There is no CFTC contract for GBPAUD, EURJPY, or any other FX cross. The Commitments of Traders report covers futures, and no cross trades as a future. So this builds one — and then tells you exactly what it is and isn't worth.
Each USD leg is indexed 0–100 against its own three-year range of weekly reports — the classic COT Index. The cross score is then simply:
score = base leg − quote leg
→ −100 … +100
GBPAUD with GBP at 82 and AUD at 18 scores +64: speculators are heavily long the pound and heavily short the Aussie, so the crowd is leaning hard toward the base currency of that cross.
Base 82 / quote 18 and base 96 / quote 32 both score +64, and they are not the same market. One is two crowded legs pulling together; the other is one leg pinned at the very top of its three-year range against a counterpart sitting mid-range. The score alone cannot tell them apart — the rail can. Read it before you read the score.
You do not have to measure the two lines by eye. The panel's Both legs row says it in words: "both legs stretched — pulling together", or "EUR is doing the work — AUD is mid-range", or "neither leg is stretched".
Measured across 21 crosses built from 7 currencies, weekly, 2000–2026.
| Score crosses… | Still there after | Accuracy | Free rate | Lift | Crosses agreeing |
|---|---|---|---|---|---|
| +60 (base favoured) | 2 reports | 55.5% | 24.6% | +30.9 | 20/20 |
| −60 (quote favoured) | 2 reports | 60.1% | 30.2% | +29.9 | 18/18 |
| +60 | 4 reports | 39.9% | 24.0% | +15.9 | 19/20 |
| −60 | 4 reports | 45.5% | 31.1% | +14.3 | 17/18 |
n=308 each side. The free rate comes from the adjacent 50–60 band — ten points below the trigger — so "high things stay high" is already inside the baseline and is not being claimed as insight.
Every figure here is published next to the rate you'd have got by chance. A hit rate without a baseline means nothing: "still at an extreme within 8 weeks" would score near 90% on any series that moves slowly, and would tell you exactly nothing.
Eighteen directional setups were tested: follow the lean, fade the lean, and trade the unwind when the score crosses back inside the band — each at 4, 13 and 26 weeks.
None survived. Against roughly 0.9 expected to clear by chance alone. No excess return exceeded 1.3% in either direction, breadth was a coin flip (6 to 16 of the 20–21 crosses with enough signals to measure), and in-sample and out-of-sample signs flipped.
That is why the colours name a currency and never a direction, why there is no buy or sell marker anywhere in it, and why the verdict line ends with "no directional claim" whenever it flags a lean.
A tool that told you a crowded cross was about to turn would be asserting something 21 crosses and 26 years of data say is not there. Plenty of COT tools make exactly that claim. This one won't.
Positioning tells you what regime you are in: whether one side of a cross is crowded, how stretched it is, and — now measured — that the state will most likely still be there in a fortnight.
What it does not tell you is where to act. That needs a price-location layer: support and resistance structure, envelope bands, pivots. There is none in this script.
Used properly, this is a context filter. It tells you which crosses have a stretched crowd behind them, so you know which charts deserve attention — and then your own system decides everything else.
21 crosses built from all pairings of EUR, GBP, JPY, CHF, CAD, AUD and NZD. Each leg's COT index is the
same 156-week min-max range position COTVault computes live, so the study and the dashboard cannot drift
apart. Cross prices for the directional tests were built from the USD-quoted futures as XUSD /
YUSD, close-based only — the high of a ratio is not the ratio of the highs, and a barrier test on
approximated highs and lows would be measuring the approximation rather than the market.
Persistence uses a matched adjacent-band control. Directional tests used excess return versus each cross's own unconditional drift, with per-cross breadth, a bootstrap CI resampled over crosses, and a walk-forward split at 2010. The count of setups tested is published above so you can discount for multiple comparisons yourself — which, given none of them survived, is academic.