Graveyard / Positioning
COT positioning on currencies and commodities
Do extreme futures positions in FX and commodities predict reversals?
The idea
When leveraged funds or managed money are unusually long or short a currency or commodity, the crowd is stretched. Trade against it (contrarian) or with it (follow).
How we tested it
| Data | Seven USD currency pairs with leveraged-fund positions, and five commodity ETFs (GLD, SLV, CPER, USO, UNG) with managed-money positions. Currency entries at a z-score of 1.5 and beyond (the commodity threshold is not stated in the note). Both directions were specified before the run. 104 weeks are needed for the z-score, so the development window is effectively 2012-2018. |
|---|---|
| Period | CFTC data from 2010. Holdout January 2019 to November 2026. |
| Costs | Costs and financing or borrow charged, plus a 1.5 times stress case. Currency financing used a 3-month rate differential plus 50 bp a year. |
| Holdout | January 2019 onward, untouched until the single run. |
| Control | Random twins: the signal series shifted circularly by 1 to 8 years, 300 times. Bonferroni alpha 0.025 for two variants. |
Result
| Currencies, holdout Sharpe contrarian / follow | 0.04 / −0.23 |
|---|---|
| Currencies, annual net | +0.07% / −0.43% |
| Currencies, twin p | 0.42 / 0.57 |
| Currencies, holdout entries | 63 |
| Commodities, holdout Sharpe contrarian / follow | 0.06 / −0.13 |
| Commodities, annual net | +0.46% / −1.1% |
| Commodities, twin p | 0.63 / 0.48 |
| Commodities, holdout entries / max drawdown | 37 / −21% |
Why it failed
Neither direction beats the random twins. Currency positioning gave nothing before or after 2019 (development Sharpe −0.06 and −0.07). The commodity ETFs carry their roll cost inside the price, so positioning cannot be separated from roll yield with free data.
What would make it worth retesting
Futures-level data with the term structure, so roll yield can be separated from positioning. For currencies, a second positioning source, or a signal built on changes rather than levels.
Test id: fxcot-fx7-event-v1, mmcot-cmdty5-event-v1. Past results of a backtest say nothing certain about the future.