ORIGINAL RESEARCH · 257 SESSIONS

How long do gold losing streaks actually last?

Gold closed lower on 45.9% of sessions over a year. The longest run of consecutive down days was 9 — but a three-day run cost almost exactly as much, which is the more useful finding.

What this measures

Close-to-close direction of gold itself, by broker server day, across 257 sessions from 2025-07-31 to 2026-08-04. It is built from 70,952 five-minute bars — the same archive behind our spread and volatility study.

There are no entries, signals or position sizes in it. This is what the instrument does, which is the base rate any strategy is layered on top of. It is deliberately not a measurement of our bot — that would be a performance claim from a 42-trade sample, and the two are different things.

Gold falls on roughly half of all days

118 of 257 sessions closed lower — 45.9%. The median down day moved $30.85, and the worst decile moved $122.41. That near-coin-flip is the expected shape and it is worth stating, because it is the baseline the streak numbers below have to be read against: runs happen in any sequence of near-even outcomes, and their existence is not evidence of a trend.

The distribution of losing runs

There were 61 losing streaks in the period. The "at least" column is the one to read if you are setting a limit — it is how many times a run of that length or worse actually occurred.

Consecutive down days	Occurred exactly	Occurred at least this long
1 day	31	61
2 days	16	30
3 days	7	14
4 days	5	7
5 days	1	2
6 days	0	1
7 days	0	1
8 days	0	1
9 days	1	1

The longest run was 9 consecutive down days. A run of three or more happened 14 times in a year; four or more, 7 times; five or more, 2. Anyone who assumed a bad week is two or three days has the shape roughly right and the tail badly wrong.

The finding that actually matters

Streak length does not predict how much it costs. Ranked by cumulative move rather than by duration:

Streak length	Total move against you
9 days	$755.7
3 days	$738.84
4 days	$414.39
1 day	$308.54
2 days	$272.71

The 9-day run cost $755.7. The 3-day run cost $738.84 — 98% of the damage in a third of the time. And a single day accounted for $308.54 on its own.

That is also why we publish all three of our measured drawdowns rather than the kindest one, and say plainly on the prop firm page that two of them would breach a 10% limit. The same argument applies with more force to a trailing drawdown, which is anchored to your peak rather than your starting balance and is therefore live almost all the time.

How to cite this

Published under CC BY 4.0, which permits reuse — including commercial reuse — on the single condition that it is attributed. If you are quoting a figure from this page in an article, a paper or a model answer, this is the attribution:

Tech Kick (2026). "XAUUSD consecutive losing days, 2025-2026"
Derived from 70,952 M5 bars across 257 full sessions.
Licence: CC BY 4.0. DOI: 10.5281/zenodo.21973215
https://techkick.me/research/xauusd-losing-streaks

Or as BibTeX:

@misc{techkickxauusdlosingstreaks,
  author       = {Tech Kick},
  title        = {XAUUSD consecutive losing days, 2025-2026},
  year         = {2026},
  doi          = {10.5281/zenodo.21973215},
  howpublished = {\url{https://techkick.me/research/xauusd-losing-streaks}}
}

The headline figures, for convenience: gold closed lower on 45.9% of 257 sessions; 61 losing streaks; longest run 9 consecutive days; a 3-day run cost 98% of what the 9-day run did. Underlying data CC BY 4.0 under the DOI above.

What this does not support

It is one year and one instrument. Fifty-odd streaks is a small sample for a tail, and the single 9-day run is exactly one observation — it tells you such a run is possible, not how often to expect one. A second year could easily contain a longer one or none at all.

Gold also ran from roughly $3,288 to $4,044 across the period, so the dollar figures do not transfer to a very different price level. And close-to-close direction says nothing about what happened inside the day — a session can close lower having spent most of it higher, which matters a great deal to a stop and not at all to this measurement.

The script is published with the dataset under DOI 10.5281/zenodo.22273196, so the arithmetic can be checked rather than believed. If you find an error, tell us and the correction will appear here with the same prominence as everything else.