Delta Terminal Research · Paper #3 · October 2026

Does the Asia Range Breakout Fail 73% of the Time?A 2,303-breakout test on XAU/USD

A rule many gold day traders carry into the London session, and one our own morning briefing repeated: no conviction before 14:30 Paris, because about 73% of breaks of the Asia range before 14:25 fail, so a morning range exit is more likely a trap than the start of a trend. We took the rule apart into its three claims, wrote down what we expected, counted every break of the Asia range over three and a half years against two placebo levels and the afternoon, and let three auditors attack the first run. The 73% is real. It is also true of any level, at any hour.

Key findings

Yes, the number reproduces. A wick beyond the Asia high or low between 09:00 and 14:25 is followed by an M5 close back inside the range within 30 minutes 75.7% of the time [72.2; 78.9] on normal days (n = 630); the original script re-run on the frozen store gives 74.6%, and 73.1% when cut at the 641 normal days it had in July 2026.
But it is neither "Asia" nor "morning". The same definition gives 69.0% for afternoon breaks of the same range, 76.0% for phantom levels half a range-width outside the Asia range that nobody watches, and 73.9% for the London first-hour range. It is the M5 poke-and-fade base rate around any level: 48.6% of those wick breaks are bars that already close back inside.
At the close, a coin flip. With a close break, the day ends back on the original side 48.2% of the time [44.1; 52.3] in the morning and 43.1% in the afternoon; phantom levels 48.9%, London first hour 48.4%, a driftless walk 47.3%. The distance from the level at the close is symmetric (median +0.03 ATR, IQR [−0.25; +0.31]), and the afternoon move is uncorrelated with the morning's (ρ = +0.02).
The hour explains nothing at the horizon that matters. Morning minus afternoon failure, close breaks: +6.9 points [+0.2; +14.1] at 30 minutes, +5.1 [−2.2; +12.0] at the close; +19.3 in 2023 and about zero every year since. Traversal of the whole range within four hours: 9.8% versus 7.8% (+2.0 [−2.2; +6.2]); the pre-registered "34% vs 10%" gap at end of day was a time budget.
Trap and trend are equally likely. After a morning close break the day closes at least 0.25 ATR beyond the level 29.4% of the time and at least 0.25 ATR back on the other side 24.5%; two thirds of breaks never see the opposite edge broken; one in four is reversed through the whole range by the US session. 76.5% of trend days had their first close break before 14:30.

What traders believe

The Asia range is the high and low built overnight, here between 01:00 and 08:55 Paris. The London morning then "tests" it, and a well-known rule says those tests mostly fail: wait for New York, because a morning break is a liquidity grab, not a trend. Our own briefing stated it in numbers on 20 October 2026: "No conviction before 14:30. Statistically, about 73% of Asia-range breaks before 14:25 fail. A range exit this morning is therefore more likely a trap than the start of a trend." The 73% came from an earlier script of ours, written for a study of pre-release mornings: a "break" was any wick beyond the Asia high or low, a "failure" one M5 close back inside the range within the next six bars.

The sentence contains three claims. C1, the number: about 73% of morning breaks fail, and that is a property of the break rather than of the way it was counted. C2, the hour: breaks before 14:30 fail more than breaks of the same range after 14:30. C3, the reading: a morning break is more often followed by a return into the range or a trip to the opposite edge than by continuation to the close, and is not, in the main, how trend days start.

We wrote our expectations down before computing anything. C1: the 73% is largely a definition artefact; with a close break and a reasonable horizon the failure rate falls to 45–60%, and at the close the day stays on the break side 50–60% of the time. C2: afternoon breaks fail at a similar rate (within ten points) at equal horizon; what changes in the afternoon is amplitude, not reliability. C3: part of the morning breaks are the start of trend days; the trap is mostly the wick without a close; and we expected a narrow Asia range and a decisive close to carry real information. Half of that survived.

How we tested it

Mechanical definitions

A trading day runs 01:00 to 23:00 Paris; the Asia range is the high and low of the 01:00–08:55 bars; end of day is the last bar before 23:00. Two break definitions, both published: W, the first bar whose high exceeds the Asia high (or low undercuts the Asia low), the original study's rule; C, the first bar whose close is beyond the level. First break per side and per window; the morning window is 09:00–14:25, the afternoon window 14:30–23:00 on the side the morning left untouched. Four failure horizons: E30 and E60, at least one close back on the original side within six or twelve bars; E_win, before the end of the window; E_EOD, the day's close back on the original side (inside the range or beyond the opposite edge). Plus traversal, the opposite edge touched, at fixed horizons of 2 h, 4 h and 5 h 25 and at end of day; continuation, a day close at least 0.25 prior-day ATR beyond the level; the maximum excursion before the first close back; and the minutes until it. Every normalisation uses the previous day's 14-day ATR.

Three controls, defined before the effect

The hour control tests C2: the same definitions applied to afternoon breaks of the same range, compared at equal horizons only. Phantom levels test whether the Asia edge is special: a level placed half a range-width above the Asia high and below the Asia low, a price nobody has drawn, broken in the same window and scored the same way (median distance from price 0.21 ATR). The London first-hour range, 09:00–09:55, broken between 10:00 and 14:25, tests "recent range" against "Asia range". The 73% is a normal-day figure, so every headline cell uses the 742 days without a US release and the 181 release days are shown separately.

Pre-registration and three audits

Definitions, windows, horizons, controls and the seven headline metrics were frozen in writing on 20 October 2026 before the first result; the pre-registered run produced 614 comparisons under the rule "one published estimate, one comparison". Three independent reviews then attacked it: a blind replication that recounted every one of the run's 4,368 event rows from the raw bars and matched them all, with every headline cell within one point; a code audit, which cleared look-ahead and symmetry and found two blocking flaws in how we compared the morning with the afternoon; and an adversarial pass that tried to make each emerging conclusion disappear, then tried to rescue the briefing sentence by searching 61 morning strata for one that would support it. The corrections and the post-hoc work add 384 labelled comparisons; no pre-registered cell was altered.

Results

A. The number reproduces as a definition

On the frozen store the original script gives 74.6% fake breaks on 744 normal days; truncated at its 641st normal day, 73.1%, the published figure. Under the study's own filters the same cell, wick break plus close back inside within 30 minutes, reads 75.7% [72.2; 78.9] on 630 morning breaks. The number is honest. It is also a function of two choices.

The first is the horizon. Any "at least one close back inside the interval" rate rises with the interval, by construction: 75.7% within 30 minutes, 80.8% within an hour, 87.5% before 14:25. Only the state at the close describes the day, and there the wick-break "failure" is 50.2%. The second is the break itself. 48.6% of wick breaks are pokes whose own bar closes back inside; for them the 30-minute failure is 90.2%, against 62.0% when the bar closes beyond. At the close the two populations are indistinguishable, 50.7% and 49.7%. Requiring a close beyond the level removes the pokes: 62.9% at 30 minutes, 71.5% at an hour, 81.3% before 14:25, 48.2% [44.1; 52.3] at the close (n = 571).

Line chart of failure rate against four horizons for wick and close breaks, morning and afternoon: all four series rise from 56 to 76 percent at 30 minutes to 77 to 88 percent before the end of the window, then fall to 43 to 50 percent at the day close. — F2. Failure rate by horizon for wick and close breaks, morning and afternoon, normal days. The rate is set by the horizon and by the break definition; the hour moves it by a few points at 30 and 60 minutes and nothing at the close. End-of-window cells (shaded) compare a 5 h 25 window with an 8 h 30 one and are shown, not contrasted.

Our pre-written expectation was half right: the mechanism is the poke plus the short horizon, and the close-break rate at 30 minutes landed in the 45–60% band. But "a reasonable horizon" does not lower a return rate, it raises it; what falls is the state at the close, and it falls to a coin flip.

B. The hour not confirmed

Same range, same definitions, breaks of the side still untouched at 14:30: 295 close breaks and 294 wick breaks on normal days. At 30 minutes, afternoon close breaks fail 55.9% against 62.9% in the morning, +6.9 points [+0.2; +14.1]; wick breaks 69.0% against 75.7%, +6.7 [+0.3; +13.4]. At an hour, +6.0 [−0.5; +12.8]. At the close, 43.1% against 48.2%: +5.1 [−2.2; +12.0] for close breaks, +1.2 [−6.1; +7.8] for wick breaks. Removing the 62 days on which the US clock was an hour out of step with Europe changes no cell by more than a point.

The pre-registered run also showed morning breaks reaching the opposite edge of the range before the close 34.0% of the time against 10.2% in the afternoon, a +23.8-point gap with an interval far from zero, and we had read it as the morning's unreliability. The code audit pointed out that a 09:30 break has up to fourteen hours to cross the range and a 16:00 break less than seven. At equal horizon the gap is gone: within two hours, 3.3% against 5.4% (−2.1 [−4.9; +0.6]); within four hours, 9.8% against 7.8% (+2.0 [−2.2; +6.2]); within 5 h 25, the length of the morning window, 15.6% against 9.5% (+6.1 [+1.5; +10.8]), where the afternoon is already truncated by the close, with 58 bars available out of 65. For wick breaks the afternoon traverses more at two hours (−7.6 [−11.6; −3.8]) and the same at 5 h 25 (0.0). Inside its own window, a morning break crosses the range before 14:25 only 6.5% of the time: when the "trap" closes, it closes in the US session, and no more often than after an afternoon break.

Grouped bars of traversal rate for morning and afternoon close breaks at two hours, four hours, five hours twenty-five and end of day: 3.3 vs 5.4, 9.8 vs 7.8, 15.6 vs 9.5, then 34.0 vs 10.2 in the shaded unequal-horizon cell. — F4. Share of close breaks reaching the opposite edge of the Asia range within a fixed number of bars, morning against afternoon. The end-of-day cell is the pre-registered one; it compares a fourteen-hour budget with an eight-and-a-half-hour one and is kept as a description, not a contrast.

What does differ in the afternoon is amplitude, in both directions. The median excursion beyond the level before the first close back is 0.075 ATR for a morning close break and 0.135 ATR for an afternoon one; two hours after the break the close sits at least 0.25 ATR beyond the level 10.5% of the time in the morning and 21.7% in the afternoon (−11.2 [−16.8; −6.0]), while "close back at or below the level" at the same clock reads a flat 42.2% against 40.3%. The US session moves more; it does not fail less. Year by year, the end-of-day gap in favour of the morning is +19.3 points [+4.3; +32.5] in 2023, then −0.5, +1.0 and 0.0 in 2024, 2025 and 2026. What remains attributable to the hour is a re-close rate seven points higher in the first 30 to 60 minutes, with an interval that touches zero, and nothing at the close.

C. Trap or trend refuted

Take the 571 morning close breaks on normal days and ask where the day ends, in prior-day ATR beyond the broken level. The median is +0.026 ATR, the interquartile range [−0.25; +0.31]: a distribution centred on the level. The day closes at least 0.25 ATR in the break's direction 29.4% of the time [25.8; 33.3] and at least 0.25 ATR back the other way 24.5%; 46.1% of days end within a quarter of an ATR of the level. The "failures" are not small re-entries: only 22.9% of them end within 0.10 ATR of the level, against 16.6% of the "successes". They are the left half of a symmetric distribution. A gaussian walk with no drift, started at the breaking close with the day's remaining bars, would put the close back on the original side 47.3% of the time; the observed figure is 48.2%. And the afternoon does not undo the morning, it ignores it: the correlation between the 09:00–14:25 move and the 14:30–close move is +0.02 over 742 normal days, and after a morning move of at least 0.3 ATR the afternoon goes the other way 48.4% of the time, against 49.7% unconditionally.

Histogram of the day close minus the broken level in ATR for morning close breaks of the Asia range, overlaid with phantom-level and London-first-hour outlines: all three are symmetric around zero with the same spread. — F3. Where the day closes relative to the broken level, signed in the break direction, for morning close breaks of the Asia range and of the two placebo levels, normal days. Phantom median +0.009 ATR, London first hour +0.012: the same distribution.

The doctrine's "trap" has two possible meanings and we scored both. The first is the opposite edge touched. Decomposing the 571 breaks by what happens to the other side of the range: 9.1% (52) are morning chop, both edges broken before 14:25, on 26 days or 3.5% of normal days; 23.8% (136) see one edge broken in the morning and the other in the afternoon, the break reversed through the whole range by the US session, with the close back on the original side 95.6% of the time and continuation 1.5%; 67.1% (383) never see the other edge broken, with the close back inside 29.2% [24.9; 34.0] of the time and continuation 39.9% [35.2; 44.9]. The second meaning is the deep failure, a close at least 0.25 ATR back on the other side: 24.5%, against 29.4% continuation. Under either definition the trap is no more frequent than its opposite.

Did the morning break "pay" before failing? Half of the morning close breaks never extend more than 0.075 ATR beyond the level before the first close back (about four dollars on a fifty-dollar ATR); 17.9% reach 0.25 ATR, 8.9% reach 0.5 ATR; the median time to the first close back is 15 minutes and 8.9% never close back. Phantom levels: 0.081 ATR, 20.7%, 10.3%. London first hour: 0.065, 17.7%, 7.4%. Identical.

And the trend days. A trend day is one whose close sits at least one ATR from the 01:00 open: 11.5% of normal days (85 of 742). Every one of them broke the Asia range on a close, which is arithmetic, a one-ATR move leaving a range whose median width is 0.42 ATR. 76.5% of them [66.4; 84.2] had their first close break before 14:30 (65 of 85). The control reads the other way: 73.1% of non-trend days also had their first close break before 14:30, so an early break is simply the norm. Conditionally, a morning close break turns out to be the start of a same-direction trend day 10.7% of the time (61 of 571) and of an opposite-direction one 1.6%. Three trend days in four began with a morning break; one morning break in nine began a trend day. Neither reading supports "more likely a trap than the start of a trend"; the only defensible sentence is symmetric.

D. Width, strength and the placebo levels base rate confirmed

Every cell of the headline table has a placebo twin and the twins agree. Wick plus 30 minutes: Asia morning 75.7%, phantom 76.0%, London first hour 73.9%; the differences are −0.3 [−6.8; +6.5] and +1.9 [−2.2; +6.0]. Close plus end of day: 48.2 / 48.9 / 48.4%, differences −0.8 and −0.3. Continuation: 29.4 / 26.6 / 28.9%. No Asia-minus-placebo gap leaves a band of three points around zero. The pokes are there too: 54.7% of phantom wick breaks close back inside on their own bar, and their 30-minute failure is 94.3%.

Grouped bars for four populations, Asia morning, Asia afternoon, phantom levels and London first-hour range: wick-plus-30-minute failure 75.7, 69.0, 76.0 and 73.9 percent; close-plus-end-of-day failure 48.2, 43.1, 48.9 and 48.4 percent. — F1. The same two cells for the Asia range in the morning, the Asia range in the afternoon, phantom levels half a width outside the range and the London first-hour range, normal days, Wilson 95%. The dotted line is a coin flip.

We had expected a narrow Asia range to make a break more durable. Traversal of the range by the close does fall with width, 48.9% for the narrowest tercile, 31.6% and 16.0% for the others, but the maximum adverse excursion in ATR is flat across terciles (median 0.261 / 0.247 / 0.260 ATR), and feeding each day's width into the pooled excursion distribution predicts 47.6 / 31.3 / 15.1%, within a point of what was observed. A narrow range does not pull price back; its far edge is simply closer. The one residual that is not geometry runs against the expectation: continuation rises with width, 25.1% to 37.2%, one cell in thirty-one, intervals almost disjoint, a lead to confirm and not a result. A decisive close beyond the level (top tercile, more than 0.028 ATR) triples the median time to the first close back (5 to 30 minutes) and the excursion (0.028 to 0.157 ATR) and changes nothing at the close: 50.0 / 46.3 / 48.1% back on the original side, 27.8 / 29.3 / 31.2% continuation. Strength changes the quarter hour, not the day.

Asia width tercile	Breaks	Traversed EOD	Geometry predicts	Median MAE	Cont. ≥ 0.25	Back at EOD
Narrow (0.29 ATR)	219	48.9%	47.6%	0.261 ATR	25.1%	51.6%
Middle (0.42 ATR)	196	31.6%	31.3%	0.247 ATR	28.1%	46.9%
Wide (0.64 ATR)	156	16.0%	15.1%	0.260 ATR	37.2%	44.9%

Morning close breaks, normal days; terciles of width in prior-day ATR (median per tercile in brackets; edges 0.357 and 0.496 over all valid days). MAE = maximum adverse excursion to the close, in ATR. "Geometry predicts" is the share of days whose pooled maximum adverse excursion would exceed their own range width.

What the audit did to our own numbers

The first run had three findings in the morning's disfavour that looked solid. Two were artefacts of our comparison, one survived in a weaker form, and the briefing sentence that started the study did not survive at all.

The traversal gap, +23.8 points. Pre-registered, interval [+18.4; +29.0], and we had listed it as the morning's main weakness. It compared a morning break's fourteen hours of remaining day with an afternoon break's eight and a half. At four hours it is +2.0 [−2.2; +6.2]. The end-of-day cell is kept and labelled unequal horizon; the four-hour cell is the headline.
"Afternoon re-breaks fail far less than morning breaks." The first run's afternoon-all series counted 897 of 1,287 "re-breaks" at exactly 14:30: the state of a morning break still open, not a break. Counted after the first close back, the morning-minus-afternoon 30-minute gap falls from +16.7 to +11.5 points for close breaks; requiring a genuine transition bar, +3.9 [−1.4; +9.4], and +3.9 [−0.0; +7.5] at the close. The sentence is withdrawn.
The rescue attempt. The adversarial audit searched every stratum of hour block, break definition, width tercile and direction with at least 25 events, 61 in the morning and 35 in the afternoon, for a cell with at least 65% failure and a lower bound of at least 55% at the close or in traversal. None. The best morning cell is 62.5% on 32 events with a lower bound of 45%. Only the 30-minute and 60-minute definitions reach 65%, and the afternoon reaches them too.
What survived. The 73% as a base rate; the coin flip at the close, strengthened by the phantom levels and the driftless benchmark; the seven-point morning excess in the first hour, with its interval on zero; and a decline in intraday mean reversion from 2023 to 2026 (median adverse excursion to the close 0.37 to 0.19 ATR, raw traversal 48% to 20%), about half of which is the drift toward wider Asia ranges and half a real fall, visible inside 2023–24 by semester and on the release-day series too.

All 4,163 pre-registered cells of the results file are unchanged between the first run and the audited one, verified leaf by leaf; the corrections sit beside them under their own labels. We publish the first reading because the correction is the finding, and because the sentence we tested was ours.

Two panels: left, raw and width-standardised traversal falling from 48 and 41 percent in 2023 to 20 percent in 2026 with median adverse excursion falling from 0.37 to 0.19 ATR; right, the morning-minus-afternoon end-of-day gap of plus 19.3 points in 2023 with an interval above zero, then minus 0.5, plus 1.0 and 0.0 with intervals spanning zero. — F5. Year by year, morning close breaks on normal days. Left: raw traversal, traversal standardised to the pooled width mix, and the median maximum adverse excursion to the close. Right: the end-of-day failure gap, morning minus afternoon, with its day-block bootstrap interval. 2026 runs to 3 August (92 breaks).

How to use this

Nothing here is a signal; no entry, stop, target or cost was simulated. What the data offers is a calibrated expectation for a break of the Asia range on XAU/USD, by the hour at which it closes beyond the level.

Close break at (Paris time)	Breaks	Back on original side at EOD	Cont. ≥ 0.25 ATR at EOD	Opposite edge within 4 h
09:00–09:55	267	50.9%	29.6%	7.1%
10:00–10:55	109	45.9%	28.4%	5.5%
11:00–12:55	114	45.6%	28.9%	17.5%
13:00–14:25	81	45.7%	30.9%	13.6%
14:30–15:25	89	42.7%	41.6%	11.2%
15:30–17:25	130	46.9%	30.0%	8.5%
17:30–22:55	76	36.8%	18.4%	2.6%
All morning (09:00–14:25)	571	48.2%	29.4%	9.8%
All afternoon (untouched side)	295	43.1%	30.5%	7.8%

Close breaks of the Asia range, normal days. Morning rows are the first break per side between 09:00 and 14:25; afternoon rows are breaks of the side the morning did not break. The late block has little day left, which lowers both columns at once. Wilson intervals on cells of 76 to 267 events run roughly ±6 to ±11 points.

Expect the poke; do not read it. Three wick breaks in four close back inside within 30 minutes, and so do three in four at a level nobody is watching. A wick beyond the Asia high tells you that gold mean-reverts on M5 around any level, which you knew.
Expect a coin flip at the close, morning or afternoon. After a close break the day ends back on the original side about half the time, 48% in the morning and 43% in the afternoon, and continues by a quarter of an ATR or more about three times in ten, at either hour. The data does not support waiting for 14:30 on grounds of reliability; it supports expecting a larger move, in either direction, once New York is in.
Define the trap before you fear it. One morning break in four is reversed through the whole range, and that happens in the US session; two in three never see the other edge broken. If the trap is a deep failure, it is as frequent as a deep continuation: 25% against 29%.
Do not read the width or the candle. A narrow range is traversed more often because its far edge is nearer, not because price is pulled back; a decisive closing bar buys minutes, not a day.

The sentence we now use instead

"A close outside the Asia range this morning is neither a trap nor a signal. On XAU/USD since 2023, the day closed back on the original side about 48% of the time after such a break, and about 43% after an afternoon break: the hour changes nothing at the close. Continuation of at least 0.25 ATR, 29%; a return of at least 0.25 ATR the other way, 25%. The '73% fail' figure describes a wick that closes back inside within 30 minutes, at the same rate in the afternoon (69%) and at any level (76%). When a morning break is reversed through the whole range, it is the US session that does it, one break in four. No directional bias follows from the break."

Limits

One definition of the range. 01:00–08:55 Paris, the original study's convention, kept for comparability; no sensitivity on 00:00 or 09:00 edges. 01:00 is not the session open (913 of 923 days trade from 00:00); "open" here means the 01:00 bar.
Release calendar. The list of US release days has 94, 61, 23 and 3 dates for 2023 to 2026, so "normal" days in 2025–26 include genuine data days. 2023–24 is the full-calendar period and is published separately: morning-minus-afternoon end-of-day gap +8.8 [−2.2; +18.1] there, +0.9 [−8.9; +11.2] in 2025–26. Release days traverse and continue more (48.4% and 36.6%); they do not fail more (49.0% at the close).
Horizons. Traversal by end of day is an unequal horizon between the morning and the afternoon and between hour blocks; only the fixed-horizon cells support a contrast, and at five hours or more the afternoon is itself truncated. E30 and E60 are capped at the day, not the window, so 88 morning breaks see their 30-minute window spill past 14:25 (13 rows where it changes the outcome).
Selection in the afternoon series. The primary afternoon series is the side the morning left alone, a selection of quiet mornings for that side. It is nil at the close (43.4% on days with a morning break of the other side, 42.8% without) but not at 30 minutes (63.2% against 49.7%).
Placebo geometry. The London first-hour range is narrower than the Asia range (median 0.17 against 0.42 ATR), so its traversal is not comparable and is not used; the phantom level has no opposite edge. Both are compared on the 30-minute, 60-minute, end-of-day and continuation cells only.
Terciles after the fact. Width and strength terciles are set on the sample; a trader at 09:30 does not know them. The "wide range continues more" cell is one of thirty-one.
Sixty-nine desynchronised days. In the weeks when the US had changed clocks and Europe had not, 14:30 Paris is an hour late relative to New York; those days are kept and a sensitivity without them moves no cell by more than a point. Twenty-one early-close days (20:25) are kept.
Coverage, instrument, strategy. Dukascopy spot XAU/USD, the same frozen store as Papers #1 and #2 (through 3 August 2026); spot is not futures; no volume; gold rose over most of the sample, though the 09:00-to-close session drift averages +0.02 ATR. No strategy was simulated, so there is no expectancy, no R, no win rate. 614 pre-registered and 384 post-hoc comparisons, no multiplicity correction: about fifty "effects" at the 5% level would be expected by chance, and the seven-point first-hour gap, the trend-up continuation cell and the wide-range continuation cell are candidates until shown otherwise. No cell in the tables above has fewer than 76 events; the 2026 cells (92 breaks) and the trend-day argument (85 days) remain fragile.

Methodology appendix

Definitions

Day and range: trading day = Paris calendar day, bars 01:00 ≤ t < 23:00; Asia range AH/AL = high/low of the 01:00–08:55 bars (at least 30 bars); EOD = last bar before 23:00. Days with fewer than 100 bars, no prior-day ATR or no prior-day regime are dropped.
Breaks: W = first bar with high > AH (low < AL); C = first bar with close > AH (close < AL); one per side × window × definition. Morning window 09:00–14:25 inclusive; afternoon window 14:30 ≤ t < 23:00 on the side with no same-definition morning break (primary series). The all-afternoon series, including re-breaks of a side already broken in the morning, is a published sensitivity; after the audit a re-break is the first bar at or after max(14:30, first close back + 1), with a strict variant requiring the re-break bar itself to be a transition.
Failure: a close back on the original side of the broken level on bars after the break bar. E30 = within 6 bars, E60 = within 12, both capped at EOD (a window-capped E30 variant reproduces the original code); E_win = before the end of the window; E_EOD = the day's close on the original side, inside the range or beyond the opposite edge.
Traversal: opposite edge touched by a wick after the break bar, within 24, 48 or 65 bars (2 h, 4 h, 5 h 25) capped at EOD, or by EOD (labelled unequal horizon). Continuation: day close ≥ 0.25 × ATR14(D−1) beyond the broken level. Excursion (MFE): maximum extension beyond the level from the break bar to the bar before the first close back, or to EOD; in ATR and in range widths. Survival: minutes from the break bar to the first close back; censored at EOD when none. Adverse excursion to EOD: maximum wick excursion back through the level after the break, in ATR.
Normalisation and strata: w_atr = (AH − AL) / ATR14(D−1), Wilder, daily, lagged one day; terciles of all valid days (edges 0.357 / 0.496). Break strength = (close − level) / ATR, terciles of morning close breaks (0.011 / 0.028). Hour blocks 09:00–09:55, 10:00–10:55, 11:00–12:55, 13:00–14:25, 14:30–15:25, 15:30–17:25, 17:30–22:55. Prior-day regime from the daily ADX14/DI classification reused from earlier work. Release days from the house event calendar (181 unique dates).
Controls: phantom levels AH + 0.5 × (AH − AL) and AL − 0.5 × (AH − AL), broken 09:00–14:25, scored identically (no opposite edge); London first-hour range = high/low of the 09:00–09:55 bars, broken 10:00–14:25 (923 days with a range, 919 with a break). Trend day: |close(EOD) − open(01:00 bar)| ≥ 1 × ATR14(D−1); first close break = first bar 09:00–23:00 closing outside the Asia range on either side.
Driftless benchmark (post-hoc, order of magnitude): for each morning close break, the probability that a gaussian walk with σ per bar = ATR(D−1) / √264 × 0.55, started at the breaking close and run over the remaining bars, ends at or below the level; averaged over breaks.

Data

XAU/USD spot, Dukascopy, 253,965 five-minute bars from 3 January 2023 to the close of 3 August 2026, the frozen store shared with Papers #1 and #2; 940 calendar days, 925 with at least 100 bars (the 15 dropped are 14 Sunday-evening stubs and the truncated last day), 923 valid: 742 normal, 181 release days; 255 / 259 / 258 / 151 per year. Median Asia width 0.42 ATR. Breaks: morning 801 wick / 724 close; afternoon primary 390 / 388, which is the 2,303 of the subtitle, over all 923 days; the headline cells use the normal-day subsets (630 / 571 and 294 / 295). Placebo breaks: phantom 289 / 236, London first hour 1,247 / 1,118. Seed 42; two consecutive runs are byte-identical.

Statistics

Wilson 95% intervals on every proportion; day-block bootstrap (1,000 resamples of whole trading days, paired draws on both sides of a difference) on every headline difference, since a day can contribute up to two breaks per definition; the bootstrap governs where the two disagree (they agree within half a point on every published cell).
Multiplicity and integrity: 614 pre-registered comparisons (one per published estimate in metrics 1 to 7) and 384 post-hoc comparisons in a separate labelled register; cells under 25 events are flagged indicative and none appears in this paper. The blind replication reproduced all 4,368 event rows of the first run and every headline cell within one point; the first-run results file and the audited one agree on all 4,163 pre-registered leaves. Every number here is reproducible from asia_breakout_study.py v2 and asia_breakout_stats.json; the hour-block traversal column in the table above is a recount from the published event file.

No indicator is infallible. Nothing here is financial advice — every trading decision is your own.

Past frequencies are not future probabilities. This paper describes what XAU/USD did between January 2023 and August 2026 under precisely stated definitions; it does not describe what it will do, and it does not evaluate any trading strategy.

Research like this is what the Pro tier funds.