(3) of raw. The same function runs in the backtest and in the live trader; the package ships it.

Signal. A completed 10-minute bar, stamped by its close time, whose PO closes beyond ±100 while the prior bar's PO was inside. Signal close from 09:30 to 10:00 ET inclusive. Four signal slots a day: 09:30, 09:40, 09:50, 10:00. One position at a time.

Entry. Market at the next 1-minute open after the signal close, in the cross direction. Above +100 → long. Below −100 → short.

Exit, whichever comes first. (a) A completed 1-minute bar, opened after the fill minute, whose PO reads 95 or more against the position (long: 1m PO ≤ −95; short: 1m PO ≥ +95) → out at the next 1-minute open. (b) The opposite 10-minute ±100 cross → out at that bar's close. (c) Flat at 15:59 ET. In practice (a) always arrives before (b): across all 511 trades the opposite 10m cross never fired first, so the strategy has two exits, the 1-minute extreme and the close.

What the 1-minute exit does. It is a trend-day-failure detector, not a price stop. It fired on 298 of 511 trades (58%) at an average of -9.3 pts. The 213 trades that never saw an adverse 1-minute extreme ran to the close at an average of +74.2 pts. Taking profit at a favourable 1-minute extreme destroys the edge (PF 1.89 → 1.40 in the July sweep); the edge is in holding through the day.

Cost. 1.0 NQ point per round trip ($20 at $20/pt), applied to every trade on this page unless a panel sweeps the cost on purpose. No price stop and no target in the backtest. The live sleeves carry a 200-point catastrophic server-side stop that the backtest does not model; section 9 sizes what it would have done.

2 · Trade list511 rows
Every trade, 2008-03-24 → 2026-10-16, timestamps ET to the minute. Entry and exit prices are the raw fill levels on the ratio-adjusted continuous series (point differences are exact; absolute levels of old bars are shifted by the roll adjustment). Net = gross − 1.0 pt. MAE/MFE are the worst and best excursion from the fill, measured on 1-minute highs and lows.
⬇ Download trade CSV (10:00 window)⬇ CSV, original 11:00 window (1011 rows)
Cumulative net points, 1 contract (top) and drawdown from peak (bottom)
cumulative net ptsdrawdown ptsDeepest: 971 pts (−$19,428), 2019-01-15 → 2021-02-23. Longest underwater: 2009-07 → 2013-04.
Filter
DateSideSignalPO10EntryExitExit pxWhyHoldNet ptsNet $MAEMFECum pts
3 · Distribution, sides, exits
Net points per trade, 511 trades
loserswinnersSkewed right on purpose: the 1-minute exit caps most losers inside −50 pts and the right tail runs to +794.
Both sides carry. Longs PF 2.61 on 255 trades, shorts PF 2.51 on 256. Shorts make more per trade (+29.0 vs +22.0 pts) because the big down days run further. The two exit types split the book cleanly: the 1-minute-extreme exits are the loss sink (PF 0.60, 298 trades, -2,773 pts in total) and the hold-to-close trades are the engine (PF 12.2, 213 trades, +15,816 pts).
4 · Costs and dollars
The whole trade population re-costed at 0 to 5 points per round trip. Break-even cost is the average gross, 26.5 pts per trade, about five times the base assumption. The dollar columns translate the base case to the vehicles we actually run it on.
Cost / RTPFAvg ptsTotal pts
Vehicle$ / ptTotalMax DD2019→ / yr
Darwinex Zero row is the live probe size (0.05 lots NDX CFD). The CFD spread is wider than NQ's 1.0-pt assumption; section 10 compares paper fills on NDX quotes to the NQ backtest trade by trade.
5 · Year by year
positive yearnegative yearNet pts per year, 1 contract. Five negative years out of nineteen: 2009, 2011, 2015, 2019, 2020.
YearnWin%PFAvg ptsNet ptsNet $Max DD pts
2008–2018 is a different regime: PF 1.64 on 299 trades, +1,670 pts in eleven years. From 2019 the same rule makes +11,374 pts on 212 trades (PF 2.96). 2020 is the worst year (−603 pts): the March crash printed 1-minute extremes against short entries that were right on direction. The rule did not change; the market's intraday range did.
6 · Signal window and exit variants
The one change made after the prereg froze (2026-07-22) is the signal cutoff, 11:00 → 10:00 ET on 2026-08-29. Here is the full gradient, plus the base rule without the 1-minute exit, so the contribution of each piece is visible.
Signal cutoff, everything else fixed (±100 entry, 1m ≥95 exit, flat 15:59)
Last signaln/ yrWin%PFAvg ptsTotal ptsMax DDt
Exit and threshold variants
VariantnPFWin%AvgTotalMax DDtPF 2019→PF 2025→2026 pts
Two honest reads. The 10:00 cutoff trades total points for efficiency: the 11:00 window makes more in total (+17,018 vs +13,044 pts) on twice the trades, at PF 1.95 instead of 2.55. The 1-minute exit raises PF but not total: without it, the same 511 entries make +13,022 pts at PF 2.25 with a smaller max drawdown (691 vs 971 pts). The exit converts a few large held losses into many small ones; it wins on profit factor and on win-rate stability, not on drawdown. We keep it because it is the piece that reads trend failure early and because every year since 2021 it has been neutral to positive. That is a judgment, stated as one.
7 · Parameter neighbourhood
Entry threshold × 1-minute exit threshold at −20 / −10 / 0 / +10 / +20 %. Each cell is a fresh simulation from raw bars. Colour = profit factor. Gold outline = the deployed cell.
All 25 cells are profitable, PF 1.49 to 2.55. The deployed cell's four direct neighbours sit at PF 2.40 / 2.23 / 2.12 / 2.35: the largest drop is 17%, well inside the 30% island rule. The entry threshold is the load-bearing parameter: ±80 doubles the trade count and gives back most of the edge per trade; ±120 keeps the PF but halves the frequency.
8 · Robustness diagnostics
The checks we run on every candidate before Pedro decides. None of them kills a strategy on its own; they are reported so the decision is informed.
Calendar completeness
Trade-sequence resample (10,000 shuffles)
MAE / MFE
Gate ablation and random baseline
Random-direction baseline. Same 511 signal times, coin-flip side, same exits, 100 seeds: median PF 1.40, 95th percentile 1.71, median total +4,092 pts. Zero of 100 random books reached the strategy's +13,044. The baseline itself is positive because the exit asymmetry (cut fast, hold winners to the close) has value on its own in a trending index. The cross direction adds about +17.5 pts per trade on top.
9 · Out-of-sample blocks and recent windows
The rule has three fitted constants (±100, 95, 10:00) chosen on history. These blocks apply the fixed rule forward through time without refitting; the recent windows weight what matters most, whether it is working now.
BlocknWin%PFNet ptsMax DD
WindownWin%PFAvgNet ptst
The edge gets stronger in the newer data, not weaker: PF 1.51 → 1.70 → 2.24 → 3.70 across the four blocks. 2025→: 53 trades, PF 4.43, +4,794 pts. 2026 YTD: 23 trades, PF 11.1, +3,504 pts, 70% winners. The catastrophic 200-pt stop the live sleeves carry would have bound on 12 of 511 trades; stopping those at −200 changes the total by -1,072 pts.
10 · Paper and live recordsince 2026-07-24
The frozen prereg (2026-07-22) runs as a friction-free paper sleeve on Darwinex NDX CFD quotes, and three live mirrors trade the same signal at probe size. Everything below is the record as logged by the traders, not a replay.
Paper sleeve, NDX quotes, 1.0 pt frictionrunning
DateSideSignalPO10FillExitWhyNet ptsCumWindow
Prereg gates: kill at 13 closed trades if PF < 0.7; evaluate at 26 (GO = PF ≥ 1.3 and net > 0). Paper is at 9 of 26. The 08-21 trade came from the original 11:00 window, before the 08-29 change.
Same dates, paper NDX vs NQ backtest
DateSideSignalPaper exitPaper netBacktest exitBacktest netGap
All 8 matched dates agree on side; 7 of 8 on exit type (08-04: the paper trader read a 1-minute extreme at 15:57, two minutes before the backtest's 15:59 close). Exit minutes agree to within two minutes on every one. Largest point gap 33 pts (NDX vs NQ basis and futures-vs-cash drift over the day). This is the second-feed check: the signal engine behaves the same on a different instrument's quotes.
Darwinex Zero, NDX CFD, 0.05 lotsentries paused
DateSideSignalExitWhyP&L $
3 fills, $118 total at $0.50/pt with a $100 hard cap per trade. An account-wide entry pause on the Darwinex book blocked most sessions since 2026-07-30; exits always run. The 08-21 adverse exit lost $92.95, inside the cap.
TTP, QQQ shares on the NDX signalrunning
DateSideSignalQtyExitWhyEst $
7 fills, $1,061 estimated. Two "cat" exits are the $100-a-day breaker on that account, not a strategy exit. Dollar figures are the trader's estimate from NDX levels; broker statements are the truth.
Hyperliquid XYZ100 perphalted
Two fills (2026-08-13 long +$27, 2026-08-21 short −$3). Neither exit was a strategy exit: one was the exchange-side catastrophic stop, the other the dead-man flatten when the NDX signal feed went stale. The sleeve is halted by the account floor watcher since 2026-11-05. Not evidence either way.
11 · Live probation contract

Envelope. From the resample in section 8, a book of the next 26 trades (the prereg evaluation point) is expected at +664 pts net with a one-sigma band of ±524 pts on the sum, and a drawdown inside it of 564 pts (median) to 870 pts (95th percentile). A paper or live book inside the band is the sleeve behaving. Below the band is a costs, fills or feed question first, an edge question second.

What is watched weekly. Signal count against the four daily slots (silence is a feed or time-handling bug until shown otherwise), paper-vs-backtest match on every shared date, average winner alongside win rate (a hollowed-out edge keeps its hit rate), and live fill minus paper fill.

When the question gets asked. The frozen gates: 13 trades with PF below 0.7, or 26 trades with PF below 1.3 or net below zero. The rule decides when the question is asked. Pedro decides the action. Nothing is re-optimised inside a drawdown.

What isn't in this backtest

Backtest instrument ≠ live instrument. Every number in sections 2–9 is NQ futures on a ratio-adjusted continuous series. The live sleeves trade NDX CFD, QQQ shares and a perp. Section 10's matched-date table is the bridge.

Fills are next-minute opens. No intrabar fills, no slippage beyond the 1.0-pt cost. Section 4 shows the edge survives 5 pts per round trip; the live traders enter at market within the minute after the 10m close.

Three fitted constants. ±100 and 95 were chosen on the full history in July 2026. The 10:00 cutoff was adopted on 2026-08-29 after @etbent1's train/test split and our NQ confirmation; the prereg froze the 11:00 version, whose trade list and statistics are in the package and in section 6. Section 7 shows the neighbourhood around all three.

The 1-minute exit is evaluated on bars that open after the fill minute, exactly as the live trader does. The July discovery code may have included the fill minute's own bar; the difference is a handful of exits.

The live traders flatten at 15:57, the backtest at 15:59. Section 6 includes the 15:57 variant: PF 2.53 vs 2.55, total −38 pts.

The 200-point catastrophic stop is not in the headline. Section 9 sizes it: 12 trades reached 200 pts of adverse excursion in 511.

Data ends 2026-11-03. The paper sleeve's 11-05 short (+1,110 pts on NDX) is in section 10 but not in the backtest.

Live record is thin. 9 paper trades, 3 Darwinex fills, 7 TTP fills. Enough to check parity, not enough to judge the edge. The gates above say when it is.