Delta Terminal Research · Paper #4 · November 2026
Is Gold Rising, or Is the Dollar Falling?A 3.5-year attribution test on XAU/USD — and what the answer does not change
Every move in XAU/USD can be split, by an accounting identity, into a part that is gold and a part that is the dollar. Our free TradingView indicator, the Gold Strength Basket, does exactly that: it prices gold in six currencies other than the dollar, calls the average of those returns the gold part, and calls the remainder the dollar part. The natural next question is whether the split predicts anything. Does a morning rise that is "broad gold" hold into the close more often than a rise that is "mostly the dollar falling"? We wrote our expectation down, classified 544 normal mornings over three and a half years, controlled for the size of the move, ran four placebos and let three auditors attack the first run. The label does not predict what follows. The size of the move does.
Key findings
- Raw, the dollar-driven mornings do come back twice as often. On 544 normal mornings (01:00–14:25 Paris, 2023–2026), a move classed "mostly dollar" had returned to its starting price before the close 58.5% of the time [43.4; 72.2], against 27.0% [21.9; 32.9] for a "broad gold" move: a gap of −31.5 [−46.9; −15.9] points.
- Because they are half the size. "Mostly dollar" mornings are rare (7.5%, 41 of 544, 6 to 12 a year) and small: 0.22 prior-day ATR on average against 0.50 for broad gold, and none above 0.76 ATR. The pooled amplitude curve alone predicts a return rate of 51.2% for moves of that size and 26.7% for moves of broad-gold size. Adjusted for size, regime, direction and year, the gap is −6.7 [−22.5; +6.5] points.
- At equal size, a dollar of "dollar" and a dollar of "gold" hold exactly as well. The cleanest test puts the gold part and the dollar part of each move, in ATR, side by side in one model: the difference in their effect on still being on the move side at the close is −0.06 log-odds per ATR [−1.73; +1.81], and on having returned to origin +0.18 [−1.41; +1.44]. Shifting a tenth of an ATR from the gold part to the dollar part at constant size changes the probabilities by −0.1 and +0.3 points, intervals of about ±2.5. The same holds in the Asia window, the London window, on US-release days and the next morning.
- The dollar share is a by-product of size. The dollar part of a morning is about 0.09 ATR whatever happens, which is also its level on mornings with no move at all (0.088 ATR): it is a noise floor. Divide a roughly fixed number by the size of the move and you get the dollar share, hence the correlation of −0.48 between the two. "Mostly dollar" and "fight" are, for the most part, mornings on which gold did not move: 71% and 74% of them are under 0.25 ATR.
- Nothing downstream either. The afternoon does not give the dollar part back (about 8% of it, interval covering zero), and on US-release days the composition of the pre-release move does not predict the reaction: 0 favourable cells out of 288, every class near 50%. What the indicator does show, correctly, is where a move comes from: two thirds of the first hour after a US release is dollar (median dollar share 0.70 against 0.31 on a normal morning).
What the indicator shows, and what traders read into it
Gold is quoted in dollars, so a rise in XAU/USD can be gold getting dearer everywhere or the dollar getting cheaper against everything. The Gold Strength Basket separates the two. For a window of time it computes the return of gold priced in EUR, GBP, JPY, CHF, AUD and CAD, averages the six with equal weights, and calls that the gold part; the dollar-priced return minus the gold part is the dollar part, which by construction equals the average move of the basket against the dollar. The identity is exact. With a 25% threshold, every move with a magnitude of at least 0.10% falls into one of four classes: broad gold (the dollar part is at most a quarter of the move, and gold and dollar agree in sign), mostly dollar (the gold part is at most a quarter), tilt (both pull the same way, neither dominates) and fight (gold and dollar pull opposite ways).
The indicator measures; it does not forecast. But a reading invites a belief, and it is a belief many gold traders hold: a rise that is "real gold", bid in every currency, carries conviction and should persist, whereas a rise that is only the dollar sagging is borrowed and should fade when the dollar bounces. Phrased as hypotheses before any computation: H1, on normal days a morning move classed broad gold holds into the close better than a move of the same size classed mostly dollar; H2, on US-release days the reaction to the number continues a broad-gold pre-release move more often than a mostly-dollar one; H3, both are symmetric for rises and falls. The shared null: composition adds nothing beyond the size of the move and the prior-day regime.
We wrote our bet down first. H1: weak or nil once size is controlled, a return-rate gap between −5 and +5 points with an interval covering zero, and if anything "mostly dollar comes back a little more", without conviction. H2: plausible but probably limited by sample size. H3: symmetric. The trap we expected: the classes differ in size, so any raw gap has to survive matching or regression on amplitude. That trap turned out to be the whole result.
How we tested it
Mechanical definitions
A trading day runs 01:00 to 23:00 Paris on five-minute bars. The morning move is the change from the close of the last bar before 01:00 to the close of the 14:25 bar; the same anchors give an Asia window (to 08:55), a London window (09:00 to 17:25) and, on release days, a pre-release window from 09:00 to the last bar before the release. Gold in each currency is XAU/USD divided by EUR/USD, GBP/USD, AUD/USD or multiplied by USD/JPY, USD/CHF, USD/CAD, at the same two instants; the gold part is the equal-weight mean of the six returns, the dollar part the remainder, the dollar share the dollar part divided by the absolute move. Three things are then recorded at each horizon after the window (+1 h, +2 h, 17:25, the close, the next morning 08:55): the state, whether the close is still on the move's side of the starting price; the return, whether at least one five-minute close has crossed back through the starting price; and the follow-through, the further move in units of the original one, with the best and worst excursion in ATR. The size of the move is measured in the previous day's 14-day ATR, and is the control variable throughout. For H2 the reaction is the signed move from the last pre-release close to one hour after the release and to the close, and "continuation" means it has the same sign as the pre-release move.
Controls defined before the effect
Size and regime. Three pre-registered ways of holding size equal: stratification by amplitude tercile (Mantel–Haenszel weights, with equal weights published beside), a logit regression of each outcome on class, amplitude, prior-day regime, direction and year, reported as the average marginal effect, and 1:k matching on amplitude within 0.10 ATR and regime. Four placebos. P1 uses the previous valid day's FX for the same window with today's gold, so the label carries no information about today's currencies; P2 permutes the class labels across days 1,000 times; P3 flips the sign of each currency's return at random; P4 recomputes everything with an "exotic" basket of NOK, MXN and ZAR, descriptive only since those pairs contain the dollar too. Robustness. Basket weights (equal, DXY-like, median), log returns, thresholds of 15% and 35%, a 0.25 ATR gate, year by year, FX tolerance, early-close days, and the US-release days as a separate population. The headline is restricted to normal days, with the 180 US-release days (NFP, CPI, PPI, retail sales, FOMC) analysed separately, because large dollar moves cluster on them.
Pre-registration and three audits
Hypotheses, the written bet, definitions, windows, horizons, controls and the ten headline metrics were frozen on 23 October 2026 before the first result; the run produced 3,723 pre-registered comparisons under the rule "one published estimate, one comparison". Three independent reviews then attacked it. A blind replication recomputed every window from the raw bars, wrote its numbers down before opening our results file, and matched 2,514 common windows row for row on class and every outcome. A code audit cleared look-ahead and the identity, and found one blocking flaw: in the random-sign placebo, the move was read from an unsorted table, so 99.97% of the placebo rows were compared with another day's move. It was fixed, and the sentence we had drawn from it is withdrawn below. An adversarial pass then tried nine ways to break the emerging conclusion, and nine ways to rescue the hypothesis; the nine attacks were replayed inside the script as a labelled post-hoc block of 224 comparisons. No pre-registered cell was altered except where an audit found a bug.
Results
A. The raw gap real, and made of size
Of the 544 normal mornings that cleared the 0.10% gate, 244 were broad gold (44.9%), 41 mostly dollar (7.5%), 236 tilt (43.4%) and 23 fight (4.2%). At the close, a broad-gold move was still on its own side 86.9% of the time [82.1; 90.6] and a mostly-dollar move 78.0% [63.3; 88.0]; at least one five-minute close had come back through the starting price before 23:00 for 27.0% of broad-gold moves and 58.5% of mostly-dollar ones. The raw gaps, broad gold minus mostly dollar, are +8.8 [−4.8; +23.1] points on the state and −31.5 [−46.9; −15.9] on the return (day-block bootstrap). Taken alone, that is the belief, confirmed and significant: a permutation of the labels gives p < 0.01 on the return gap.
It is also a comparison between moves of 0.50 ATR and moves of 0.22 ATR. Half of the mostly-dollar mornings are under 0.20 ATR; half of the broad-gold ones are over 0.43. The median favourable and adverse excursions after the window are the same in both classes (0.27 / 0.23 ATR for broad gold, 0.29 / 0.24 for mostly dollar): the afternoon travels the same distance either way, and a quarter of an ATR is simply enough to undo a 0.22 ATR move and not a 0.50 ATR one.
B. Size, not label confirmed four ways
Pool the four classes and sort the 544 mornings by size instead. The return-to-origin rate falls from 66.7% [57.3; 74.8] for moves of at most 0.15 ATR to 25.0% between 0.40 and 0.55 ATR and 5.7% [2.2; 13.8] between 0.55 and 0.75; above three quarters of an ATR, 2 of 58 mornings came back. The state at the close climbs the same curve, from 67.6% to 100%. A logit on log-amplitude fitted to all classes together is the "amplitude curve"; read at each class's mean size it gives the expected rate for that class if composition carried no information. Broad gold observed 27.0%, expected 26.7%; mostly dollar observed 58.5%, expected 51.2%, a residual of +7.3 points [−7.2; +22.2]; tilt −2.0. On the state the residuals are +0.3, +3.6 [−10.2; +16.5] and +0.8. Estimating the curve without the class being tested changes the expected values by about a point.
The three pre-registered size controls say the same thing in three dialects. Stratified by amplitude tercile with Mantel–Haenszel weights, the return gap is −15.6 [−33.3; +1.4] and the state gap +2.0 [−13.5; +17.5]; with equal stratum weights, −17.6 [−40.9; −0.5] and +7.9 [−9.6; +31.6], both dragged by a top tercile that contains 4 mostly-dollar mornings against 121 broad-gold ones. In the only tercile where the two classes have comparable numbers (the small one, 40 against 25), the gaps are −1.0 on the return and −11.5 on the state: the mostly-dollar mornings hold slightly better. The regression-adjusted average marginal effect is −6.7 [−22.5; +6.5] on the return and −2.5 [−13.4; +9.7] on the state. Matching gives −16.7 [−38.5; +4.8] and +12.6 [−10.5; +28.3], but with 41 controls serving 168 treated mornings the effective number of controls is 17.2, and 76 broad-gold mornings above 0.40 ATR have no match at all; the matched estimate is published as indicative and is not the estimate of record.
The cleanest test came from the adversarial audit. Instead of four labels, take the two continuous ingredients of every morning, the gold part and the dollar part, each in ATR and signed in the direction of the move, so that they add up to the size of the move. Regress the outcome on both, plus regime, direction and year. If composition mattered, an ATR of dollar would carry a different coefficient from an ATR of gold. On the state at the close the coefficients are +4.18 per ATR of gold and +4.12 per ATR of dollar, a difference of −0.06 [−1.73; +1.81]; on the return, −5.45 against −5.28, a difference of +0.18 [−1.41; +1.44]. In probability terms, moving a tenth of an ATR from the gold part to the dollar part at constant size changes the chance of still being on the move side by −0.1 points [−2.4; +2.2] and the chance of a return by +0.3 [−2.5; +2.6]. The next morning at 08:55: −0.80 [−2.35; +0.76] and +0.33 [−1.16; +1.50]. The Asia window: +0.35 and −0.07. The London window: +0.50 and −0.21. Mornings on US-release days: −0.83 and −1.93, wider intervals, all covering zero. The hypothesis is not under-powered; it is refuted to a precision of about two and a half points on the headline population.
C. Why "mostly dollar" means "small" mechanical
The dollar share is the dollar part divided by the size of the move, and the dollar part barely varies. Across the 544 classified mornings its median magnitude is 0.089 ATR; across the 97 mornings that did not clear the 0.10% gate, where by definition nothing happened, it is 0.088 ATR. The six currencies never move in lockstep, so the basket's average return against the dollar always differs from gold's by roughly a tenth of an ATR of noise. Sort the mornings into amplitude deciles: the median dollar part stays between 0.07 and 0.14 ATR from the smallest decile to the largest while the median gold part rises from 0.08 to 0.85, and the median dollar share falls from 0.82 to 0.14. The Spearman correlation between the size of a move and its absolute dollar share is −0.48. The share of mostly-dollar mornings by amplitude decile runs 16, 9, 13, 16, 11% in the lower half and 2, 4, 2, 2, 0% in the upper half; the share of broad gold rises from 9% to 76%.
"Fight" is the same phenomenon one step further. Its 23 mornings have a median size of 0.16 ATR, 74% of them under a quarter of an ATR, and in every one of them at most three of the six currencies moved with gold. The typical disagreement between the six currency returns inside the basket is 0.18% (median standard deviation) against a median move of 0.49%: on a move of 0.16 ATR, ordinary dispersion between currencies is enough to manufacture a "conflict". On the state at the close, fight sits off the amplitude curve by −17.3 points, with an interval of [−38.5; +3.9]; we report it as what it is, small moves on which the basket disagrees, and claim nothing further. Finally the identity itself behaves: the gold and dollar parts are almost uncorrelated across mornings (+0.07 Pearson), so the split does not create an artefact, and the dollar part accounts for about 15% of the variance of the morning move.
D. Does the afternoon give the dollar part back? no
A second version of the belief is directional: whatever the dollar lent in the morning, it takes back in the afternoon. Decompose the move from 14:25 to the close into its own gold and dollar parts, signed in the direction of the morning move, for the 537 mornings with FX at both instants. After a mostly-dollar morning whose dollar part averaged +0.21 ATR, the afternoon dollar part averages −0.025 ATR [−0.077; +0.027]. Regressing the afternoon dollar part on the morning's two parts across all classes, the morning dollar part comes back at a rate of −0.077 [−0.170; +0.016] per unit: about 8% of itself, with an interval that covers zero. The afternoon gold part, meanwhile, continues at +0.19 per unit of morning gold part and +0.20 per unit of morning dollar part: whatever the origin of the morning move, the afternoon adds the same small gold-side drift. And the afternoon is made of gold: the standard deviation of its gold part is 0.39 ATR against 0.17 for its dollar part. The afternoon does not undo the dollar; it mostly ignores it.
| Morning class | n | Morning gold / dollar part (ATR) | Afternoon gold part | Afternoon dollar part |
|---|---|---|---|---|
| Broad gold | 242 | +0.47 / +0.03 | +0.030 [−0.023; +0.081] | −0.001 [−0.021; +0.019] |
| Mostly dollar | 41 | +0.01 / +0.21 | +0.019 [−0.075; +0.118] | −0.025 [−0.077; +0.027] |
| Tilt (gold + dollar) | 231 | +0.30 / +0.02 | +0.035 [−0.017; +0.085] | −0.000 [−0.024; +0.021] |
| Fight (gold vs dollar) (indicative) | 23 | −0.13 / +0.33 | −0.100 [−0.236; +0.044] | −0.116 [−0.240; −0.013] |
| All classes | 537 | +0.34 / +0.05 | +0.026 [−0.006; +0.059] | −0.007 [−0.022; +0.007] |
Afternoon = 14:25 to the close, decomposed with the same basket, in prior-day ATR and signed in the direction of the morning move; a negative value means that part gave ground. Day-block bootstrap intervals. Post-hoc (adversarial audit, attack 4).
E. US releases non-conclusive, and nothing points its way
On the 98 days with a 13:30 or 14:30 Paris release (NFP, CPI, PPI, retail sales) and a classifiable move from 09:00 to the last bar before it, the reaction one hour after the release continued the pre-release move 47.7% of the time [33.8; 62.1] when that move was broad gold (n = 44) and 58.3% [32.0; 80.7] when it was mostly dollar (n = 12): a gap of −10.6 [−41.7; +25.4] points, permutation p 0.53, in the direction opposite to the hypothesis. At the close, 63.6% against 58.3%, +5.3 [−26.4; +38.1], p 0.77. All classes together: 51.0% at one hour and 57.1% at the close. Twelve mostly-dollar days is a small sample, but the failure is not only one of power: the continuous version of the same test on this population gives +2.71 [−2.50; +9.22], and every class sits within a few points of a coin flip.
The adversarial audit tried to rescue it. Two pre-release windows (from 09:00 and from 01:00), six release types, six definitions of the reaction (total, gold part, dollar part, at one hour and at the close) and four class contrasts: 288 cells. 2 reached p < 0.05 where 14.4 would by chance; both are CPI days with twelve and thirteen observations, and both run against the hypothesis. Cells with at least 20 observations per group, a gap above 20 points and p < 0.05: 0.
What the release days do show, descriptively, is where the reaction comes from. In the hour after a release the median dollar share of the move is 0.70 (NFP 0.66, CPI 0.57, PPI 0.72, retail sales 0.77), and 59% of reactions are more than half dollar; on a normal morning the median share is 0.31 and 32% of moves are more than half dollar. Two thirds of a US-data reaction in gold is a dollar move. That is a sentence the indicator can display; it is not a forecast.
F. Symmetry, placebos and robustness nothing moves
Rises and falls. The raw return gap is larger on rises (−40.3 points, 322 mornings) than on falls (−24.5, 222), which is what a bull market in gold does to the large broad-gold moves; the difference of the differences is −15.7 [−47.0; +15.8] on the return and −0.6 [−28.9; +25.5] on the state, and the class-by-direction interaction in the adjusted model is +0.01 [−1.13; +1.11] and −0.33 [−1.09; +0.54]. Symmetric.
Placebos, and what they taught us. We had written that the placebos should "cancel any gap". P1 does not: with yesterday's currencies and today's gold, so that the label knows nothing about today's dollar, the raw return gap is still −17.7 [−40.2; +4.3] points, 56% of the observed one. That is the proof that the raw gap is not about the dollar: a large move in XAU/USD makes the dollar share small whatever the currencies did, so "broad gold" keeps selecting the large moves. Adjusted, the P1 gap is +5.8 [−14.9; +22.1]. P2, the label permutation, confirms that the raw gap is real (p < 0.01) and that the stratified one is borderline (p 0.04, with one stratum of four). P3, the random-sign basket, is the placebo the code audit found broken in the first run; corrected, it produces placebo classes of a different size (about 25 "broad gold" against 230 "mostly dollar" per draw), and is published as a shape check, not a test. P4, the NOK/MXN/ZAR basket, agrees with the primary one on the class 55.1% of the time and reproduces the structure: raw return gap −13.2 [−28.2; +2.5], adjusted +1.2 [−13.6; +16.4]. The only control that cancels the gap is the amplitude control, because the gap is amplitude.
Robustness. The adjusted gap, broad gold minus mostly dollar on the return to origin, under each variant of the definitions, one at a time:
| Variant | n (BG / MD) | Raw gap, state | Raw gap, return | Adjusted, state | Adjusted, return |
|---|---|---|---|---|---|
| Base: 6 currencies, equal weights, 25% threshold, 0.10% gate | 544 (244 / 41) | +8.8 | −31.5 | −2.5 [−13.5; +8.9] | −6.7 [−22.9; +7.1] |
| DXY-like weights, SEK in, AUD out | 543 (236 / 32) | +13.6 | −26.9 | +1.9 [−10.1; +16.1] | −2.2 [−17.8; +11.7] |
| Median of the six returns | 544 (255 / 35) | +12.0 | −26.4 | +1.3 [−11.5; +16.5] | −5.4 [−21.3; +11.6] |
| 15% threshold | 544 (152 / 25) | +10.8 | −33.7 | −1.2 [−16.6; +13.3] | −9.3 [−28.7; +10.0] |
| 35% threshold | 544 (294 / 60) | +10.4 | −23.1 | +0.0 [−8.8; +9.6] | −1.6 [−14.3; +9.1] |
| Log returns | 544 (244 / 41) | +8.8 | −31.5 | −2.5 [−13.5; +8.9] | −6.7 [−22.9; +7.1] |
| FX tolerance 60 min (recovers part of 2023) | 574 (252 / 45) | +8.7 | −30.8 | −2.8 [−13.2; +9.0] | −6.1 [−22.1; +6.5] |
| Without the 12 Mondays anchored at the 02:00 open | 534 (236 / 40) | +8.9 | −29.5 | −2.6 [−13.9; +9.2] | −4.7 [−21.2; +8.9] |
| Without the 31 US early-close days | 524 (232 / 40) | +8.7 | −29.1 | −3.6 [−15.2; +8.7] | −2.7 [−19.9; +11.3] |
| Without the clock-desynchronised weeks | 505 (227 / 33) | +11.5 | −36.3 | −0.5 [−12.2; +13.1] | −11.4 [−29.4; +4.7] |
Morning move, normal days, close of day. Adjusted = average marginal effect of a logit on class, amplitude, prior-day regime, direction and year, day-block bootstrap intervals with 300 regression draws for every row of this table; the headline estimate quoted in the text uses 1,000, so the base row's bounds differ from it by Monte-Carlo noise of about half a point. The 0.25 ATR gate leaves 12 mostly-dollar mornings and the single years 6 to 12; those rows are indicative and not shown. The prior-day regime, in every regression, changes nothing: mostly-dollar mornings are a little more often "range" days (41% against 27%), which is another way of saying small.
What the audit did to our own numbers
The first run reached the same conclusion as this paper. Three of its supporting numbers did not survive as stated, one of its choices had not been declared, and our pre-written bet was wrong in one place.
- The random-sign placebo was broken. The code audit found that the first run read each morning's move from a table that had been re-sorted, so 99.97% of the placebo rows compared a randomised basket with another day's move. Every number from that block, and the sentence we had drawn from it, "the placebo does not re-centre the gap", were artefacts. Corrected, the placebo turns out to be structurally unable to test anything, because flipping signs shrinks the gold part and makes "broad gold" the rare, small class (25 against 230 per draw). It is kept as a shape check and labelled as such.
- The release-day pool mixed in FOMC days. The first run's "all releases" pool (114 days) included FOMC days whose pre-release window runs to 19:55 and whose "+1 h" is 20:55. Its gap of +2.3 points in favour of the hypothesis rested on three mostly-dollar FOMC rows. Restricted to 13:30/14:30 releases (98 days) the gap is −10.6 [−41.7; +25.4]; FOMC days (16 classifiable) are a separate indicative block that this paper does not cite.
- An undeclared anchor. On 12 Mondays in 2026 the session opened after 01:00 and the script silently anchored the move at the 02:00 open; the blind replication caught it. Eight of those mornings were broad gold and all eight favoured the hypothesis. Declared, and published as a sensitivity: without them the adjusted gaps are −2.6 [−13.9; +9.2] and −4.7 [−21.2; +8.9].
- The matched estimator was weaker than it looked. 41 mostly-dollar controls for 168 matched broad-gold mornings, one control reused up to 28 times, an effective sample of 17.2. It is published with that number and demoted to indicative; the regression is the estimate of record.
- The bet. Right on H1 (nil after size, residual sign as announced), right on H3, and too indulgent on H2, which we had called "plausible but under-powered": at 98 days every cell is a coin flip and the continuous version is at zero, so the honest verdict is "nothing points its way". Our expectation that the placebos would cancel the gap was mis-posed: a placebo that destroys the currency information does not destroy the size information, and the gap was size.
Between the first run and the audited one, 3,020 leaves of the results file changed value, every one of them traceable to a named correction, and the headline table, the Asia and London tables, the direction split and the class and outcome of every morning are unchanged row for row. The register was rebuilt at the same time: 3,723 pre-registered comparisons (904 on fewer than 25 events, flagged) and 224 post-hoc ones.
How to use this
No entry, stop, target or cost was simulated; nothing here is a trading rule. What the data offers is a calibrated expectation for a morning move in XAU/USD, and the variable it is calibrated on is the size of the move, not its composition. This is the table worth keeping:
| Size of the 01:00–14:25 move (prior-day ATR) | Mornings | Still on the move side at the close | Returned to origin before the close |
|---|---|---|---|
| ≤ 0.15 | 108 | 67.6% [58.3; 75.7] | 66.7% [57.3; 74.8] |
| 0.15–0.22 | 70 | 78.6% [67.6; 86.6] | 44.3% [33.2; 55.9] |
| 0.22–0.30 | 83 | 74.7% [64.4; 82.8] | 48.2% [37.8; 58.8] |
| 0.30–0.40 | 83 | 83.1% [73.7; 89.7] | 36.1% [26.6; 46.9] |
| 0.40–0.55 | 72 | 87.5% [77.9; 93.3] | 25.0% [16.4; 36.1] |
| 0.55–0.75 | 70 | 94.3% [86.2; 97.8] | 5.7% [2.2; 13.8] |
| 0.75–1.00 | 41 | 100.0% [91.4; 100.0] | 2.4% [0.4; 12.6] |
| > 1.00 | 17 | 100.0% [81.6; 100.0] | 5.9% [1.0; 27.0] |
Normal days, all four classes pooled, n = 544; Wilson 95% intervals. The two rows above 0.75 ATR hold 41 and 17 mornings. Post-hoc (the pre-registered terciles, with edges at 0.22 and 0.43 ATR, tell the same story in three rows).
- Read the indicator for what it measures: where the move comes from. A morning rise that is gold in every currency and one that is the dollar sliding are different events, and the indicator tells them apart exactly. On US-release days it will show that about two thirds of the first hour is dollar. That is useful context; it is not a probability of anything.
- Do not read the class as a persistence filter. At equal size, a mostly-dollar move and a broad-gold move held into the close and into the next morning equally often, to within about two and a half points, in Asia, in London, on quiet days and on data days. A "broad gold" label is not confirmation; a "mostly dollar" label is not a warning.
- Read the size. Under 0.15 ATR, two mornings in three come back to their starting price before the close; between 0.40 and 0.55 ATR, one in four; above 0.55 ATR, about one in twenty. If a class seems to carry information, check its amplitude first: "mostly dollar" and "fight" are, nine times in ten, small moves.
- Do not expect the afternoon to "give the dollar back". It does not, measurably; the afternoon is dominated by gold-side noise whatever the morning was made of.
The sentence we now use instead
"The Gold Strength Basket tells you where a move comes from, not where it goes. On 544 normal mornings since 2023, a 'mostly dollar' move returned to its starting price before the close about 59% of the time against about 27% for a 'broad gold' move, and only because it was half the size (0.22 against 0.50 ATR): at equal size, an ATR of 'dollar' held as well as an ATR of 'gold', to within about two and a half points. What predicted persistence was size: under 0.15 ATR, two moves in three came back; above 0.55 ATR, about one in twenty. On US-data days, the reaction to the release was about two thirds dollar whatever the pre-release move was made of."
Limits
- The class of interest is rare. 41 mostly-dollar mornings on normal days, 12 / 12 / 11 / 6 per year, 12 on release days; any mostly-dollar cell split by year, stratum or release type is indicative, and the two-and-a-half-point precision comes from the continuous model, not from the class counts.
- The 2023 FX gap. The currency feed is missing about a quarter of its hours from February to July 2023, mostly in the US session. A window is used only if all six pairs have a close within 10 minutes of both instants, which leaves 55.3% of 2023 mornings valid against 96.5%, 100.0% and 97.4% in the following years, and almost nothing from March to June 2023. No window is ever classified on a partial basket. A 60-minute tolerance adds 30 mornings and moves no adjusted gap by more than a point; "excluding February–July 2023" is not a meaningful sensitivity, since only 7 windows from that period are in the base sample.
- One basket family. Six currencies, equal weights, AUD in and SEK out, the indicator's default. DXY-like weights and the median change which mornings are "mostly dollar" (32 and 35 instead of 41) without changing the reading; the exotic basket agrees with the primary one only 55% of the time. The "dollar part" is the average move of a basket against the dollar over a window; it is not an economic attribution, and the identity says nothing about which leg led.
- Release days. 98 classifiable pre-release moves, three or four mostly-dollar ones per release type. The class of a pre-release move depends heavily on where the window starts (only 21 of 44 broad-gold moves from 09:00 are still broad gold when measured from 01:00). FOMC days are outside the pool. Removing the six days on which the release fell at 13:30 Paris gives −19.7 points at one hour.
- Horizons and clocks. The +1 h and +2 h horizons of the morning window (15:25, 16:25) cross the US open, by construction of the protocol; the London window has only five and a half hours to the close, which lowers every return rate there. 31 US early-close days and 69 days whose last bar is 21:55 are kept, with sensitivities published.
- Multiplicity. 3,723 pre-registered and 224 post-hoc comparisons, no correction: roughly 200 "effects" at the 5% level are expected by chance. The candidates are named: a −39.8-point return gap in the 0.20–0.37 ATR overlap cell (16 mostly-dollar mornings, family-wise p 0.19, sign reversed in the neighbouring cell), a London return gap of +9.6 [−0.5; +18.7] after adjustment that runs the other way, and a pre-release state coefficient of −5.13 [−13.95; +0.02]. The U-shaped profile of the dollar-share deciles is not a dose–response: its two extreme deciles are the smallest moves and the residuals against the amplitude curve are not monotone.
- Coverage, instrument, strategy. Dukascopy spot XAU/USD, the same frozen store as Papers #1 to #3, through 3 August 2026; HistData bid quotes for FX; spot is not futures; no volume; gold rose over most of the sample, which is why the raw return gap is larger on rises. No strategy was simulated, so there is no expectancy, no R and no win rate. Correlation is not causation, and an accounting identity is not a model of who moves whom.
Methodology appendix
Definitions
- Windows: trading day = Paris calendar day, bars 01:00 ≤ t < 23:00, at least 100 bars, with a prior-day ATR14 and regime. Morning 01:00–14:25 (H1), Asia 01:00–08:55, London 09:00–17:25, pre-release 09:00 → release − 5 min (14:25, or 13:25 in the weeks when the US had changed clocks and Europe had not). Anchor t0 = close of the last bar before the window (the indicator's px[1] convention; on 12 Mondays of 2026 with a 02:00 session open, the first bar's open, flagged and published as a sensitivity); t1 = close of the window's last bar; r = close(t1)/close(t0) − 1.
- Attribution: XAU in currency c = XAUUSD / cUSD for EUR, GBP, AUD and XAUUSD × USDc for JPY, CHF, CAD; gold part g = equal-weight mean of the six returns; dollar part d = r − g; dollar share s = d / |r|. Classes at threshold 25%: broad gold if sign(g) = sign(r) and |d| ≤ 0.25|r|; mostly dollar if |g| ≤ 0.25|r|; tilt if sign(g) = sign(r) and neither; fight otherwise; mixed = tilt ∪ fight. A window is classified only if |r| ≥ 0.10% (the indicator's minMove gate); below it, "quiet". Exactly one class per window, verified.
- Outcomes from t1 to horizon h (+1 h and +2 h = the bar whose close falls 60 / 120 minutes after t1, 17:25, the close = last bar before 23:00, next day 08:55): state S_h = close_h on the move's side of close_t0; return R_h = at least one M5 close in ]t1; h] beyond close_t0 against the move; follow-through F_h = sign(r) × (close_h − close_t1) / |close_t1 − close_t0|; MFE / MAE in ATR. Amplitude = |close_t1 − close_t0| / ATR14(D−1), Wilder, daily, lagged. H2 reaction = sign(r_pre) × (close_h − close_t1), continuation if positive; "+1 h" = close of the bar stamped release + 55 min.
- Controls: amplitude terciles of the base population (edges 0.224 / 0.431 ATR), Mantel–Haenszel primary and equal weights beside; logit / OLS on class + amplitude + prior-day regime (ADX14/DI, four states) + direction + year, average marginal effect; 1:k ≤ 3 matching on regime and amplitude within 0.10 ATR, one treated morning = one weight, effective control count published. Amplitude curve: pooled logit on log(amp) and log(amp)², with leave-class-out variants. Continuous model: logit on g_dir + d_dir (the two parts in ATR, signed with the move, summing to amplitude) + regime + direction + year; model B adds d_dir to a flexible function of amplitude and reports the AME of a 0.1 ATR transfer.
- Placebos: P1 = FX ratios of the previous valid row of the same window and the same clock (up to five rows back), today's gold; P2 = class labels permuted across days, 1,000 draws, seed-per-cell; P3 = each currency return × ±1 drawn per day and currency, 200 draws, shape check; P4 = USDNOK / USDMXN / USDZAR basket, descriptive.
Data
Gold: 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 to #3; 940 calendar days, 925 with at least 100 bars, 923 with a prior-day ATR and regime: 743 normal and 180 US-release days; 255 / 259 / 258 / 151 per year. Currencies: HistData one-minute bid quotes for EURUSD, GBPUSD, USDJPY, USDCHF, AUDUSD, USDCAD (primary basket), USDSEK (DXY-like variant), USDNOK, USDMXN, USDZAR (P4), aggregated to five minutes with the house convention. HistData's own XAUUSD series matches the Dukascopy store bar for bar (median close difference 0.00 dollars, five-minute return correlation 0.99999 at lag zero, 0.006 at one bar either side), which establishes the time alignment; it also means the two feeds are the same source, so the FX data is not an independent price check. HistData's timestamps follow European summer time, contrary to its documentation; that was verified empirically against the gold store and corrected before any computation.
The honest part. From February to July 2023 the HistData feed is missing whole hours at random, about a quarter of them, mostly 13:00–23:00 Paris. A window is valid only if all six primary pairs have a close at most 10 minutes before both t0 and t1: 2,561 of 2,949 day-windows are valid at 10 minutes, 2,669 at 30 and 2,781 at 60; 544 normal mornings and 119 release-day mornings clear both the FX rule and the 0.10% gate. The headline uses the 10-minute rule; the 30- and 60-minute variants are published and change nothing. Release calendar: rebuilt from the BLS, Federal Reserve and Census schedules, 198 releases on 182 dates (180 in the store; the two Good Fridays are not), release time computed per date from 08:30 / 14:00 New York. The calendar used in our earlier studies had three mislabelled dates (a retail-sales day marked NFP, two retail-sales days marked CPI) and omitted most of 2025–2026; it was not reused, and the three errors are recorded.
Statistics
- Wilson 95% on every proportion; day-block bootstrap (1,000 resamples of whole trading days, seed 42, one shared weight matrix) on every difference, 1,000 regression draws for the headline and 300 for sensitivities, 500 for the post-hoc block; permutation (1,000) for H2 and P2, 2,000 within-stratum permutations for the family-wise test of the common-support search. The bootstrap governs where it disagrees with Wilson. Cells under 25 events are flagged indicative and none carries a claim in this paper.
- Register: 3,723 pre-registered comparisons (deduplicated from 3,990 raw values) and 224 post-hoc, no multiplicity correction. Every number here is reproducible from gold_attribution_study.py v2 and gold_attribution_stats.json (60.5 s, two consecutive runs byte-identical); the identity r = g + d holds to machine precision on every row, and a hand recomputation of one morning (3 June 2024: r +0.233%, g +0.165%, d +0.068%, class tilt) matches the script.
Appendix table — what the Pro version of the indicator may display
Class × horizon, morning move on normal days, as historical frequencies with their sample size and Wilson interval. These are descriptive; if displayed, they belong next to the amplitude of the current move, never as a filter. Fight (n = 23) is indicative at every horizon.
| Class | Horizon | n | On move side | Wilson 95% | Returned to origin | Wilson 95% | Mean follow-through |
|---|---|---|---|---|---|---|---|
| Broad gold | +1 h (15:25) | 244 | 96.7% | [93.7; 98.3] | 6.1% | [3.8; 9.9] | +0.02 |
| Broad gold | +2 h (16:25) | 244 | 93.4% | [89.6; 95.9] | 14.3% | [10.5; 19.3] | −0.18 |
| Broad gold | 17:25 | 244 | 91.8% | [87.7; 94.6] | 18.9% | [14.4; 24.2] | −0.08 |
| Broad gold | Close (EOD) | 244 | 86.9% | [82.1; 90.6] | 27.0% | [21.9; 32.9] | −0.04 |
| Broad gold | Next day 08:55 | 244 | 82.4% | [77.1; 86.6] | 34.8% | [29.1; 41.0] | −0.14 |
| Mostly dollar | +1 h (15:25) | 41 | 92.7% | [80.6; 97.5] | 12.2% | [5.3; 25.5] | +0.24 |
| Mostly dollar | +2 h (16:25) | 41 | 87.8% | [74.5; 94.7] | 31.7% | [19.6; 47.0] | +0.55 |
| Mostly dollar | 17:25 | 41 | 82.9% | [68.7; 91.5] | 41.5% | [27.8; 56.6] | +0.41 |
| Mostly dollar | Close (EOD) | 41 | 78.0% | [63.3; 88.0] | 58.5% | [43.4; 72.2] | +0.21 |
| Mostly dollar | Next day 08:55 | 40 | 72.5% | [57.2; 83.9] | 65.0% | [49.5; 77.9] | +0.57 |
| Tilt (gold + dollar) | +1 h (15:25) | 236 | 90.3% | [85.8; 93.4] | 15.3% | [11.2; 20.4] | +0.02 |
| Tilt (gold + dollar) | +2 h (16:25) | 236 | 82.6% | [77.3; 86.9] | 28.8% | [23.4; 34.9] | +0.06 |
| Tilt (gold + dollar) | 17:25 | 236 | 80.9% | [75.4; 85.4] | 33.9% | [28.2; 40.2] | −0.04 |
| Tilt (gold + dollar) | Close (EOD) | 236 | 80.1% | [74.5; 84.7] | 39.8% | [33.8; 46.2] | +0.08 |
| Tilt (gold + dollar) | Next day 08:55 | 236 | 76.7% | [70.9; 81.6] | 47.5% | [41.2; 53.8] | +0.19 |
| Fight (gold vs dollar) ⚑ | +1 h (15:25) | 23 | 91.3% | [73.2; 97.6] | 8.7% | [2.4; 26.8] | +0.52 |
| Fight (gold vs dollar) ⚑ | +2 h (16:25) | 23 | 65.2% | [44.9; 81.2] | 52.2% | [33.0; 70.8] | −0.26 |
| Fight (gold vs dollar) ⚑ | 17:25 | 23 | 60.9% | [40.8; 77.8] | 56.5% | [36.8; 74.4] | −0.62 |
| Fight (gold vs dollar) ⚑ | Close (EOD) | 23 | 56.5% | [36.8; 74.4] | 56.5% | [36.8; 74.4] | −0.88 |
| Fight (gold vs dollar) ⚑ | Next day 08:55 | 23 | 52.2% | [33.0; 70.8] | 65.2% | [44.9; 81.2] | −1.04 |
| Mixed (tilt + fight) | +1 h (15:25) | 259 | 90.3% | [86.1; 93.4] | 14.7% | [10.9; 19.5] | +0.06 |
| Mixed (tilt + fight) | +2 h (16:25) | 259 | 81.1% | [75.9; 85.4] | 30.9% | [25.6; 36.8] | +0.03 |
| Mixed (tilt + fight) | 17:25 | 259 | 79.2% | [73.8; 83.7] | 35.9% | [30.3; 41.9] | −0.09 |
| Mixed (tilt + fight) | Close (EOD) | 259 | 78.0% | [72.6; 82.6] | 41.3% | [35.5; 47.4] | +0.00 |
| Mixed (tilt + fight) | Next day 08:55 | 259 | 74.5% | [68.9; 79.4] | 49.0% | [43.0; 55.1] | +0.08 |
Source: metric9_indicator_table, a digest of the pre-registered follow-through cells (same definitions, no new estimates). Follow-through in units of the original move (mean; the medians are near zero for every class). ⚑ = fewer than 25 events. Next-day cells use the following trading day in the store.
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 and six currency pairs did between January 2023 and August 2026 under precisely stated definitions; it does not describe what they will do, and it does not evaluate any trading strategy.
Research like this is what the Pro tier funds.