🥇 Gold · published every week

Sessions backtests, gold

Each week we publish every US AM session Atlas Sessions (GC) read, captured twice: once at the open, when the tables state what they expect, and once at the close, when the outcome is settled. Right calls, wrong calls, and the sessions that split the difference.

These sessions are not in the sample. The Atlas Sessions (GC) tables were measured on gold history ending 6 August 2026. Every session reviewed on these pages happened after that date, so nothing here was available to the tables when they were built. This is where the out-of-sample record gets written, week by week, in public.

The running record

One row per week, newest first. Three scores, because the indicator makes three separable claims about a session. Click a week to read it in full.

WeekSessionsFirst sweepModal branch Levels takenStated P(take)RealisedInside sessions
w/c 14 September 2026 55 / 55 / 5 5 / 1050%50%0
w/c 7 September 2026 54 / 52 / 5 6 / 1054%60%1
w/c 31 August 2026 54 / 54 / 5 5 / 1055%50%0
w/c 24 August 2026 52 / 52 / 5 6 / 1047%60%0
w/c 17 August 2026 54 / 52 / 5 8 / 1054%80%0
w/c 10 August 2026 54 / 53 / 5 6 / 1048%60%0
Since 10 August 2026 3023 / 3018 / 3036 / 6051%60%1
How to read this table. Each row is one week of five sessions. Three scores, because the indicator makes three separate claims.
  • First sweep is a two-way call: of the levels that were touched, did the side the tables favoured go first.
  • Modal branch is the strictest score: did the four-way outcome land on the branch rated highest, out of high only, low only, both sides and inside.
  • Levels taken scores the high and the low independently, two events per session, against the marginal probability on each.
  • Stated P(take) is the mean of those marginal probabilities: the rate the tables expected the levels to go at.
  • Realised is the rate they actually went at, levels taken divided by the two events per session. Read it against Stated: the two should stay close as the record grows, and a gap that persists in either direction is the calibration finding.
  • Inside counts the sessions that took neither level.
A session can miss the modal branch and still resolve both its levels as expected, which is why all three scores are here rather than one summary figure.

This is a running tally over a small number of sessions, 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. The place to judge calibration remains the indicator's own realised versus expected row, running on your own chart over your own history.

Why US AM, on gold

Atlas Sessions (GC) reads all four gold sessions. These reviews fix on one, and the reason is measured rather than inherited from the NQ work.

2.30×
range concentration, against 0.97× for the next best session
35%
of the sessions' combined range traversed, in 15% of their minutes
209
minutes long, the shortest of the four by a wide margin
SessionClockMinutesShare of minutes Share of rangeRange concentration
Asia (SGE)18:00 ET to 13:30 CST44932.6%20.3%0.62×
Eurasia13:30 CST to 08:00 ET38928.3%27.3%0.97×
US AM08:00 to 11:30 ET20915.2%35.0%2.30×
US PM11:30 to 17:00 ET32923.9%22.5%0.94×

Measured over 9,984 gold session instances from 2015 to 2026. Range concentration is a session's share of range divided by its share of minutes, so 1.00 means a session covers exactly its proportional amount of ground and nothing more. Eurasia at 0.97 and US PM at 0.94 landing either side of 1.00 is a useful check that the method is not manufacturing the result. Ranges are expressed in ATR rather than points throughout, because gold's own range has grown roughly tenfold across this history and raw points cannot be compared across it. Gold needs three clocks, not one: the Asia and Eurasia boundary is set in Shanghai time, which has no daylight saving, so it moves against the New York clock twice a year.

Two honest qualifications

Eurasia forecasts better than US AM, and we still chose US AM. On calibrated forecast skill Eurasia beats it on every measure we ran, landing the modal branch 60.1% of the time against 54.1%. But that is a symptom of Eurasia being a quieter session rather than of the indicator working better there: with 32% less range reaching levels that sit at much the same distance, more of Eurasia's outcomes are settled by geometry alone, which is easier to predict and less useful to trade. Ranking sessions by how predictable they are systematically rewards the ones where little happens.

Asia is gaining ground and we are watching it. Gold's price discovery has been migrating east for years, and Asia's share of range has climbed about five points since 2015 to 18, the only one of the four that is rising. It still has not overtaken anything: per minute it remains the least dense session of the four, and it earns its share by running for 449 minutes rather than by being busy. If that changes, this page changes with it.

How to read the charts

Every chart on these pages is the indicator as it renders live, with nothing added and nothing removed.

  • The two levelsThe previous session's high and low, carried forward as draws on liquidity. For US AM that means Eurasia's. They are the only two levels being scored.
  • The % on a levelAn exclusive share, not the chance that level gets taken. Eurasia High | 11% means 11% of comparable sessions took the high and left the low alone. The chance the high is taken at all is high only plus both sides. This is the single most misread number on the chart, so it is worth pausing on.
  • The four-way rowsTake high only, take low only, take both sides, inside session. Mutually exclusive, always summing to 100, measured over comparable historical sessions at the same distance.
  • A ~ before a %Gold specific, and not a typo. The tilde marks a cell whose shares moved materially between market eras, so the indicator is telling you it is quoting an approximate figure rather than a settled one. Gold's history contains a genuine structural break in 2015, when the London fixes became electronic auctions, and the tables are built on the period after it. Cells that still look unstable say so.
  • First sweep H / LGiven that something is taken, which side goes first. A separate two-way question from whether each side goes at all.
  • Previous session summaryThe context the forecast is conditioned on: the prior session's range, whether it swept its own high or low or stayed inside, and where price opened relative to its midpoint.
  • TAKEN 08:20The latch. Once a level is taken the label records the time and stops updating, so the close capture carries the full outcome.
  • Med pen up / dnMedian penetration: how far past the level price typically travelled once it was taken. Measured drift, not a target.
  • Missing session nameThe previous session's box carries no name label, because its name and its high label anchor at the same point and print on top of one another. A known cosmetic fault, queued for the next release, and it affects no number on the chart.
  • TimeframeCharts here are 5-minute for legibility. The levels and percentages are identical on any timeframe, being defined by clock time and price rather than by bars, and this was verified across 1m and 15m before release. Only the precision of the TAKEN stamp follows the bar size.

The weeks

Newest first.

Week commencing Monday 14 September 2026

Five modal branches from five and five first sweeps from five, the first perfect week the gold record has had

The first time the gold tables have gone five from five on both calls; the best weeks before this one read 4 of 5 on each. Every morning took exactly one side, three the high and two the low, which is not itself a first: w/c 31 August was the mirror image, three low and two high, and still scored 4 of 5 on both, so a one-sided week is what a perfect modal week needs on this record rather than enough on its own. The stated marginals also separated cleanly for the first time, every taken level at 59, ~65, 72, 77 or 81 percent and every untaken one at 11, 12, ~32, 43 or 51, on ten events; the tildes are Tuesday's, whose cell carries the study's era flag. Levels still went only five of ten, joint lowest with w/c 31 August, and Tuesday's high was crossed by six tenths of a point.

Read the week →
Week commencing Monday 7 September 2026

The first inside session the gold record has produced, rated 7%, and a 112 point Friday

Monday's US AM took neither Eurasia level, the first inside session in twenty-five gold sessions, finishing 4.6 points under the high and 4.4 above the low on a branch rated 7%; the NQ book printed an inside session the same morning on a table that rated inside at 0%. With no level reached there is no first sweep to call, so the week's one first-sweep miss is a miss by construction: on the four days that swept anything the called side went first every time. The modal branch landed twice, Tuesday and Thursday; of the three misses, Monday took neither level and the other two were the called side going with company. Friday ran 112 points on a 30.2 point Eurasia, close to four times the range the levels were drawn from, with both levels gone by 08:40.

Read the week →
Week commencing Monday 31 August 2026

Four from five on both axes, and the first gold week where levels went less often than stated

Four first sweeps and four modal branches from five, the best modal week gold had produced at the time, and the one miss was Monday's double miss on the era-flagged cell, drawn for the third time in four weeks and resolving a third different way. Tuesday's 82.1 point Eurasia, the widest of the record, put 97% on the low and it went at ~08:30. Wednesday and Thursday drew the same cell on consecutive mornings and both took the high on the 08:00 bar itself. Levels went five of ten against a mean stated 55%, the first gold week under the stated rate, leaving the four-week record at twenty-five of forty against 51%.

Read the week →
Week commencing Monday 24 August 2026

Two first sweeps from five, the weakest gold week yet, and both sides landed on a 4% rating

The third gold week on record and the weakest so far: two first sweeps from five, with all three misses on calls of 64 to 70%, and two modal branches from five. Three cells repeated from earlier weeks, and one of them resolved exactly as it had the first time, Thursday going low only against the same 70% call on the high. Friday rated inside at 33%, the highest the gold record has shown, and instead took the high at ~09:55 and the low at ~10:05 either side of the Warsh speech, on a both-sides share of 4%. Levels went six of ten against a mean stated 47%, more often than stated for the third week running.

Read the week →
Week commencing Monday 17 August 2026

Four first sweeps from five again, and both sides kept landing on ratings of 13 to 29%

Three of five mornings took both levels on a branch the tables rated between 13 and 29%, which leaves both sides at four from ten over two gold weeks against a mean rating near 13%. The one first-sweep miss was a 65% call on the narrowest Eurasia of the record. Two tables repeated from the week before, one resolved the same way and one did not, and Friday's 97% call went at 11:25 with the inside branch five minutes from landing.

Read the week →
Week commencing Monday 10 August 2026

Four first sweeps from five, and both misses landed on inflation day

The first gold week on record. Four of five first sweeps called correctly and three of five modal branches, with both of the modal misses falling on the two mornings carrying an 8:30 inflation release. Wednesday took its high five minutes before CPI printed and its low on the release itself, which is not a branch any table rates highly. Nothing stayed inside all week, on a set of forecasts that expected it on up to 29% of mornings.

Read the week →
Next

Week commencing Monday 21 September 2026

Published at the end of the trading week.

What a week of five sessions can and cannot show. Ten level events cannot confirm a frequency measured over hundreds, and any one week can flatter or bury a perfectly well-behaved statistic. These pages exist so the record accumulates where you can see it, including the weeks that go badly. The place to judge calibration is the indicator's own realised versus expected row, running on your chart, over your history.