Weekly backtests
Each week from August to September 2026 we published every 09:30 cash-open window Atlas IB produced: the reading it gave a few minutes after the window closed, and what price actually did afterwards. Wins, misses and the sessions where it declined to lean, all of them.
The running record
One row per week, newest first. Every 09:30 window Atlas IB produced, at both durations. The first two weeks were read on Auto; from w/c 17 August the reviews use a fixed target multiple of ×0.5, the setting we now recommend for NQ. Click a week to read it in full.
| Did it offer a lean? | What price did next | Atlas score | |||||
|---|---|---|---|---|---|---|---|
| Week | Windows | 🧲 Atlas mode | Abstained | ✅ | ❌ | ∅ Neither | ✅ / ❌ / ∅ |
| w/c 28 September 2026 | 10 | 6 | 4 | 4 | 0 | 6 | 3 / 0 / 3 |
| w/c 21 September 2026 | 10 | 9 | 1 | 6 | 1 | 3 | 5 / 1 / 3 |
| w/c 14 September 2026 | 10 | 8 | 2 | 6 | 2 | 2 | 5 / 2 / 1 |
| w/c 7 September 2026 | 10 | 7 | 3 | 2 | 0 | 8 | 1 / 0 / 6 |
| w/c 31 August 2026 | 10 | 6 | 4 | 5 | 2 | 3 | 3 / 1 / 2 |
| w/c 24 August 2026 | 10 | 8 | 2 | 3 | 3 | 4 | 3 / 1 / 4 |
| w/c 17 August 2026 | 10 | 7 | 3 | 4 | 1 | 5 | 3 / 0 / 4 |
| w/c 10 August 2026 | 10 | 3 | 7 | 1 | 1 | 8 | 1 / 0 / 2 |
| w/c 3 August 2026 | 10 | 5 | 5 | 6 | 1 | 3 | 4 / 0 / 1 |
| Since 3 August 2026 | 90 | 59 | 31 | 37 | 11 | 42 | 28 / 5 / 26 |
- Windows is the starting count: ten readings a week, five sessions at 60 minutes and the same five at 30 minutes.
- Did it offer a lean? splits those ten. In 🧲 Atlas mode the indicator put a magnet on one rail and offered a direction. Abstained means it did not; the rails stayed grey. The two add up to the window count.
- What price did next scores all ten against the side with the larger stated share. ✅: that side's target was reached first. ❌: the other side's target was reached first. ∅: neither target was reached by the close. These three also add up to the window count. An abstained window is scored here too, even though no lean was offered on it.
- Atlas score is the score that matters most. It repeats the three outcomes, ✅ / ❌ / ∅, for the Atlas mode windows alone: how many reached the target they leaned to, how many reached the opposite target, and how many reached neither by the close. The three add up to the Atlas mode count. So 3 / 1 / 2 means six leans were offered: three were paid, one went the other way and two went unreached.
This is a running tally over a small number of windows, published as it accumulates. It is not a performance record and not a return: no entries, exits, position sizing or costs are modelled anywhere on these pages. Ninety windows cannot confirm or refute frequencies measured over thousands, and the place to judge calibration remains the indicator's own realised-versus-expected row running on your own chart.
Why the 09:30 cash open
Because it is the hardest window on the board, and rather than assert that, we measured it.
Atlas IB offers 46 anchors, every half hour from 18:00 through 16:30 New York time. We ranked all of them on one measure: the average one-minute high-to-low range inside the window, expressed as a share of price so that different price eras compare on equal terms, across the full 2009–2026 dataset.
09:30 comes first. Over sixty minutes it averages 9.8 basis points of range per minute against 8.2 for 10:00, the runner-up, and roughly 3.6 × the day's median window. Over thirty minutes the shape is identical: 10.6 against 8.9, about 3.9 × the median.
And it is not an artefact of the present regime. Split the seventeen years into 2009–2015, 2016–2021 and 2022–2026 and 09:30 ranks first in every one of them, at both durations, with the same 16–25% margin over 10:00 each time.
Which makes it the hardest test
A statistic that holds up in quiet tape is not saying much. The open is where the range is widest, the targets are furthest away in absolute points and a session can invalidate a lean within minutes. If the measured frequencies are going to fail anywhere, this is where.
It is where the numbers earn their keep
It is also the window most people actually trade, and the one where a plan is hardest to form in the moment. Having the split and the target distance fixed a few minutes after the window closes is worth more here than anywhere else in the day.
Why a 60-minute and a 30-minute window
Atlas IB is not fixed to the cash open. The window anchor can be set to any of the 46 half-hourly slots from 18:00 through 16:30 New York time, at a duration of either 30 or 60 minutes, so you can point it at whatever part of the day your routine actually covers, from the Asia open to the afternoon.
For these reviews we fix on two: 09:30 + 60m and 09:30 + 30m. Holding the window constant is what makes one week comparable with the next, and the cash open is the demonstration that means the most. Showing both durations is deliberate, because they are separate measurements of the same morning and they regularly disagree, which is itself part of the picture. The 30-minute window commits earlier and to a nearer target; the 60-minute window waits longer and aims further. Neither is the "right" one. Pick the one that fits how you trade, and read these pages for what both were saying.
How to read the charts
Every chart on these pages is the indicator as it renders on a live chart, with no annotations added, nothing removed.
- The boxThe window's high and low. Its header gives the range in points, the average range of the last five sessions, and a behaviour tag, 💰 ACCUMULATION, ⚖️ NORMAL or 🚀 EXPANSION, comparing this window with those five.
- ▲ / ▼ badgesThe two dotted rails are the extension targets, projected above and below the box. Each badge carries the share of comparable historical windows in which that side's target was reached first, with the same split measured over shorter horizons in brackets: 2h and 4h.
- The line beneath▲ up first · ▼ down first · ∅ neither, plus n, the number of comparable historical windows behind the figures. The three shares are measured to the session close and sum to 100 before rounding; each is rounded to a whole percent, so a printed set can read 99 or 101.
- Target distanceThe rails sit a multiple of the window's range beyond it, and the multiple is printed in every capture's status table. From w/c 17 August the reviews use ×0.5, the setting we now recommend for NQ. The first two weeks ran on Auto, which stepped the multiple by regime: 1.0 × for a compressed window, 0.5 × for a normal one, 0.25 × for an expanded one.
- abstainWhen neither side clears the lean test (a 62% share of the comparable sessions that resolved, with at least half resolving at all) both rails go grey and are labelled abstain. The indicator is saying it has no directional lean for this window. These are common, and they are not failures.
- 🧲 Atlas modeWhen one side does clear both bars, the favoured rail carries a magnet. These are the windows where the history hands you a genuine lean.
- ✅ ❌ ◇The outcome mark, printed at the bar where the race resolved. ✅ the higher-share side was reached first, ❌ the other side got there, ◇ both in the same minute. A 🧲 in front of the mark means it was an Atlas-mode window; a bare mark means the window had abstained and the mark is grading the larger of two shares that were both weak. If neither target is reached before the session closes, no mark prints, and that is the ∅ outcome.
- Faint guidesThe dotted 15m / 30m / 60m lines beyond each rail are median run-on distances, how far price typically drifted past that target over those horizons. They are drift, measured; they are not a continuation forecast.
The weeks
Newest first.
Six leans, four abstains all at 61%, and every rail reached was reached by 10:45
Six of the ten readings cleared the lean test, three long and three short; the four abstains each printed 61% directional, one point under the floor. Four windows reached a rail, all on the favoured or larger side and all by the 10:45 bar, and no lean was crossed against; the other six reached neither all day. Thursday's half hour was paid on the first bar after its box, equalling the earliest resolution on the record, and Monday's hour at 10:45 equals the earliest an hour has resolved. Tuesday read one 180-point box two ways, an upper larger share at the hour and a lower one at the half hour, and price reached neither rail.
Read the week →Nine leans of ten, the most on the record, and every half hour reached a rail while three hours reached nothing
Nine of the ten readings cleared the lean test, six long and three short, more than any week the record has published; Thursday's half hour was the only abstain, at 58%. Seven windows reached a rail, six on the favoured or larger side and one against. All five half hours resolved, the first clean sheet for the shorter window, and the three blanks were all hours, including Wednesday's 372.75-point box, the widest since the multiple was fixed, whose rail sat 186 points away and was never within 192. Thursday's two hits are one 12:15 bar, Monday's hour was cleared by 0.625 points, and Friday's half hour was crossed against at 10:15 by 3.75 points, the earliest cross on the record. Three cells drew twice and each pair matched. A week proves nothing on its own.
Read the week →Eight of ten windows reached a rail, and two of the four Atlas-lean crosses on the record are here
Eight of the ten readings cleared the lean test, five long and three short. Eight windows reached a rail before the close, more than any week the record has published, and only two ended inside both targets, the fewest so far. Of the eight leans, five were paid and two were crossed against: no week had paid more than four or been crossed against more than once. Counted across all ten windows rather than the leans, the two crosses are behind the three of w/c 24 August. Friday's half hour was settled on the first bar after its box closed, at 10:00, the earliest resolution on the record, while Thursday lost both windows off one 15:50 high, by 15.8 points and by 10.5. Two cells drew twice and split: 56 / 28 / 16 paid on Monday and was crossed on Tuesday, 44 / 20 / 36 paid on Wednesday and reached nothing on Thursday.
Read the week →Eight of ten windows reached no rail, and the two ticks are the same Monday move
Seven of the ten readings cleared the lean test, four long and three short, and one was paid: Monday's hour, on the smallest sixty-minute box the record has, 44.5 points in a Labor Day session that closed at 13:00 ET. Its 11:20 fall crossed the half hour's rail on the same bar, so the two ticks on the page are one price move. Eight windows reached neither target, equalling the quietest week on the record, and the only week yet with no cross came within 1.5 points of having one: Thursday's half hour leaned short and the day's high stopped just under the upper rail. Thirty-six Atlas-mode windows on the record: fifteen paid, nineteen unreached, two against.
Read the week →Six leans, three paid, one beaten on payrolls Friday, and two boxes that served four windows
Six of ten windows cleared the lean test, all but one long, and every resolution came by 11:20. Three leans were paid, two reached neither, and Friday's half hour, the strongest cell on the page at 65 / 16 / 19, was marked against by the same bar that ticked the hour's abstained larger share. On Thursday and Friday the hour did not extend the half hour, so both durations shared rails and settled together. Twenty-nine Atlas-mode windows on the record: fourteen paid, thirteen unreached, two against.
Read the week →Eight leans, three paid, and the first lean marked against, on the last bar of the day
The most Atlas-mode windows on any week so far: eight of ten cleared the lean test, five short and three long. Three reached their target, four reached neither, and Wednesday's hour was marked against when the 16:55 bar crossed the upper rail, the same bar that paid the half hour's long lean. The 28 / 53 / 19 half-hour cell returned for a fifth and sixth session and split them. Twenty-three Atlas-mode windows on the record: eleven paid, eleven unreached, one against.
Read the week →Seven short leans, three paid, none beaten, and the same half-hour read four days running
The busiest Atlas-mode week so far: seven of ten windows cleared the lean test, every one of them short. Three reached their target, four reached neither, and none resolved against the lean. The 30-minute window returned an identical 28 / 53 / 19 on four consecutive mornings and split them two and two. The first week captured at ×0.5.
Read the week →Seven abstains from ten windows, and a week that moved without going anywhere
The hardest week we have published. Eight of ten windows reached neither target, one Atlas-mode lean from three was paid, and Tuesday travelled three hundred points off the box without completing a run in either direction. The abstains were the useful part.
Read the week →Three Atlas-mode mornings, one red mark, and an NFP week that refused to travel
Ten windows across five sessions. Every Atlas-mode race that resolved went to the favoured side; the week's only ❌ landed on a window that had already abstained; and Thursday's 60-minute target went unreached all session, finally trading through eight minutes after Friday's payrolls print.
Read the week →Monthly from October 2026
The weekly pages ended with the week of 28 September. The record continues as a monthly update.