MEASURED · 3,079 DAYS · 12 YEARS
How often does gold reverse the previous day’s direction?
A little more often than chance, and not by enough to trade. Across 3,079 trading days of XAUUSD daily bars from 2014 to 2025, gold closed in the opposite direction to the day before 51.7% of the time. The same days shuffled into random order reversed 49.9%. Fading yesterday’s direction still averaged −0.0007% a day before costs, because the days that reversed were smaller moves than the days that carried on.
Big days did not reverse more: after moves of at least twice an ordinary day, 50.3% reversed. And in the latest year the answer depended on the clock. On US gold days, 52.3% of 222 days reversed; measured as UTC days, the same dates gave 46.4%.
How a reversal was counted
A day is up if it closed above the previous trading day’s close and down if it closed below. Each day is paired with the one before it. If the two point different ways, that is a reversal; if they match, a continuation. Monday is compared with Friday’s close, so the weekend gap counts as part of Monday.
Twelve years of daily bars. The broker’s XAUUSDm daily bars from Exness, January 14, 2014 to December 31, 2025, the file behind our Monday and Friday study. They are UTC days, so each one starts at 8 p.m. ET (7 p.m. CT) in US summer and 7 p.m. ET (6 p.m. CT) in winter, which is midnight GMT. Gold went from $1,243 to $4,319 over the sample.
The control. Gold rose on 52.7% of these days. When up days outnumber down days, chance alone does not give 50%, so every rate here is compared with the same days shuffled: each day keeps its own move, and only which day came next is randomized, 10,000 times. That is a random walk built from gold’s own days.
Size. A move’s size is its absolute change divided by the median absolute change of the 20 trading days before it. 1x is an ordinary move for that month; a big day is 2x or more.
Twelve years: slightly more reversals than chance
| Days counted | Days | Reversed | 95% interval | Shuffled days | p |
|---|---|---|---|---|---|
| Every day | 3,079 | 51.7% | 50.0-53.5% | 49.9% | 0.036 |
| After an up day | 1,623 | 49.1% | 46.7-51.5% | 47.3% | 0.036 |
| After a down day | 1,456 | 54.7% | 52.1-57.2% | 52.7% | 0.036 |
| After a big day (2x+) | 694 | 50.3% | 46.6-54.0% | 49.9% | 0.85 |
| Candle color (close vs open) | 3,078 | 51.9% | 50.1-53.6% | 49.9% | 0.027 |
p is the share of 10,000 shuffles at least as far from their own average as the real figure. The candle color row uses each bar’s close against its own open instead of the previous close.
1,593 of 3,079 days reversed. In 95% of shuffles the rate stayed between 48.1% and 51.6%, and 360 of 10,000 came as far from the shuffle average as the real 51.7%. So the tilt is probably not luck, but it is 1.8 percentage points. Reading candle colors instead of closes gives nearly the same answer, 51.9%.
The up and down rows are the same result seen from two sides, not two findings. Down days were followed by an up day 54.7% of the time, which sounds like a rule. But shuffled days did that 52.7% of the time, simply because gold rose more days than it fell. Both rows sit about two points above their shuffles, the same gap as the overall figure. The correlation between one day’s return and the next was -0.004, inside the ±0.035 that chance gives.
Why a 51.7% win rate earned nothing
Take the simplest fade: at each day’s close, bet against that day’s direction for one day. This is arithmetic on closing prices, not a backtest. It won on 1,593 of 3,079 days and averaged −0.0007% a day, against an average daily move of 0.653%. Shuffled days averaged between −0.034% to +0.030% in 95% of runs, so the real figure is unremarkable (p = 0.92).
The reason is size. Days that reversed moved 0.631% on average; days that carried on moved 0.678%. The fade won slightly more often and lost slightly more each time it lost. Over 12 years the daily results added up to −2.05%. Split in half, they added up to +15.2% in 2014-2019 and −17.2% in 2020-2025: a rule that was ahead after six years and then gave it all back. All of this is before spread, commission and swap, which a real account pays on every one of those days. A swap-free account drops only the swap, and Exness’s swap-free rules decide which accounts keep that status.
Big moves do not reverse more
The usual version of the idea is to fade a big day, not every day. The data shows the opposite shape: the small tilt toward reversal was largest after the smallest moves and gone after the largest.
| Yesterday’s move | Days | Reversed | Shuffled days | Fade, average a day |
|---|---|---|---|---|
| Under 0.5x ordinary | 832 | 52.8% | 49.9% | +0.005% |
| 0.5x to 1x ordinary | 718 | 52.1% | 49.9% | −0.001% |
| 1x to 2x ordinary | 835 | 51.6% | 49.8% | +0.006% |
| 2x to 3x ordinary | 401 | 50.6% | 49.8% | +0.007% |
| 3x or more ordinary | 293 | 49.8% | 49.9% | −0.045% |
After the quietest days, 52.8% reversed; after moves of three times an ordinary day or more, 49.8%, level with the shuffles. Fading those 293 big days averaged −0.045% a day. Splitting by direction changes nothing: after big up days 48.9% turned down (shuffled: 47.3%), after big down days 51.8% turned up (52.7%).
Runs did not make a turn more likely either. After one, two, three, and four or more days the same way, the next day reversed 51.2%, 51.6%, 55.7%, 50.0% of the time. The three-day row (375 days, p = 0.021) is the only one of the 12 subgroups on this page under the 5% line. Testing 12 subgroups at 5% gives about 0.6 false alarms on average, and the next row fell back to 50.0%. For how long down-day runs last, see our losing streak study.
Year by year
| Year | Days | Reversed | 95% interval | Up days |
|---|---|---|---|---|
| 2014 | 230 | 55.2% | 48.8-61.5% | 46.1% |
| 2015 | 260 | 50.8% | 44.7-56.8% | 46.9% |
| 2016 | 259 | 50.2% | 44.1-56.2% | 52.1% |
| 2017 | 258 | 53.9% | 47.8-59.9% | 58.1% |
| 2018 | 258 | 53.5% | 47.4-59.5% | 48.8% |
| 2019 | 258 | 50.4% | 44.3-56.4% | 55.4% |
| 2020 | 259 | 52.1% | 46.1-58.1% | 58.7% |
| 2021 | 258 | 49.2% | 43.2-55.3% | 50.0% |
| 2022 | 259 | 55.6% | 49.5-61.5% | 52.1% |
| 2023 | 259 | 49.0% | 43.0-55.1% | 49.4% |
| 2024 | 261 | 51.7% | 45.7-57.7% | 56.3% |
| 2025 | 260 | 49.6% | 43.6-55.7% | 57.3% |
9 of 12 years reversed more than half the time, and in none of them did the 95% interval leave out 50%. With about 260 days a year, a single year cannot separate a two-point tilt from noise. 2014 starts on February 13, 2014, because the file’s first weeks were used for the size baseline.
The last year, on the US clock
71,402 five-minute bars from the same broker, cut into US gold days that run from 6 p.m. ET (5 p.m. CT; 22:00 GMT in US summer, 23:00 in winter) the evening before to 5 p.m. ET (4 p.m. CT; 21:00 or 22:00 GMT). Every one of the 250 full days ended on the 4:55 p.m. ET bar, in summer and winter, which confirms the bars’ clock. Days missing more than an hour of bars were dropped (11: US holidays, early closes and the ends of the archive), and so was any pair touching one. That leaves 222 pairs from September 3, 2025 to August 4, 2026.
52.3% reversed, against 49.5% for shuffled days. With 222 days, 95% of shuffles fell anywhere from 43.2% to 55.9% (p = 0.41). Fading yesterday averaged −0.08% a day before costs: reversal days moved 1.14% on average, continuation days 1.42%.
Now measure the same 222 dates as UTC days, the way the twelve-year file is cut. The direction agreed on 208 of them. The other 14 were enough to move the reversal rate to 46.4%, below half. The day boundary moved the answer by 5.9 points, more than three times the twelve-year tilt. “Yesterday’s direction” is not one fact: it depends on where the daily bar starts.
What to do with this if you run a bot
Do not add a fade-yesterday rule, or its opposite. The direction count leans 1.8 points toward reversal, and the money did not follow: −0.0007% a day over 3,079 days before costs, and a different sign in each half of the sample. Trading with yesterday’s direction is the same bet reversed, and it has the same problem.
Test a filter against shuffled days, not against 50%. If your EA only buys after a down day, gold’s rise alone made that look right 52.7% of the time in data with no day-to-day memory at all. Judge it on the size of the wins and losses, and on enough trades: see how many trades it takes to judge an EA.
Check which day your daily logic reads. Anything that uses the previous D1 bar depends on where that bar starts. On the same year, 14 of 222 days changed direction between the US day and the UTC day. Before moving an EA between brokers, the Tech Kick Trading Bot included, check where each broker’s daily bar starts. Trading from the US? Read whether US traders can use MT5 trading bots first: the instrument measured here is a spot gold CFD.
Method and limits
One broker, one spot CFD, close to close. Days with no change at all would break a pair; there were none. The shuffle keeps each day’s move and randomizes only which day follows which, so it holds gold’s drift and its spread of daily moves fixed. The fade figures are sums of daily percentage returns, not compounded, and they include no costs. Whether losing days cluster for a strategy, rather than for the price, is a separate test, covered under equity curve trading. Every figure comes from one script, research-gold-previous-day-reversal.py, run on the same archive as our other studies.
Common questions
Does gold usually reverse after a big up day?
No more than after any other up day. After up days at least twice an ordinary move, gold closed lower the next day 48.9% of the time (364 days, 2014 to 2025), against 47.3% when the following day was picked at random. After moves of three times an ordinary day or more, in either direction, 49.8% reversed.
Is a reversal due after several days in the same direction?
Not on this data. After a run of one, two, three, and four or more days the same way, the next day reversed 51.2%, 51.6%, 55.7%, 50.0% of the time. Three days is the only row that stood out, and four or more fell back to 50.0%. For how long down-day runs last, see our losing streak study.
Does buying gold after a down day work?
It looked right 54.7% of the time: that share of down days were followed by an up day. Shuffled days did nearly as well, 52.7%, because gold rose on 52.7% of all days. Buying after down days earned +0.045% a day before costs, inside the +0.010% to +0.077% that gold's rise alone produced in shuffles.
Are gold’s daily returns autocorrelated?
Barely. The correlation between one day's return and the next was -0.004 over 3,079 days, inside the ±0.035 band that pure chance gives a sample this size. In the year to August 2026 on the US clock it was 0.021, inside ±0.132.