← Stage Two ResearchGold Regime Signal
September 11, 2026 · gold at $4,371 (2026-09-11) · trend-led regime read
52/100NO EDGE
0 · trend broken25no edge75trend confirmed · 100
No clear trend. The middle of the range carries no usable edge — and note this model barely beats a bare 200-day trend rule even at its best, so the honest read here is “no signal.”
REGIMEGold is currently tracking real yields inversely (link intact) — context only; the score is trend-led, not rate-driven · strongest scored block: momentum / trend
Beats a 200-day trend rule?
NO
Walk-forward 2007-2026, the composite scores IC +0.07 against forward 3-month gold — versus +0.09 for a bare “price above its 200-day average” rule. It matches the naked rule, it doesn't beat it. This is a trend read with extra context, not an edge over trend.
What the score is made of
Momentum / trend-1% 1m, +4% 3m, +20% 12m, below 200-dayYELLOW · 22/40
Gold trends for months at a time — multi-horizon returns plus price vs its 200-day average. This is the ONLY block that reliably tracks forward gold (walk-forward IC ≈ +0.09), so it now carries the score. Everything else that used to be scored just re-measured this same trend (miners, silver and ETF-flow are all 0.5-0.75 correlated with it), so they were demoted to context to stop triple-counting one signal.
Positioning (COT)managed-money net long 86th percentileRED · 4/10
Speculative crowding from the CFTC report — a contrarian fragility gauge. Scored (lightly) and it also hard-caps the composite when crowding is extreme. Note: making it two-sided (crowded longs + broken trend = a sell) was tested and rejected — only ~6 such months in 19 years and no forward edge, so it stays a cap, not a trigger.
Risk / uncertaintyVIX 36th pct, HY spreads flatYELLOW · 4/7
Gold's safe-haven bid shows up in stress regimes. Scored lightly as a tail amplifier — quiet markets leave it neutral, so it only nudges the score up when VIX or credit is genuinely elevated.
Context — shown, not scored
These used to be scored, but each either re-measured the gold trend the momentum block already captures (miners, silver and ETF-flow are 0.5-0.75 correlated with it) or carried no forward signal (real yields, structural liquidity). Kept for color; they no longer move the number.
Opportunity costreal yield +2.46%, dollar -1.7% 3mo, breakevens 2.40%YELLOW · context
10y real yield, broad dollar, breakevens — gold's contemporaneous macro backdrop. Context only, NOT scored: validation found real yields don't predict gold's 3-month-forward direction (the relationship is same-day, not a lead), and the stationary rate-change has ~0 forward IC. Shown so you can see the backdrop; it does not move the score.
Confirmation (miners, silver)miners +16% vs gold, silver -9% vs goldGREEN · context
Is the broader complex confirming the move? Miners (GDX) lead via operating leverage; silver is the higher-beta cousin. Context only, NOT scored — these are ~0.5-0.6 correlated with the gold trend itself, so scoring them was counting the same move twice. Useful as a divergence read (miners lagging is an early warning), not as independent signal.
ETF flow (GLD)accumulation, +10d net 3moGREEN · context
On-balance volume on GLD — a volume-flow proxy for ETF demand. Context only, NOT scored: OBV is price×volume, so it's ~0.75 correlated with the gold trend (the most redundant of the demoted blocks). Actual GLD tonnage would be cleaner but SSGA blocks it server-side. A price-up-on-weak-flow divergence is still a useful distribution warning.
Structural biasnet liquidity +62bn/4wk, M2 +5% YoYGREEN · context
Slow liquidity backdrop (Howell framework) plus money-supply growth. Context only, NOT scored — it's a generic risk-asset input with ~zero forward gold IC at 1-3 months, kept purely as slow-backdrop color.
How the score is built
A leakage-free walk-forward (2007-26) found the old “seven scored blocks” were mostly one signal in three costumes — momentum, miner/silver confirmation, and ETF flow are all “has the metals complex been going up,” 0.5-0.75 correlated, so half the score was the gold trend counted three times (and the blend was actually worse than momentum alone). So v3 scores only the blocks that carry forward signal: Momentum / trend (40), dominant; Positioning / COT (10), contrarian, which also caps the score when crowding is extreme; and Risk (7), a stress amplifier. Confirmation, ETF flow, opportunity cost (real yields/dollar/breakevens) and structural liquidity are kept as context — shown, but they no longer move the number.
The bands, with Wilson 95% intervals on the 19-year sample: 75+ (TREND CONFIRMED) preceded gains 66% of the time [CI 53-77%, n=53] — a CI that straddles the 64% base rate, so even the top band is not cleanly distinguishable from “gold usually drifts up.” 25 or below (TREND BROKEN) has essentially never fired, and when gold does get that washed out it has tended to bounce — so it's a mean-reversion watch, not a sell. The honest use is the NO EDGE default.
Read this before acting
This is close to unforecastable with free data. At its best the model matches a bare 200-day trend rule and no more, and its one mild edge — the TREND CONFIRMED band — has a win-rate whose confidence interval overlaps the base rate. Treat it as a descriptive regime read, not a forecast. There is deliberately no fire threshold and no scorecard entry here; that would imply a forward claim this doesn't earn.
Why so weak: gold's marginal price-setter for the last few years has been official-sector central-bank buying, which is price-insensitive and reports quarterly with revisions — not something a free daily model can front-run. It reads the regime, not the next tick, and it can't see shocks (a central-bank surprise or liquidation scramble moves gold in a day regardless). Treat a NO EDGE score as exactly that — no signal, not a forecast of calm.