A range, a sweep that fails, an inverse fair value gap to confirm it, then the real move — AMD, or the Power of 3. We coded it, found every setup in seven futures markets from 2019 to mid-2026, and measured what price did next. Two of the three phases are real. The one that is supposed to pay you is a coin flip.
The model has four phases. Price accumulates in a range. It then manipulates — taking out one side of that range and closing back inside it, the stop hunt. An inverse fair value gap confirms the turn: a gap that price closes back through, flipping it from support to resistance or the reverse. Then price distributes in that opposite direction, and that leg is the trade.
The range is where the teaching is vaguest, so we tested three readings of it, and every result below holds for all three:
Markets: NQ, ES, YM, GC, SI, CL and RTY, on five- and fifteen-minute candles. The sequence is common: about 60% of sessions produce a sweep that closes back inside the range, and an inverse fair value gap confirms roughly 75% of those.
The controls. A rate on its own says nothing, so every setup was run against four comparisons through the identical code:
The bar. Every test was written down before the data was touched. A result had to beat those controls with a confidence interval that excludes zero, and survive a correction for running thousands of tests — test enough things and some look good by luck.
One disclosure: the sweep and the inverse fair value gap had each been measured on their own in earlier studies here, and both came back empty. What had never been tested is this packaging — the range as the thing being swept, the session structure, and holding the trade into the close.
ATR — how far the market typically moves in one candle. Using it instead of dollars puts the Nasdaq and silver on the same scale.
R — the amount risked on a trade. “−0.34R per trade” means the average trade lost a third of what it risked.
A variation — one complete version of the rule: a market, a timeframe, a definition of the range, a sweep window, a stop and a target.
Costs — commission plus one tick of slippage on entry and exit, charged on every trade.
This test has no trade in it — no entry, no stop, no costs, nothing to argue about. Take every completed setup, stop at the moment the gap confirms, and measure two things: how far price travels in the direction the model promises, and how far it travels the other way. If distribution is real, the first number has to beat the second.
| After the setup · 1,514 setups, seven markets | The model’s way | The other way | Goes its way |
|---|---|---|---|
| Next hour | 1.96 ATR | 2.01 ATR | 49.9% |
| Next three hours | 3.65 ATR | 3.71 ATR | 49.7% |
| To the close | 5.80 ATR | 6.01 ATR | 49.9% |
| All three horizons | — | — | 49.8% |
A random moment at the same hour goes the same way 50.4% of the time. Of the 21 market-and-horizon cells tested, 0 were more likely than a coin flip to go the model’s way.
And the move isn’t even bigger. The total distance price covers after a setup is 0.99 times a normal move at the same hour. Volatility clustering after a stop hunt would at least have been something real; it isn’t there, because the setup requires a quiet range to form in the first place.
We traded it anyway, in every reasonable form: seven markets, two timeframes, three definitions of the range, two sweep windows, three stop sizes and eight exits — fixed targets from one to twelve ATR, the opposite side of the range, and simply holding to the close. That is 2,688 variations and roughly 710,000 trades, with commission and slippage charged on every one.
| Exit · averaged over every market and setting | Per trade | Before costs | Win rate |
|---|---|---|---|
| Target 5 ATR | −0.318R | +0.019R | 17.6% |
| Target 3 ATR | −0.320R | +0.016R | 24.9% |
| Hold to the close | −0.332R | +0.005R | 9.7% |
| Target 2 ATR | −0.340R | −0.004R | 32.1% |
| Target 8 ATR | −0.345R | −0.009R | 13.0% |
| Target 1 ATR | −0.353R | −0.017R | 46.6% |
| The opposite side of the range | −0.357R | −0.021R | 26.7% |
| Target 12 ATR | −0.358R | −0.022R | 10.9% |
Before a cent of cost, every exit sits at zero — between +0.019R and −0.022R. The rule does not lose because we charged too much; it has no edge to pay the friction with. Letting it run does not rescue it either: the biggest targets are among the worst, because the trade gives back the stop long before it reaches them.
Against its own controls. The full rule loses 0.34R a trade. Its exact mirror image — same setups, opposite direction — loses 0.39R. A range copied from five days earlier loses 0.34R. And the inverse fair value gap, in the version where the model does best, makes the trade 0.03R worse than just trading the sweep with no confirmation at all.
Yes, and it is worth showing why. The best of the 2,688 was the Nasdaq on five minutes, consolidation box, London sweep, one ATR stop, held to the close: +0.34R over 828 trades. Taken alone that has a p-value of 0.032. To survive 2,592 tests it would need a p-value below 0.00002 — it is off by three orders of magnitude.
It is the only market where it works. The same rule: NQ +0.11R, RTY −0.14R, CL −0.22R, GC −0.23R, YM −0.31R, ES −0.36R, SI −0.57R. A real mechanism does not switch off when you change contract.
It is direction and drift, not the setup. Inside that cell, longs make +0.37R and shorts +0.11R, and it loses money in 2019, 2020, 2021 and 2023 before making it all back in the years the Nasdaq trended hardest. It is “hold a long into the close in a bull market”. In the same cell, the plain sweep with no gap required makes +0.24R, so the confirmation is not doing the work either.
Ranges exist and sweeps happen — about 60% of sessions produce one. What does not exist is the move that is supposed to follow. Price after the setup goes the model’s way 49.8% of the time, travels 0.99 times a normal distance, and no version of the trade survives once you count how many were tried. The accumulation range is not special: one copied from five days earlier performs the same. The inverse fair value gap subtracts.
| What we predicted before running it | What happened |
|---|---|
| 1 · The sequence pays nothing after costs | right — 0 of 2,688 variations survived |
| 2 · The inverse fair value gap adds nothing | right — it subtracts 0.03R |
| 3 · The manipulation direction carries no information | right — the trade backwards performs the same |
| 4 · The accumulation range is not special | right — a five-day-old range matches it |
| 5 · Distribution is not a real phase | right — 49.8% directional, 0.99× the size |
The same brutal test — matched controls, real costs, one honest out-of-sample shot — run on what we sell: edges that survive it.
View our strategies