Delta Terminal Research · Paper #1 · September 2026

Do Fair Value Gaps Really Get Filled?A 55,739-gap test on XAU/USD

We counted every three-candle imbalance on gold over three and a half years and asked whether the gap changes what price does next, or whether a comparable zone with no gap behaves the same way. The fill rate is real. The magnet is not.

What traders believe

A Fair Value Gap is the band left between candle one and candle three of a fast move: their wicks do not overlap, so a range of prices printed only once, inside candle two. The popular reading is that this "imbalance" must be repaired, and two claims travel with it. The gap is a magnet: price comes back to fill it almost always, and does so because the gap is there. The gap is a reactive zone: on the way back, price bounces in the direction of the original move.

Both are testable. Before computing anything we wrote down our expectation: a high raw fill rate (above 70% by end of day on M5), a placebo rate almost as high, a difference under five points, and no information added by the gap once the move behind it is held fixed. On the bounce claim we had no strong prior, and said so.

How we tested it

Pre-registration

Definitions, horizons, controls, headline metrics and statistics were frozen in writing on 25 August 2026, before any result existed: 198 pre-registered comparisons. Everything computed afterwards at the auditors' request sits in a separate block labelled post-hoc (228 comparisons) and is labelled as such here.

The placebo zone

For every real FVG we placed three ghost zones on random other days from the same prior-day regime, at the same time of day, with the same normalised distance, width and side, and measured them identically: 167,207 placebo zones for 55,739 M5 gaps. If price returns to a zone whether or not an imbalance exists there, the placebo says so.

The no-gap control

To separate the gap from the move, we took 94,795 three-candle sequences with the same net direction but no gap, gave them the equivalent target, and matched them to real FVGs by decile of displacement and of return distance, both in prior-day ATR units.

The independent audit

Three reviews attacked the first results: a replication that recounted every headline from the raw event files and re-derived sixteen events bar by bar (zero discrepancies); a code audit with nine tests for look-ahead, horizons, daylight-saving and symmetry; and an adversarial pass that tried to make each effect disappear. The third succeeded twice, and changed this paper's conclusions.

Results

A. The magnet not confirmed

The raw numbers match the folklore: on M5, 90.7% of gaps are touched and 85.3% fully filled before the close of their birth day, 94.0% by the next session, 96.8% within five. The pre-registered comparison against all placebo zones gave 85.3% versus 83.8%: +1.45 points [+0.97; +1.90]. Small, as expected, but clear of zero.

The audit found that the placebo was a mixture. Our framing fixed distance, width and side, but not where the placebo's birth bar sat relative to its zone. A real gap's third candle has just left the zone and never traded inside it; 43.4% of placebo bars had already travelled into theirs, and a quarter had crossed the far edge. One ratio, q, indexes how far the birth bar had already moved toward the zone; a real FVG has q = 1. Figure 1 shows why it matters: fill rates climb monotonically with q, from 75.9% for zones never approached to 92.9% for zones already crossed. The honest comparator, the 21,286 zones with q between 0.8 and 1.25 (a real third candle's geometry, no imbalance beneath), fills 84.7% of the time.

Against that comparator the FVG advantage is +0.54 points [−0.01; +1.10] at end of day, −0.21 [−0.57; +0.19] at one day, −0.21 [−0.50; +0.07] at five. On M15: +0.14 [−1.15; +1.40]. No single year exceeds +0.9. For scale, a price traded within the last M5 bar is revisited before the close 89.6% of the time: gold mean-reverts intraday relentlessly, and the gap adds nothing.

The no-gap control agrees. Pre-registered, gaps filled 2.17 points less often than their matched no-gap twins [−2.59; −1.74]. But a no-gap target is a price the third candle has just traded near, so 56.4% of them are hit on the very next bar against 23.6% of FVGs. Drop everything filled within three bars from both groups and the difference is −0.49 [−1.17; +0.31]; on M15, +0.04. Our pre-registered expectation, a positive difference under five points, held; the corrected figure is about half a point.

B. The reactive zone not confirmed

The bounce claim was measured on the 25,467 M5 gaps that price returned to without filling on the same bar: did price then travel at least one gap-width back in the gap's direction before the fill? Pre-registered result: 85.6% for FVGs, 83.7% for placebo, +1.82 [+1.31; +2.35].

It was an artefact of where the window started: it included the bar that touched the zone, and that bar had usually arrived from more than one width away, so the "excursion" was its own path in. A metric that counts the approach as a rebound will confirm anything. Start the window on the next bar and the rate is 57.3% for FVGs against 56.8% for placebo (+0.46 [−0.20; +1.14]); M15 reads 57.6% versus 58.0%. Against the q-matched subset alone the M5 difference is +1.4 [+0.3; +2.5]: one cell among many, with a negative M15 counterpart and no same-day fill advantage beside it. We read it as noise. Price bounces off a fresh FVG about as often as off a line drawn anywhere nearby.

C. The release-window exception holds

One pre-registered cell stands out, in the opposite direction. Gaps born between 14:25 and 16:00 Paris on the 181 days with a US 08:30 ET release fill 74.7% by end of day against 81.4% for placebo: −6.7 points [−10.7; −3.3], n = 857. Reassigning the "release day" label to 2,000 random date sets, stratified by year, never produced a difference this large (p = 0.0005). The same clock window on ordinary days shows the opposite sign (+2.9, n = 3,197): it is the release, not the hour.

The post-hoc breakdown puts almost all of it in the first fifteen minutes: −16.2 points for the 169 gaps born 14:30–14:45, against −1.2 to −6.8 in neighbouring slices. A gap printed on the release candle is the footprint of an impulse that tends to keep going into the close; the return is deferred, not cancelled (91.9% filled at five days versus 94.1%). It is uneven across years (−9.5 in 2023, −1.5 in 2024, −5.8 in 2025): robust in aggregate, fragile in any season. M15 agrees in direction (57.9% versus 66.9%, n = 252, p = 0.007).

D. Fill is a speed effect, not a level effect holds

Follow the fill curve over time (Figure 2). Against the pooled placebo the FVG group is behind at five minutes (−1.7, because pre-penetrated placebos fill instantly), leads by +2.6 at one hour, then gives it back: +1.5 at the close, +0.2 the next day. Against the comparable placebo the lead is +1.8 from the first bar, shrinks steadily to +0.5 at the close, and is −0.2 the next day.

Either way the shape is the same: a fresh FVG is filled somewhat sooner than a comparable zone, by one to three points in the first hours, and no more often once the day is done. The five-day comparison (+0.08 [−0.15; +0.29]) is not a ceiling effect: it stays within a fraction of a point in strata well below 100%, such as the Overlap block (94.4% versus 94.4%).

An exploratory lead not pre-registered, to be confirmed

The audit surfaced one pattern we did not plan to look for and do not yet treat as a result. Rank each FVG by the violence of the move that created it (three-candle range over prior-day ATR) and the fill gap versus comparable placebo is not flat: +3.9 points in the softest decile, between +1 and +5 through the middle, −1.3 in the ninth, −11.5 [−14.1; −8.6] in the most violent. By tercile: +3.3, +1.7, −2.3. M15 shows the same shape, its top decile at −17.8.

Two cautions. This is a post-hoc cut, one of 228, and the top decile is where the matched sample is thinnest. It is also consistent with the release-window result and with what the pre-registered width split was really showing: the widest gaps differed from placebo because the move behind them was violent, not because they were wide. If it holds, the informative variable is the displacement, not the gap. It gets its own pre-registered study before we say more.

How to use this

Nothing here is a signal; no entry, stop or target was simulated. What the data gives you is a calibrated expectation: an FVG marks a fast move after which gold usually comes back, at a rate that depends far more on timeframe and time of day than on the gap.

End of day is the last bar before 23:00 Paris; a gap born at 20:00 has less day left, which is most of what the NY-late row says. The 162 M5 gaps born after 23:00 are left out for that reason. Placebo rates by block move in step.

• Expect the fill; do not trade for it. An M5 gap will most likely be filled today and almost certainly within a week. So will a zone with no gap.
• Adjust by session. Nine in ten Asia and London M5 gaps close the same day; seven in ten NY-late gaps do.
• Distrust the release candle. A gap printed in the quarter hour after a US data release is the one case where the same-day expectation should be marked down.
• Do not expect a bounce. A one-width reaction on the first return happens 57% of the time, as it does for placebo zones.
• Do not read width as strength. The narrowest third of M5 gaps fill 89.9% of the time, the widest third 79.4%; placebo tracks both.

Limits

• Tiny gaps. Three in ten M5 gaps are narrower than 0.66% of prior-day ATR, roughly a quarter-dollar at the sample's median ATR: spread territory on spot. Their 99.5% "respect" rate is a floor effect and was not used.
• Placebo calendar. A placebo zone carries the year and news status of the day it was placed on, not of its source gap.
• Coverage. 3 January 2023 to 3 August 2026: 940 trading days, 925 kept; 277 M5 gaps censored at the five-day horizon only. Gold rose for most of the period; bull- and bear-side fill rates are close (85.1%, 85.5%), but a different regime is a different sample.
• Instrument. Dukascopy spot M5, resampled to M15. Spot is not futures; volume is tick count and was not used.
• No strategy. No entries, stops, targets or costs. This tests a geometric claim, not a trade.
• Multiplicity. 198 pre-registered comparisons, 228 post-hoc. The release-window effect is reported because it survives a permutation test, not because one cell cleared a threshold.
• Overlap and clock. 68.1% of M5 triplets share a bar with another; 17.1% of zones intersect an earlier open zone; each gap is followed independently and the bootstrap resamples whole days. Blocks are Paris time; "Open" is the data store's opening hour; end of day does not shift with US daylight saving.

Methodology appendix

Research like this is what the Pro tier funds.

Delta Terminal Research · Paper #1 · Data through 3 August 2026 · Framing frozen 25 August 2026 · Audit closed 6 September 2026