Delta Terminal Research · Paper #2 · September 2026

Do Order Blocks Hold?A 14,187-block test on XAU/USD

The order block is the centrepiece of Smart Money Concepts: the last opposite candle before a move that breaks structure, read as the footprint of institutional orders that price will come back to and respect. We detected every such block on gold over three and a half years, built two controls and let three auditors attack the first numbers. Price does come back. The candle leaves no measurable trace, and the block does not hold.

Key findings

Yes, price returns. 86.1% of M5 order blocks on XAU/USD are revisited ("mitigated") before the end of their birth day, 94.4% by the next day, 97.1% within five; median 20 trading minutes. M15: 78.9% / 90.0% / 94.8%, median 45 minutes.
But not because it is an order block. Against placebo zones with the same geometry and recent path, each block set against its own twins, the same-day difference is −0.70 points [−1.31; −0.02]; at equal local volatility, +0.09 [−0.73; +0.95]; five days later, +0.21.
The opposite candle leaves no trace. Keep the structure-breaking move but let the last candle be the same colour as the move, and the zone behaves like an order block: −0.30 points [−0.91; +0.30] same day, +0.07 at five days.
The block does not hold. Price traverses the block, the doctrine's "invalidation", in 74.4% of cases before the close and 94.3% within five days. Still unbroken at the close: 13.6% of revisited blocks; at five days, 2.8%, the same as placebo (2.8%).
One law for FVGs and order blocks. Binned by distance from price at birth, order blocks, placebos, same-colour zones and Paper #1's Fair Value Gaps fall on one curve: about 98% revisited within 0.005 ATR, 90% at 0.02–0.04 ATR, 60% beyond 0.15 ATR. Distance and the clock set the odds, not the pattern.

What traders believe

In the SMC and ICT vocabulary a bullish order block is the last bearish candle before an impulsive rally that breaks a prior high. That candle is where institutions accumulated, the story goes, and unfilled orders remain in its range: price is expected to return to it ("mitigation") and then continue in the direction of the original move; a block that price cuts straight through is "invalidated".

Three testable claims. A: price returns to the block almost always, and because the block is there, not merely because gold mean-reverts intraday toward anything it just left. B: on the first return price bounces and the block holds. C, the one that makes an order block an order block rather than any recent range: the opposite colour of the last candle carries information a same-coloured candle would not.

We wrote our expectations down before computing anything. A: a high raw return rate (70–85% by end of day on M5), a placebo almost as high, a difference under five points. C: no difference, within about two points. B: no strong prior; a block is a real traded range that could plausibly hold, but Paper #1 had just shown a "bounce" metric on FVGs to be noise.

How we tested it

A mechanical definition

A structure break is a bar closing above the highest high of the ten bars before the candidate candle. A bullish order block is the last bearish candle c0 followed by a structure break within five bars, without price trading back below c0's low in between. The zone is c0's full range; the block is born at the close of the breaking bar, the first moment it exists without look-ahead. Bearish blocks are the mirror image.

Two controls

The trajectory-matched placebo tests claim A. For each block we placed up to three ghost zones on other days with the same prior-day regime, time of day, normalised distance, height and side, and, crucially, the same recent path: the placebo day's price had to have occupied the ghost zone the same number of bars earlier, to a comparable depth. A zone price has just left in the same way, with no opposite candle and no identified break. 35,317 such zones; 13,152 of 14,187 blocks found at least one twin.

The same-colour control tests claim C. Same rules, except that the last candle before the move is the same colour as the move. 47,227 such zones on M5, compared to real blocks within bins of displacement by distance.

Pre-registration and three audits

Definitions, horizons, controls and headline metrics were frozen in writing on 7 September 2026 before the first result; the pre-registered run produced 394 comparisons. Three independent reviews then attacked it: a replication that recounted 121 lines from the raw event files and re-derived twenty zones bar by bar against the price data (zero discrepancies); a code audit for look-ahead, horizons and symmetry, which cleared the detection and found one blocking flaw in how we aggregated the placebo comparison; and an adversarial pass that tried to make every effect disappear. It mostly succeeded. The corrections add 402 paired and 256 post-hoc comparisons, kept in separate labelled blocks.

Results

A. The return specific return not confirmed

86.1% of M5 blocks are revisited before the end of day, 94.4% by the next session, 97.1% within five. Half of all returns happen within 20 trading minutes, a quarter within 5, three quarters within 80. On M15: 78.9% / 90.0% / 94.8%, median 45 minutes.

Part of that is mechanical: in 68.1% of blocks the breaking bar's own low already reaches into the zone at birth, and 40.0% of those are "mitigated" on the very next bar. Clean-start blocks are revisited 78.1% of the time by the close, the rest 89.8%.

The pre-registered comparison against the pooled placebo gave 86.1% versus 88.7%: −2.59 points [−3.17; −2.03]. Our expectation had been a gap under five points and it held, but the sign was a surprise: at face value an order block returned less often than a random zone. That number is superseded, for the reasons in the box below. The paired estimator, each block against the mean of its own twins, one block one weight, gives −0.70 points [−1.31; −0.02] at end of day, +0.36 [−0.16; +0.88] the next day and +0.21 [−0.18; +0.60] at five days; on M15, −0.05 [−1.33; +1.35] the same day. Twins with a comparable local volatility before birth: +0.09 [−0.73; +0.95]. Twins whose own birth bar also cleared its prior ten-bar extreme: −0.06 [−0.96; +0.88]. Alternative definitions (twenty-bar break, body-only zone, three-bar cap): −0.20, −0.60, −0.47.

Read strictly, the primary estimator says a block is revisited very slightly less often on its birth day than a comparable zone, seven tenths of a point with an interval that just clears zero, and no differently from the next day on; match the local volatility and even that residual is gone. Gold intraday goes back to almost anything it recently left; the order block is one such thing.

B. The hold not confirmed

This is the practical part. If the block is a wall, price should come back and stop. By the end of the birth day 74.4% of M5 blocks have been traversed, price having gone through the far edge; 89.3% by the next day, 94.3% within five. The median traversal comes 70 trading minutes after birth. On M15, 60.8% the same day.

Among blocks that price did revisit, the share still unbroken at the close is 13.6%; 5.4% at the next close; 2.8% after five days. Placebo zones read 12.5% / 5.1% / 2.8%. The pre-registered same-day gap (+1.07 [+0.31; +1.78]) shrinks to +0.59 [−0.26; +1.33] when paired and to −0.49 when volatility-matched; at five days the paired gap is −0.18 [−0.56; +0.23]. M15 same-day paired: +2.25 [+0.46; +3.92], becoming +1.38 [−0.88; +3.66] at equal volatility and +0.50 [−0.44; +1.47] at five days; we do not read "holds better on M15" out of that.

Stacked bars for order blocks and placebo zones at end of day, next day and five days: traversed 75.5 to 94.9 percent, held 11.6 percent falling to 2.6 percent, not yet revisited 12.9 percent falling to 2.5 percent; placebo bars are the same within one point. — F4. What has happened to a block by each horizon, matched blocks against their own twins. Red is the doctrine's "invalidation"; the teal slice, revisited and held, is 11.6% of matched blocks at the close and 2.6% five days later. Placebo zones show the same slices.

The bounce metric agrees. Among blocks revisited without being traversed on the touching bar, price travelled at least one block-height back in the direction of the move, counting from the bar after the touch, in 54.2% of cases; placebo zones, 55.4% (paired −1.05 [−2.20; +0.07]). We had no strong prior on B and the data supplies none in the block's favour: it holds to the close about one revisited block in seven, to the end of the week about one in thirty-five, like a zone drawn at the same distance on another day.

C. The opposite candle, the SMC core refuted

Everything above compares a block to a zone with no structure break. Claim C is narrower and more important: holding the move fixed, does the colour of the last candle matter? Standardised to the block distribution of displacement and distance, same-colour zones are revisited 86.4% of the time by the close against 86.1% for blocks: −0.30 points [−0.91; +0.30]. Excluding everything revisited within three bars, where mechanics dominate: −0.66 [−1.69; +0.48]. At five days: +0.07 [−0.23; +0.36]. Traversal by the close: 74.4% versus 74.0%. The one-height bounce: 54.2% versus 54.2%. M15: −0.12 [−1.31; +1.08].

The control reproduces the block's quirks too: clean-start blocks 78.1%, clean-start same-colour zones 78.6%; breaking bar already in the zone, 89.8% against 90.8%. The one cell that looked like a block-specific weakness, moves that took the full five bars to break structure (−6.7 points against paired placebo, n = 388), is shared: 70.9% for those blocks, 71.5% for five-bar same-colour zones. Our pre-registered expectation was about zero. It held. Whatever an order block does, a range whose last candle went the "wrong" way does too.

D. One law for FVGs and order blocks holds

Paper #1 found that a Fair Value Gap is filled at the rate a comparable zone would be. Figure 2 puts the two studies on one axis: same-day revisit rate by distance from the closing price at birth to the zone's near edge, in prior-day ATR. Four families, one curve: within 0.005 ATR, 98.0 / 98.2 / 97.9 / 97.7% for blocks, placebos, same-colour zones and FVGs; at 0.02–0.04 ATR, 90.1 / 90.0 / 90.1 / 90.5%; at 0.08–0.15 ATR, 75.0 / 78.3 / 76.9 / 78.3%; beyond 0.15 ATR, 60.6 / 57.4 / 61.5 / 63.3%.

Line chart of same-day revisit rate against seven distance bins for order blocks, placebo zones, same-colour zones and Fair Value Gaps; all four fall from about 98 percent to about 60 percent along the same curve. — F2. Same-day revisit rate by distance from price at birth, in prior-day ATR, for the four families. The order-block curve sits one to three points under the others in the middle bins, the residual Section A traces to local volatility; the shape is shared.

The clock does the rest. Blocks born in Asia are revisited 93.4% of the time by the close; London, 91.9%; the overlap, 80.3%; late New York, 71.1%. Placebo zones and FVGs fall off across the sessions the same way; the paired gap is +0.05 in Asia, −0.04 in London, +0.34 in the overlap. Late New York reads −3.25 [−4.88; −1.49] on the day and +0.36 the next day: a block born with little day left, not a property of the block.

What the audit did to our own numbers

The pre-registered headline was −2.59 points: order blocks revisited less than placebo, interval well clear of zero, and we were preparing to publish it as a mild negative property of the pattern. The corrected reading runs −2.59 pooled, −0.70 paired, +0.09 volatility-matched, +0.21 at five days.

Why the pooled estimator was biased. The trajectory match is demanding, so a block ends up with zero to three accepted twins depending on its geometry: blocks whose zone sits close to price found three easily, blocks whose zone sits far from it often found none (7.3% of blocks, revisited only 72.9% of the time). The pooled comparison counted every placebo once, so on the placebo side easy blocks weighed three times and hard blocks nothing. Decomposed: −2.59 = −1.03 from dropping unmatched blocks + −0.70 genuine paired difference + −0.85 from the weighting. The pooled number was correct as computed; it answered a question we had not meant to ask.
Why local volatility explains the residual. A structure break needs a compressed range to break out of: the average true range of the twelve bars before birth has a median of 0.043 ATR for blocks against 0.048 for their placebos, and a quieter tape appears to revisit a nearby level a little more slowly in the first hour. Keep only twins whose pre-birth range is within 0.8–1.25 of the block's and the same-day gap is +0.09, flat from the first hour (Figure 3). Placebos that themselves sit at a structure break give −0.06.
What we checked so as not to fool ourselves twice. With 200 draws instead of 40, 98.5% of blocks are matched and the gap is −0.88 [−1.44; −0.28] on the day, +0.05 at five days: a slight early delay, nothing at the horizon. The pooled run's "2026 deterioration" (−4.6) was the same artefact; by birth year the paired gap is +0.18 / −0.40 / −0.86 / −2.67 for 2023 to 2026 on M5 and flat on M15. We label the 2026 cell unstable, not a trend.

The pre-registered block of the results file is untouched, verified leaf by leaf against the audited script; the paired and post-hoc blocks sit beside it. Both are published because the correction is the finding.

Line chart of the order-block-minus-placebo revisit gap over time under four estimators: pooled starts at minus 6 points and ends at minus 0.7 next day; paired starts at minus 3 and ends at plus 0.4; volatility-matched starts at minus 1.2 and is zero from one hour; structure-break placebos are flat at zero. — F3. The revisit gap, block minus placebo, from five minutes to the next day under four estimators. The pooled curve is the superseded pre-registered reading; the volatility-matched and structure-break curves sit on zero from the first hour.

Five bars with bootstrap whiskers: minus 2.59 pooled, minus 1.55 pooled on matched blocks only, minus 0.70 paired, plus 0.09 volatility-matched, plus 0.21 paired at five days. — F1. One comparison, five estimators, same data. The first two are the pre-registered pooled estimator and the same estimator on matched blocks only; the last three are the audited readings.

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.

Block born in (Paris time)	M5 revisited by EOD	M5 traversed by EOD	M5 blocks	M15 revisited by EOD	M15 blocks
Open 00:00–01:00	93.0%	86.6%	574	84.3%	191
Asia 01:00–09:00	93.4%	86.8%	5,146	88.7%	1,788
London 09:00–14:30	91.9%	84.3%	3,402	86.7%	1,119
Overlap 14:30–17:30	80.3%	61.3%	1,876	71.3%	773
NY late 17:30–23:00	71.1%	49.9%	3,156	53.1%	793
All blocks	86.1%	74.4%	14,187	78.9%	4,673

End of day is the last bar before 23:00 Paris, so a late block has less day left. The 33 M5 blocks born after 23:00 are omitted. Placebo rates by session move in step.

Expect the revisit; do not credit the block for it. An M5 block will most likely be revisited today and almost certainly within a week. So will a zone with no block. Read the distance instead: within 0.02 ATR of price, nine times in ten; beyond 0.15 ATR, six times in ten, block or not.
Expect the traversal. Three blocks in four are cut through before the close, nineteen in twenty within five days. Any reading that needs the block to stay intact is contradicted three times in four by the close, and the minority that holds is not distinguishable from placebo.
Do not read the candle. Its colour changed nothing we could measure; a block "refined" on that basis is a range like any other.
Adjust by session, not by pattern. Nine Asia and London blocks in ten are revisited the same day; seven late-New-York blocks in ten. That spread dwarfs anything attributable to the block itself.

Limits

One definition. Ten-bar break, at most five bars from candle to break, full candle range. Sensitivities agree, but "order block" has as many definitions as teachers and we tested one mechanical version.
Raised levels. In 19.2% of blocks the opposite candle's own high is a fresh ten-bar high, so the zone sits almost on price (median distance 0.015 ATR, 93.5% revisited). Excluding them: −0.94 [−1.62; −0.14] paired on the day, +0.22 at five days.
What the placebo matches. Geometry, clock, regime and recent path; not local volatility, which is why that appears as a post-hoc correction. It tests the block and its structure break together; only the same-colour control isolates the candle. 7.3% of blocks found no twin within forty draws. A placebo carries the year and news status of the day it lives on, not of its source block.
A Monday cell. Blocks born on Mondays show a paired gap of −2.61 [−4.12; −1.13], one of five weekday cells, post-hoc and unexplained; reported, not used.
ATR convention. The prior-day ATR uses full Paris calendar days, not the 01:00–23:00 house day named in the framing document; immaterial for a normaliser, but the framing was imprecise.
Coverage and instrument. Dukascopy spot XAU/USD, the same frozen store as Paper #1 (through 3 August 2026), so the two studies are directly comparable; spot is not futures; no volume was used. 940 days, 925 kept; 76 M5 blocks could not be followed to five days. Gold rose over most of the sample.
No strategy, many comparisons. No R multiples, no expectancy. 394 pre-registered, 402 paired and 256 post-hoc comparisons; any single interesting cell should be read against those counts.

Methodology appendix

Definitions

Structure break (BOS): bar b with close(b) > max High of the ten bars preceding c0 (bullish); close(b) < min Low (bearish). N = 20 as a counted sensitivity.
Bullish order block: last bearish candle c0 such that a BOS occurs within c1…c5 and price never trades below low(c0) between c0 and b. Zone = [low(c0), high(c0)]; body-only as a sensitivity. Born at close(b), tracked from b+1. Bearish symmetric.
Geometry: height h = high(c0) − low(c0); distance d = close(b) − high(c0); displacement = high(b) − low(c0); all in units of the previous day's 14-day Wilder ATR. L = bars from c0 to b (1–5).
Revisit (mitigation): first bar whose extreme enters the zone. Traversal (invalidation): first bar reaching the far edge. The 50% "mean threshold" variant was counted (M5 EOD 80.1%, paired −0.86).
Held: revisited and not traversed by the horizon; denominator = revisited by the horizon. Bounce: among blocks revisited without traversal on the touch bar, an excursion of at least one height in the move's direction from the touched edge, window from the bar after the touch to the bar before traversal (0.5 and 2 heights counted: 66.0% vs 66.1%, 38.3% vs 40.0%).
Horizons: end of day (last bar before 23:00 Paris), then one, three and five trading days; censored events excluded from that horizon and counted.
Trajectory-matched placebo: up to three zones per block on other days of the same prior-day regime and time of day (source day excluded), same normalised distance, height and side, and a path ratio q′ within [0.8 q, 1.25 q], where q = (close(b) − low(c0)) / d for the block and the analogous quantity over the previous L bars for the placebo. Up to forty draws per block; 322,412 of 361,983 draws rejected on path.
Same-colour control: identical rules with a same-coloured last candle; rates computed within 84 displacement-by-distance decile bins and reweighted to the block distribution.
Local volatility (post-hoc): mean true range of the twelve bars before the c0…b window, in prior-day ATR; volatility-matched twins have a ratio to their block within [0.8, 1.25] (14,136 placebos; 9,026 blocks). Structure-break placebos: the 32.9% whose birth bar closed beyond the extreme of the ten bars before their own window (8,317 blocks).

Data

XAU/USD spot, Dukascopy, 253,965 five-minute bars from 3 January 2023 to the close of 3 August 2026, resampled to 84,660 fifteen-minute bars; 940 trading days, 925 kept (days under 100 M5 bars dropped). 14,187 M5 blocks (7,427 bullish, 6,760 bearish); 4,673 M15; 35,317 M5 placebos (11,428 M15); 47,227 M5 same-colour zones (14,809 M15). Prior-day regime from a daily ADX14/DI classification reused from earlier work. 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) on every headline difference, so that the dozens of blocks sharing a day (38.7% of zones overlap an open one) are not treated as independent.
Paired estimator: for each block with at least one twin, its outcome minus the mean outcome of its twins, averaged over blocks so that one block has one weight; intervals by day bootstrap. It replaces the pooled estimator as the primary reading. An independent recount from the event files reproduced −0.700 and 28 further cells to the sixth decimal.
Multiplicity and integrity: 394 pre-registered, 402 paired and 256 post-hoc comparisons in separate registers; cells under 25 events are flagged indicative and not used. The pre-registered JSON block and CSV columns were verified unchanged between the original and the audited script. Every number here is reproducible from ob_study.py v2 and ob_stats.json.

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.