Gold fell 8.2% in a month: the move from 4,331 to 3,975, in data

Between 25 September and 23 October 2026 spot gold fell from 4,331.53 to 3,975.74, −8.21% in a month. The number itself is not remarkable. What is remarkable is how it fell: not in one overnight collapse, but in a dozen separate pieces, crowded into the same few hours each day, with 107 separate fights around the round number 4,000.

This article does not retell the news and does not guess at causes. It does one thing: take 444,539 tick quotes and slice them by hour and by horizon to measure where the price actually moved, by how much, and for how long. Every figure can be reproduced from the same quotes table; the sampling time and conventions are set out below.

1. Conclusions first: eight numbers you can check yourself

Compress 30 days of tick data into one table and you do not get a straight line down. You get two sharp legs, a sideways stretch and a slow bleed:

"Largest 5-minute rise/fall" means the biggest deviation from the price 300 seconds earlier at any point in the sample, taken once in each direction; both extremes coming out at exactly 39.10 USD is a coincidence, not a typo.

Four readings, all backed by the data:

• The fall is concentrated in two days. 09-28's −161.60 plus 10-22's −101.85 comes to −263.45, or 74% of the total 355.79 peak-to-trough decline. The other 20 sessions together contribute only a quarter.
• There was a real bounce in the middle. Across the 14 sessions from 09-29 to 10-16 the price went from 4,140 back to 4,161, with an intraperiod high of 4,235.83 (10-02). So this was not "a straight line down" but "plunge — range — plunge again".
• Both 5-minute extremes happened at 20:30. 20:30 Beijing is 08:30 New York, the standard window for US economic data. That is not a coincidence; sections four and five unpack it.
• The last leg was a bleed, not a crash. 10-23 fell only 25.78 USD (−0.64%), yet the price traded below 4,000 in seven separate stretches lasting more than ten minutes, closing at 3,997.89, just −0.05% from the level — ground down, not smashed through.

2. The 23 sessions, laid out day by day

This table is the backbone of the article; every later statistic comes from it and from the tick series. Columns three to six are open, high, low and close on the 08:00 Beijing convention:

The first session (09-24) samples from 12:04 Beijing and is not a complete trading day; 10-18 is a Sunday stub (quotes from 14:04, hence a range of only 21.64). Neither enters the average-range statistics in section seven. There are no weekend quotes and no rows for them; change is measured from the previous session's close to this session's close.

The row worth staring at is not 09-28 but 10-22: a high of 4,126.10 and a low of 4,021.30, with not a single print back above 4,130 all day, closing at 4,023.67 — the first close below 4,100 in the whole month (the previous 20 sessions had all closed above it), while also breaking 4,050 intraday. The loss of 4,000 the next day was a continuation of that long down candle, not a separate event.

3. Which hours it fell in: 21:00–24:00 is the main battlefield

Slice the 23 sessions by hour, compute "price at the end of the hour minus price at the start", and average across the sessions that have quotes in that hour (most hours cover 22–23 sessions). A positive number means the hour rose on average, a negative one means it fell:

Three conclusions:

• The three hours from 21:00 to 24:00 contribute −4.4 USD an hour on average. Summed, they come to −13.21, which is 103% of the average close-to-close change across 22 sessions (−12.86 USD a day). In other words, if you had held only during those three hours each day, you would have taken this month's entire decline; conversely, 23:00 to 10:00 the next morning was a net bounce (+2.28, +2.32, +1.19, −1.28, −1.24).
• The Asian session is not a "safe zone" but a "slow-bleed zone". 11:00–14:00 fell three hours in a row (−1.99, −1.85, −2.70); each hour is only half the size of the New York hours, but the direction is the same. That is this month's daily pattern: a plunge during European hours, a grind lower through Asia, an hour of bounce at 10:00, then more of the same.
• 10:00–11:00 is the only reliably positive hour in the whole month (+4.04, the highest in the table). It is the second hour after the 09:00 Tokyo open, when Asian physical buying actually shows up; the same pattern appears in the measured readings in gold trading sessions, measured.

4. The two most violent 5-minute windows both landed at 20:30

Across all 444,539 ticks, the largest 300-second deviation occurs as an extreme only twice, in opposite directions, at almost the same time of day:

20:30 Beijing is 08:30 New York — the standard release time for US economic data, jobless claims, retail sales and the like. This is not speculation: the tick series shows that on 10-14 between 20:29:30 and 20:43 the price travelled 66.10 USD, equal to 68% of that day's 97.00 range; on 10-02 it covered 39.00 USD in five and a half minutes, 39% of the day's 99.80 range.

For orders held from 60 seconds to five minutes, these two windows are not something you can price with your normal odds:

• The price jumps rather than slides. In the three minutes after 20:30 on 10-14 the mid price shifted gear directly between 4,154, 4,131 and 4,118, with a dozen dollars between each step.
• The bid/ask gap also blows out instantly at such moments: the historical record shows spread extremes of 9.1 and 57.0 USD, ten to a hundred times normal. The breakdown is in comparing gold spreads.
• The direction of the data is not predictable, but the timing is: 20:30, 21:30 (US equity open) and 23:00 (after the London fix) are three known gear changes. The mechanics are in CPI and gold and the Fed, FOMC and gold.

The only practical rule: do not open a new order in the three minutes either side of those times. Not because it must lose, but because the odds structure there is not the one you priced.

5. How big is the real 60-second and 5-minute move?

This is the table to reread more than any other. For every tick, take the difference from the mid price 60 seconds and 300 seconds earlier, take the absolute value, and tabulate the distribution (n = 444,220 / 444,080):

In plain terms: over 60 seconds the median move is just 0.47 USD; nine times out of ten it is under 1.83; to get a move above 4 USD you need the top 1% of windows. At five minutes the median is 1.10 and the 90th percentile is 4.02.

Now the direction split (a tie is defined as |change| ≤ 0.01, matching the site's settlement tie line):

This table shows something easily missed: the longer the horizon, the clearer the direction; the shorter, the larger the share of noise. At 60 seconds the gap between up and down is only 2.28 percentage points; at 30 minutes it widens to 8.62. Trend gets chopped into noise at the short end. Which is why "trading 60 seconds" and "trading 30 minutes" are not the same job, and not the same technique.

Finally, put the odds back in. The site's /api/config currently returns these payout rates and the tie line (sampled 2026-10-24 12:00): 60s 0.68, 5m 0.80, 10m 0.82, 15m 0.85, 30m 0.88, 60m 0.90, with tie_pts = 0.01. The win rate needed to break even is 1 ÷ (1 + payout):

Two details that must be worked out:

• A tie refund does not move the break-even bar. Within 60 seconds, 18.68% of windows are ties with the stake returned; those neither add to the win rate nor subtract from capital, so "of the remaining 81.32%, 59.52% must win". The bar has not moved, the sample has just got smaller.
• The measured direction was bearish, and nowhere near break-even. Over this window, buying at random won 39.52% and selling at random 41.80% — both more than 14 points below the 55.56% five-minute break-even line. In other words, during this month, entering with no directional view was a reliably negative-expectancy exercise; whether you make money depends only on directional judgement and horizon choice, not on how many orders you place.

6. The 4,000 level: not one break, but 107 fights

4,000 was the most watched level in this decline, and the data shows a different picture from intuition. Counting tick mid prices, the sample contains 107 separate stretches below 4,000 — but only seven lasted more than ten minutes, the longest 61 minutes:

"Separate stretch" is defined by the tick mid crossing 4,000, so it includes 0.1 USD round trips (100 of the 107 last under ten minutes). These are not "107 valid breakouts" — quite the opposite: the number quantifies how dense the churn around the level was.

Three things you can act on:

• The first break was not the last. After the 09:55 break on 10-23 the price was back above within 10.8 minutes, then broke seven more times. Using "first break" as a trend signal was simply wrong this month.
• The real breakdown happened after 01:05 on 10-24. The two stretches 01:05–01:58 and 02:37–03:38 total 114 minutes below 4,000 and the price never recovered them — that is the line between "cannot hold" and "can hold". But there was no acceleration afterwards: the four lows ran 3,975.74 (22:30), 3,990.99 (01:40), 3,993.88 (03:09), 3,992.26 (04:46) — lows rising, a pattern of grinding along the level, not free fall.
• The close was 0.05% from the level. 3,997.89 against 4,000 is effectively no move at all. That is the hardest kind of price for 60-second orders: a 50/50 direction with enough noise to trigger ties.

For how to draw the line technically and where to look after a break, continue with gold technical analysis: support, resistance and patterns; the three-month scenario work is in the 2027 gold outlook. The division of labour is "where" versus "which way" — do not mix them.

7. Volatility is compressing — the month's second signal

Using only the 21 complete sessions (excluding the 09-24 half session, the 10-18 Sunday stub and the 10-24 stub), compute each session's range (high minus low):

The second half's average range is 16.7% smaller than the first half's, yet the price kept falling — a combination of "compressing volatility, unchanged trend". For short-term trading the implication is direct: the same order has less room to earn, while the cost (spread, ties) is unchanged. The site's volatility methodology and magnitudes are broken down in how to calculate gold volatility; that article covers how to compute it, this one reports what it came to this month.

8. How to use this data for 60-second to 5-minute decisions

All three rules below follow from the measured numbers above rather than from experience:

1. Look at the horizon's p90 before you decide to trade. At 60 seconds p90 = 1.83 USD, meaning nine windows in ten give you less than 2 USD of room. If your aim is to capture a full move, the 60-second structure cannot support it — look at five minutes (p90 = 4.02) or longer.
2. Do not trade in the three minutes around 20:30. Both 5-minute extremes this month (−39.10 and +39.10) occurred within five minutes of 20:30, accounting for 39%–68% of the day's range. In those three minutes the price steps between gears and the odds structure is distorted.
3. Do not trade breakouts within ±2 USD of a round level. 100 of the 107 stretches around 4,000 lasted under ten minutes, so "passing through" is the norm and "holding" the exception. Either wait for 15 minutes of confirmation, or trade somewhere else.

On direction, this month's bearish tilt (21:00–24:00 averaging −4.4 USD an hour) is the same story as the dollar and real yields; the mechanics are in measured correlation between the dollar index, real yields and gold. But note: macro explains why it fell, not where the next tick goes — that is the problem of horizon and position size addressed in short-term gold tactics.

9. Conventions and limits: what this article does not do

To stop a review being read as a forecast, four things must be stated:

• It does not explain causes. This article describes price behaviour and offers no causality. Why gold fell this month requires the dollar index, real yields, central-bank buying and policy expectations — other articles' job. Forcing a news headline onto a price chart is the most common analytical error there is.
• The sample is one feed, one month. All 444,539 ticks come from a single Swissquote feed; other feeds can differ by 0.2–0.5 USD at the same moment, and cross-feed statistics would be noisier still. Thirty days cannot support conclusions about "seasonality" or "regularities".
• Quote density changed. Before 10-18 there were about two ticks a minute; afterwards about 45. Price statistics (range, up/down distribution) are unaffected, but count-based measures like "107 breaks" are inflated by density, so this article uses that number only to describe churn, never as a count of breakouts.
• The day convention differs from other pages. Here a "day" runs 08:00 Beijing to 07:59 the next morning, whereas the sessions article's "daily change" uses the calendar day. Closes for the same date can differ by a few dollars, so align conventions before comparing. For the live reading see today's gold price.

10. FAQ

Q. How much did gold actually fall this month?
A. Peak to trough −8.21% (4,331.53 → 3,975.74); measured from the first and last session closes, −6.61% (4,280.83 → 3,997.89). Sample: 2026-09-24 12:04 to 2026-10-24 04:55 Beijing, Swissquote tick mid prices, sampled at 2026-10-24 12:00.

Q. Which part of the day was worst?
A. 21:00–24:00 Beijing. Averaged across 23 sessions, 21:00–22:00 was −5.46 USD, 22:00–23:00 −4.00 and 23:00–24:00 −3.75, a combined −13.21 — more than the average close-to-close change of −12.86 across 22 sessions.

Q. How much does gold usually move in 60 seconds?
A. Median 0.47 USD, mean 0.74, p90 1.83, p99 3.99, maximum 28.10. A further 18.68% of 60-second windows barely move (|change| ≤ 0.01) and return the stake under the site's rules.

Q. Has 4,000 been decisively broken?
A. On the site's data: the first break was 09:55 on 10-23, after which 107 stretches traded below 4,000 — 100 of them under ten minutes, seven over ten minutes, the longest 61 minutes — closing at 3,997.89, −0.05% from the level. That is grinding along the level, not a clean break.

Q. Why did gold fall this month?
A. This article does not attribute causes, only reviews price behaviour. For the mechanics see the pieces on the dollar index and real yields (measured correlation), the FOMC and CPI — they explain what gold is sensitive to, not why it fell this time.