How Gold Actually Behaves: 16 Years, 95,000 Hours of Data
% of all H1 bars
Before you can have an edge in an instrument, you have to know its personality — how it moves, when it moves, and what it does by default. We characterised gold across 16 years: 4,195 daily bars and 95,214 hourly bars, from 2010 to 2026. A few numbers reframe how the instrument should be traded.
Gold ranges far more than it trends
Classifying every hour into a regime:
| Regime | Share of time |
|---|---|
| Range | 74.8% |
| Trend | 17.2% |
| Compression | 3.7% |
| Exhaustion | 3.5% |
| Chaos | 0.8% |
Three-quarters of the time, gold is ranging. It trends only about one hour in six. And the asymmetry goes further when you look at how long each regime lasts:
- Trend runs are short — a median of about two hours before the classifier flips.
- Range runs are sticky — a median of eight hours, and up to 242 consecutive hours in the data.
The takeaway is uncomfortable for trend-followers: gold's default state is mean reversion inside a range, interrupted by brief, violent trends. A strategy that assumes persistence is fighting the base rate three hours out of four. Confirming this, the probability that the next hourly candle closes the same direction as the current one is only ~48% — below a coin flip. There is no free momentum in gold; it has to be earned from context.
London hunts the Asian range — almost always
One of the most reliable structural facts in the dataset:
- London sweeps the prior Asian-session high 68% of the time.
- London sweeps the Asian low 66% of the time.
- London sweeps at least one side 97% of the time.
The Asian range is, in effect, resting liquidity that the London session comes to take. But a sweep is not a direction: after London takes the Asian high, price closes back below it about as often as it holds above (~49/51). The sweep is where the liquidity is — not where the day is going.
Volatility is not a constant — it exploded
Gold's average daily range is not stationary, and anyone using a fixed stop distance across years would have been badly wrong:
| Year | Avg daily range (pips) |
|---|---|
| 2017 | 12.5 |
| 2020 | 30.9 |
| 2024 | 33.1 |
| 2025 | 60.7 |
| 2026 (partial) | 158.3 |
Volatility also clusters — a high-range day is followed by a range ~1.24× the local norm, a low-range day by ~1.07×. High begets high. Any level, stop, or target has to be scaled to the current volatility regime, not a historical average.
When gold actually moves
Range concentrates in the New York overlap window (13:00–16:00 UTC), where the average hourly range roughly doubles versus the Asian session. The London and overlap windows are where meaningful directional displacement happens; the Asian session is mostly drift and liquidity-building.
What this means
This is characterisation, not a trade trigger, so there's little to protect — the value is in internalising the instrument's nature:
- Gold mean-reverts by default. Trends are the exception (17%) and they're short. Position sizing and expectations should reflect that.
- The Asian range is liquidity, not direction. London will almost certainly sweep it; that tells you where, not which way.
- Volatility is regime-dependent and rising. Fixed pip stops are a mistake; everything scales to current range.
Master one instrument's personality and the setups stop being generic. On gold, that personality is: patient range, brief violent trend, and a London session that comes to take the overnight liquidity.
Studies like this become the filters inside a documented playbook — the research → playbook → backtest → live loop, locked to one instrument at a time, rather than a scanner firing on everything.