IB backtests, gold
Each week from August to September 2026 we published every 09:00 initial balance read on COMEX gold, at both the 30 and the 60 minute window, and scored them: which rail the tables favoured, which one price actually reached, and what happened on the days the indicator declined to call at all.
The running record
One row per week, newest first. Ten windows a week, five at each duration.
| What price did next | Atlas score | |||||
|---|---|---|---|---|---|---|
| Week | Windows | Rail reached | Favoured rail | 🧲 Atlas mode | ✅ / ❌ / ∅ | 30m / 60m agreed |
| w/c 28 September 2026 | 10 | 9 / 10 | 6 / 9 | 7 | 6 / 1 / 0 | 4 / 5 |
| w/c 21 September 2026 | 10 | 7 / 10 | 4 / 7 | 6 | 4 / 1 / 1 | 3 / 5 |
| w/c 14 September 2026 | 10 | 7 / 10 | 5 / 7 | 1 | 1 / 0 / 0 | 2 / 5 |
| w/c 7 September 2026 | 10 | 9 / 10 | 5 / 9 | 5 | 3 / 2 / 0 | 4 / 5 |
| w/c 31 August 2026 | 10 | 5 / 10 | 4 / 5 | 8 | 3 / 1 / 4 | 4 / 5 |
| w/c 24 August 2026 | 10 | 9 / 10 | 4 / 9 | 6 | 2 / 3 / 1 | 2 / 5 |
| w/c 17 August 2026 | 10 | 9 / 10 | 6 / 9 | 5 | 3 / 1 / 1 | 2 / 5 |
| w/c 10 August 2026 | 10 | 8 / 10 | 6 / 8 | 5 | 4 / 1 / 0 | 2 / 5 |
| Since 10 August 2026 | 80 | 63 / 80 | 40 / 63 | 43 | 26 / 10 / 7 | 23 / 40 |
- Windows is the starting count: ten readings a week, five sessions at 60 minutes and the same five at 30 minutes.
- What price did next scores all ten. Rail reached is how many windows saw price touch one of the two targets by the close; the rest reached neither, which is a stated outcome, not a failure. Favoured rail is, of the windows that reached a rail, how many reached the side with the larger stated share first. It is scored on every window, whether or not the indicator called it.
- Atlas score is the score that matters most. 🧲 Atlas mode is how many of the ten windows cleared the gate (direction at least 62% of the decided share, and at least half of comparable sessions resolving) and offered a direction. ✅ / ❌ / ∅ repeats the outcomes for those 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 / 4 means eight leans were offered: three were paid, one went the other way and four went unreached.
- 30m / 60m agreed is how many of the five mornings the half-hour and the hour favoured the same side.
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.
Why 09:00, and why not the window that looks strongest
The indicator ranks all 89 window and duration cells on a Window strength row. Reading that row from the top is a trap, and the size of the trap is measurable.
| Window (60m) | Favoured rail win | No target hit | Resolves | Target, 2025-26 | Atlas gate fires | Calls landed per week |
|---|---|---|---|---|---|---|
| 09:00 | 69.4% | 32.8% | 67.2% | 11.3 pt | 71.3% | 2.6 |
| 09:30 | 67.9% | 45.3% | 54.7% | 12.0 pt | 53.5% | 2.1 |
| 10:00 | 69.8% | 45.5% | 54.5% | 10.1 pt | 61.6% | 2.3 |
| 10:30 | 70.4% | 43.7% | 56.3% | 8.8 pt | 74.8% | 2.8 |
| 11:30 | 69.4% | 37.4% | 62.6% | 7.0 pt | 70.7% | 2.6 |
| 15:30 top of the ranking | 85.3% | 87.1% | 12.9% | 4.5 pt | 0.0% | none |
Measured across 117,133 gold races from 2015 to 2026, cross-fitted so that no window is graded on the table it helped build. Favoured rail win is the share of resolved races where the larger directional share was the rail price reached first. The 15:30 row is the point of the table: it tops the entire ranking at 85.3%, and it gets there by resolving on one day in eight, so that only the races already sitting next to a rail ever finish. Across all 44 sixty-minute cells the correlation between a window's no-target-hit share and its win rate is +0.95. Nearly all of what looks like strength is selection.
Two honest qualifications
09:30 is gold's busiest hour and we did not choose it. It packs more range into its box than any other window on the clock. But in this indicator the box is the target: a wider box pushes both rails further out, so the denser hour resolves on 55% of days against 09:00's 67%, and its gate fires on 54% against 71%. Density picks the box, resolution picks the trade, and they do not agree here.
10:30 beats 09:00 on the headline number. It wins the favoured rail 70.4% of the time against 69.4%, and lands marginally more calls per week. It also leaves 44% of days unresolved against 33%, on a target a fifth smaller. We publish 09:00 because it resolves most often on the largest target, and we would rather show the window that gives you the most decided days than the one with the prettiest percentage.
How to read the charts
Every chart on these pages is the indicator as it renders live, with nothing added and nothing removed.
- The boxThe initial balance: the high and low of the 09:00 window, either its first 30 or its first 60 minutes. Everything else is derived from it.
- The two railsProjected targets, set a multiple of the box height beyond each end of it. The race is which one price reaches first, or neither.
- ×1 autoThe multiplier, and the word auto matters. It is an output of the box regime, not a setting: a compressed box gets the full multiple, a normal one a half, an expanded one a quarter. A bare ×1.0 without auto would mean the multiplier had been forced by hand, and no chart on these pages carries one.
- ▲ / ▼ / ∅Up target first, down target first, no target hit. Mutually exclusive and summing to 100, measured over comparable historical sessions. They occasionally print as 101 because each is rounded to a whole percent.
- ATLAS MODEThe gate: direction must clear 62% of the decided share and at least 50% of comparable sessions must have resolved at all. Both conditions, which is why a lopsided but rarely-resolving window still abstains. When it does not fire the rail labels say abstain.
- Window strengthThe cross-fitted favoured-rail win for this cell and its rank of 89. Read it with the no-target-hit share beside it, never alone, for the reason set out in the table above.
- Regime tagsTwo of them, on two different rulers, and they legitimately disagree. The tag on the box compares this box to the last five sessions; the tag in the table compares it to the whole history. A box can be an expansion against this week and normal against eleven years.
- The outcome markA hit stamps immediately. A no target hit can only be confirmed once the window's race has run, and is drawn at the next session open, so a Friday window that took neither side shows a blank outcome column until Sunday evening. Friday's close has already settled it.
- TimeframeCharts are 5-minute or 1-minute depending on the week, whichever reads better for the morning; the box, the rails and the percentages are identical on any timeframe that divides the window length. Only the precision of the outcome time follows the bar size.
The weeks
Newest first.
Nine of ten windows leaned down, the most one-sided gold week on the record, and Atlas mode was paid on six of its seven calls
Every half hour leaned down and Monday's hour was the only window that leaned up. Atlas mode fired seven times and was paid on six, the most hits this record has drawn in a week, four of them on quarter-multiple targets that sit close by construction. The one miss, Monday's long hour, was crossed on the same 10:40 bar that paid Monday's half hour short. Nine of ten reached a rail; Thursday's hour reached neither.
Read the week →Eight of ten windows leaned down, then the most one-sided gold week on the record, and every half hour went to its lower rail
Every one of the five hours favoured the down side, a first for one duration here, and all five half hours reached the lower rail, including the two that leaned up. Atlas mode fired on six windows, five short and one long, and was paid on four of the five that resolved: Wednesday's long was crossed on the 09:45 bar and Wednesday's hour reached nothing. Seven of ten reached a rail; the three that did not were all hours, and two of them stopped 1.7 and 0.9 points short of the lower rail. Friday's two hits came on one 10:00 bar, the first after the hour's box, the half hour carrying a 5.6 point target, the smallest on the record. Nine of the ten cells had drawn before, the most of any week, and 13 / 53 / 34 drew twice, taking it to six draws on the record. No red folder release all week.
Read the week →Nine of ten windows declined to call, the most on the gold record, and the one that called was right
The quietest gold week published here. Nine of the ten windows abstained, against a previous high of five, and the mean stated favourite was 39, the lowest of the six weeks. The single Atlas call, Monday's half hour at 63 / 28 / 9, took the upper rail on the 11:30 bar. Seven of ten reached a rail and all three that reached nothing were 60 minute windows. No box on the week was expanded on the table's ruler, a first in six weeks, and Auto moved from a half multiple on Monday to a full one from Tuesday with no setting touched. Wednesday's 60 minute lower rail, reached on the 14:10 bar, is the only result on the page inside a news window.
Read the week →Nine of ten windows reached a rail, and the strongest lean of the week went the other way
The third gold week to come in at nine of ten, with the one miss a 1.7 point gap on a Labor Day session that shut at 14:30. Atlas mode fired five times and landed three: Wednesday's 81 per cent long call, the joint strongest lean the gold tables have shown, took the lower rail instead, and Thursday's hour repeated a 46 per cent cell and fell all afternoon. Four cells repeated and three of the four resolved differently from their earlier draw. Both of Friday's CPI morning windows abstained, on different gates, and both lower rails went.
Read the week →Every half hour resolved before 10:00, and none of the five hours reached a rail
All five 30 minute windows reached a rail by the 09:55 bar and Atlas mode landed three of the four it called there. All five hours reached neither rail by the 16:55 close, three of them stopping about three points under the upper one, including both draws of the same 72 / 12 / 16 cell. NFP expanded Friday's hour to a 70 point box, the widest on the record, and the ladder cut it to a quarter.
Read the week →Nine windows reached a rail again, and Atlas mode fired six times and landed two
The first gold week on which Atlas mode missed more than it landed: three magnet-and-cross marks in a row, Tuesday's 81% short call at a quarter multiple and the same 24 / 44 / 32 hour cell fired short on Tuesday and Thursday, all three going up. Friday's two calls paid, the hour's on the first bar after the box. Four repeat cells, and the record on repeats is now mixed on purpose.
Read the week →Nine windows reached a rail, and the one lean that missed was a 50% call
Three more repeat cells, one of which split: the 23 / 50 / 27 short call paid on Tuesday inside a minute of the box closing and was wrong on Thursday. The ladder cut both of Wednesday's expanded boxes to a quarter and both long calls paid before the FOMC minutes were anywhere near. Monday's hour was the only window to reach neither rail, a 46% lean that stopped five points short.
Read the week →The same forecast, twice, resolving both ways
Three of the ten windows returned a forecast the indicator had already given earlier in the week, to the percentage point and out of the same sample. Two of those pairs resolved in opposite directions, which is the clearest thing a week of this size can show about what a 53% call actually means. CPI day expanded both boxes and the ladder cut the multiplier to a quarter, where the short call paid.
Read the week →Monthly from October 2026
The weekly pages ended with the week of 28 September. The record continues as a monthly update.