Positioning tells you what regime you are in. Your own system tells you when to trade. The research below marks where those two things line up — and publishes the hit rate for every claim it makes, next to the rate you would have got for free.
Almost every COT indicator on the market tells you a crowded reading means reversal risk. We tested it across 12 markets and 26 years: price made a 4-week low within 2 reports 46.7% of the time, against a 45.2% rate it would have managed anyway. A +1.5 point edge on 8 of 12 markets — indistinguishable from nothing.
It is worse than merely useless. Bucketing forward returns by index level, the 80–90 bucket is mildly positive (+0.84% excess at 13 weeks, the best of the ten). There is no sign flip at 80. Shading that zone red points you the wrong way, so COTVault does not do it.
This matters beyond one indicator. "Fade the crowd" is the single most repeated idea in COT education, and on this data it does not survive contact with a base rate. Everything in the rest of this guide is what was left after the same testing was applied to every other claim.
Once positioning crosses into the extreme, it stays there far longer than most traders expect.
| Still crowded (≥80) after | Accuracy | Free rate | Lift | Markets |
|---|---|---|---|---|
| 2 reports | 65.7% | 37.2% | +28.5 | 12/12 |
| 4 reports | 55.8% | 34.9% | +20.8 | 12/12 |
| 8 reports | 42.0% | 32.4% | +9.5 | 9/12 |
Softer reading — still elevated (≥70): 85.3% / 71.8% / 57.3% at 2/4/8 reports against free rates of 72.7 / 57.9 / 48.1 (n=443 trigger events). Median episode 7 reports; 47% run 8 or longer — measured over 320 crowded-long episodes, entry at 80 and exit below 70.
The free rate here is drawn from the adjacent 76–80 band — four points below the trigger. That matters: a wider 65–80 control would have shown a +31.1 lift, but part of that gap is simply "starts higher, stays higher." Against the adjacent band the extreme itself is still worth +20.8, and all twelve markets agree.
Do not fade a fresh extreme. Excess return from fading is negative in 16 of the 18 sub-8 age/horizon cells measured, both directions — the two exceptions are +0.01% and +0.06%, indistinguishable from zero with bootstrap CIs straddling it. A crowded reading is a mature trend that is still running, not a trend about to end. Positioning unwinds after price turns, not before.
This is the part that puts an edge on the chart. Teal shading marks where two independent things agree:
The measurement is a barrier race, not an event count: from the zone bar, does price reach +1 ATR before −1 ATR within 13 reports? That is a payoff question. "Makes a 4-week high" is not — a 4-week high can be +0.05% and can arrive after a brutal drawdown. A barrier race cannot hide that.
| Accuracy | Free rate | Lift | Markets | n | |
|---|---|---|---|---|---|
| Watch zone, ±1 ATR | 60.2% | 52.8% | +7.4 | 8/12 | 362 |
| Same zone, ±1.5 ATR | — | — | +9.8 | — | — |
| Same zone, ±2 ATR | — | — | +12.8 | — | — |
The edge grows as the barrier widens. That is what a real effect does — a curve-fit artefact degrades when you change the measurement. Walk-forward: +14.0 fitted before 2010, +4.6 out-of-sample 2010–2026. The out-of-sample number is a third of the in-sample one, which is honest decay, not a failure — it stays positive and the bootstrap CI over markets excludes zero at [+2.4, +13.8].
The zone appears roughly 0.5 times per market per year — 162 distinct openings across 313 market-years, 2.3% of all weeks. It is rare on purpose. A zone that paints half the time is not a location, it is wallpaper.
Crowding does exhaust. Once the marginal buyer is gone the position has to come off eventually. The question is only how long "eventually" takes — and the answer is far longer than instinct suggests.
Crowded long at resistance, asking whether price falls 1 ATR before it rises 1 ATR: 42.7% against a 47.1% base rate. A lift of −4.4 on 4 of 11 markets. That is the mirror toggle, off by default — a crowded long typically squeezes up before it breaks, so a symmetric barrier test cannot see the effect and an 8-week horizon fails too.
Red shading marks specs crowded long for 16 or more reports. Shorting that state:
| Horizon | Excess | Hit | Markets | In → out of sample |
|---|---|---|---|---|
| 13 weeks | +2.35% | 61.6% | 10/10 | 6.91 → 1.10 |
| 26 weeks | +5.22% | 69.7% | 8/10 | 8.80 → 4.48 |
| 13 wks + stoch overbought | +3.23% | 69.7% | 7/7 | 7.76 → 1.13 |
Excess return rises monotonically with episode age — wk0 −0.48%, wk4 −0.14%, wk8 +0.75%, wk12 +1.64%, wk16 +2.35%. A smooth ramp like that is the signature of a real effect; a fitted one spikes at one setting and collapses either side of it.
Treat these as two different trades. The teal watch zone is a swing location. The red mature zone is a quarterly position read — at 8 weeks it fails outright, and you should expect an adverse excursion before it resolves. Size for that, or don't take it.
The asymmetry in speed has a mechanism. A crowded short is forced — margin and unlimited loss compel covering, so it resolves quickly. A crowded long can sit indefinitely; nothing makes those holders act until the trend itself rolls. Same exhaustion, very different clock.
60.2% is a barrier-race rate, not a profit figure. It says price reached +1 ATR before −1 ATR more often than baseline. It says nothing about your entry, your stop, your size or your costs. A 60% barrier race with poor execution still loses money.
12 markets — EUR, GBP, JPY, CHF, CAD, AUD, NZD, US Dollar Index, Gold, Silver, WTI Crude, Corn — using CFTC Legacy futures-only data joined to weekly OHLC. The CFTC series itself reaches back to 1986, but the measured sample is bounded by weekly price history: 2000–2026 for ten markets, 1992 for the US Dollar Index, 2003 for NZD — 16,284 market-weeks in total. This is the same dataset and the same report series the COTVault dashboard runs on.
Every claim is reported as accuracy against a base rate, never as accuracy alone. Persistence claims use a matched adjacent-band control, because an all-bars control is meaningless for a state-transition claim — most bars already sit in the resolved state. Price claims use an all-bars control, which is valid. Any claim whose free rate exceeded 85% was rejected outright regardless of accuracy: an indicator that "predicts" a near-certain event has taught you nothing.
Guards applied to the zone: per-market breadth, a market-resampled bootstrap CI, replication at 1.5 and 2 ATR barriers, and a walk-forward split at 2010. The number of combinations tested is published above so you can discount for it yourself.