Gold has no directional edge — but its volatility keeps a schedule
We put ~20.5 million ticks of gold (XAUUSD, 365 days, ~1 tick/second) through the same honest machinery that closed the door on Crash/Boom. Three findings: direction is unpredictable at any cost; volatility is strongly structured; and exactly one conditional construct survived rigorous validation — earning the status of candidate, which is not the same word as edge.
1 · Direction is a random walk, and cheaper spreads can't fix that
Return autocorrelation ≈ 0 at every intraday timeframe (|r| < 0.05); variance ratios ≈ 1.0. The classic "fade the big bar" setups, tested with significance rather than vibes:
| Setup (1-bar-hold fade, gross) | n | per trade | t-stat | verdict |
|---|---|---|---|---|
| Fade >2σ 15-minute bar | 952 | +7.1 ± 67.9 pips | +0.10 | noise |
| Fade >2σ 1-hour bar | 229 | −90 ± 228 pips | −0.40 | noise |
| Fade prior day | 150 | −464 ± 696 pips | −0.67 | noise |
A point estimate that "clears the spread" means nothing at |t| ≪ 2. Cheaper execution cannot rescue a directional edge that does not exist.
2 · But volatility is real structure
The autocorrelation of absolute returns is ~0.37 at 5–15 minute scales, decaying slowly — volatility is forecastable even though direction isn't. And it keeps a schedule: the London/New York overlap (12:00–16:00 UTC) runs ~50% more volatile than any other session (8.0 bps average 5-minute move; 509-pip median bar range), with directional bias ≈ 0 in every session. Fat tails everywhere: excess kurtosis 28–45 across timeframes, skewed to the downside even in a year gold rose 26.6%.
3 · The one survivor — and exactly how much we trust it
A session-breakout construct — break of the London range during the overlap, stop at the range midpoint, target 1.5× the range, 15-minute bars — tested over the full year at a 5-pip spread: 311 trades, 46.6% win rate, +0.133R per trade (t = 1.74), +41.1R total, max drawdown 18.3R. Cost-robust: expectancy stays positive even at a 70-pip spread, because stops and targets are range-sized. But the profit is regime-concentrated. Bucketing weeks by directional efficiency (|net move| ÷ total movement):
| Week type | trades | win | share of P&L | per trade | t |
|---|---|---|---|---|---|
| Choppiest third | 126 | 43.7% | +15% | +0.05R | +0.38 |
| Middle third | 93 | 40.9% | +5% | +0.02R | +0.16 |
| Trendiest third | 92 | 56.5% | +81% | +0.36R | +2.65 |
The breakout is a trend-regime tool, not a standalone edge — and the regime persists for days, not weeks. Last week's trendiness predicts nothing about this week (weekly autocorrelation ≈ 0; gating on it collapsed +41R to +6R). A trailing-5-trading-day regime estimate, applied with strict causality (the filter for today computed only from days strictly before today, threshold likewise), kept it: +0.185R per trade (t = 1.70, n = 149), a smooth profile across 3/5/7/10-day horizons (+0.151/+0.185/+0.237/+0.136 — no overfit spike), stable to filter-definition changes, robust from 3-pip to 70-pip spreads, drawdown halved (18.3R → 8.4R).
4 · The honest weakness, in our own words
Walk-forward: in the first half of the year (trending), the filtered breakout earned +0.347R/trade (t = 2.16) vs +0.222R unfiltered; in the second half (trendless), +0.045R — no better than unfiltered. The filter improves selection when a trend regime exists; it cannot manufacture an edge where the underlying breakout has none. It never made any period worse, and halves exposure — a risk filter on a non-stationary underlying, validated on one year that was essentially one macro regime.
That is why this construct is on trial rather than in production: it runs in shadow against the live market, its judgment criteria were frozen before the data, and the two sister-market tests (silver and EURUSD) sharpened exactly which part we doubt. We don't know the answer yet.
Source: gold_findings.md in our research record — ~20.5M ticks, 2025-06-10 → 2026-06-10, market-hours-aware resampling and backtesting, spike gap-through fills, no lookahead; returns never measured across a session break. Context caveat carried on every conclusion: 2025–26 was a strongly trending gold year (+26.6%, ~26% annualized vol) — one macro regime.