Does COT positioning predict reversals?

No. On my data the relationship runs the other way, and the interesting part is that it runs the other way significantly.

The AUD leg returned a trigger effect of −0.0477, bootstrap interval [−0.0801, −0.0137], p = 0.0066. The interval excludes zero. The sign says crowded positioning extremes were followed by liquidation events at 3.1% against the mid-range controls — less often, not more.

I call that shape Wrong-Way Significance: an effect that clears every bar a reader checks and points at the opposite of the hypothesis it was registered to test.

Why this is the first thing I publish with its data attached

Every measurement on this site is held to three standards: the methodology is pre-registered before the first tick, the artifacts are content-hashed, and the results are recomputable. This is the first page where all three are true at once, which is why it goes first rather than something with a better headline.

Pre-registered, and the ordering is provable rather than asserted. The registration records that at the moment it was written the positioning data had never been contacted — the directory did not exist, and the single prior fetch attempt had returned an HTTP 503 and written nothing. The fetch log agrees: every attempt on that date failed. So the hypothesis, its thresholds, its horizon grid and its decision tree were fixed at a point when looking was not possible.

That matters more than it sounds. A pre-registration is worth exactly the provability of its ordering, and most are worth nothing on that test.

Content-hashed. The dataset below carries nineteen input hashes, one per source CSV, and every one of those CSVs is committed. The compute step opens no socket.

Recomputable — and this was measured rather than assumed. The claim most sites would make here is “you could re-derive this”. I re-ran the driver into a clean temporary directory and compared the output field by field against the published artifact: fourteen of fourteen result keys byte-identical, none differing. So you can download the file, check its hash against the one published here, and re-derive it from the committed inputs — and I have already done that myself and reported what came back. Nothing in the chain asks you to trust me.

What was actually measured

Seven currency books, 201 aligned weekly release stamps, latest 2026-06-30, drawn from the CFTC’s Traders-in-Financial-Futures report. Leveraged-fund net positioning is joined to each leg’s weekly price and keyed to the Friday release stamp, giving 821 release-keyed weekly observations per leg after the 2010 price join.

The frozen defaults: a 156-week percentile window, a 2× range threshold, deciles at 0.9 and 0.1, horizons of 2, 4 and 6 weeks, a 1,000-iteration block-bootstrap surrogate, four folds.

The surrogate deserves its own line, because it is where most tests of this kind are weakest. It resamples the weekly changes in a stationary block bootstrap, relabels the extremes on the surrogate while holding the real outcomes fixed, and re-runs the test. An independent-and-identically-distributed shuffle would have been easier and would have been anti-conservative — positioning is autocorrelated, and destroying that structure inflates significance. I had learned that on an earlier hypothesis and wrote the harder surrogate into this registration before running it.

The result, leg by leg

AUD: not live. Effect −0.0477, interval [−0.0801, −0.0137], p = 0.0066, and it fails the relabel surrogate. Windowing is durable. The NZD leg was registered as a uniformity control rather than a target, and the AUD effect survives it — the difference between the two legs is −0.092, so what is being seen is AUD-specific rather than generic antipodean risk flow.

JPY: a preliminary unconditioned null. Effect +0.0442, interval [−0.0152, +0.1131], p = 0.18. This is the unconditioned result, and the registered verdict for JPY requires an intervention-and-rate regime split whose feed does not exist yet. So it is reported as preliminary and not as a finding, which is the state it is actually in.

Verdict, from the tree written in advance: the single-leg positioning fuel gauge does not carry the crowding-to-liquidation signal. That was the outcome the registration declared it expected, which is the least exciting way for a pre-registered test to resolve and the most credible.

What this does not establish

The null gates rather than kills. The registered thesis is a conjunction — crowding and dealer positioning and thin liquidity — and this measures the first term alone. A null here means Layer 1 does not on its own justify the spend on Layers 2 and 3. It does not mean the conjunction is refuted, and I would have had to say that either way, because the tree deciding it was fixed first.

The JPY leg is preliminary for the reason above. The AUD result is one currency on one weekly cadence. And positioning data is a weekly position snapshot, not a price series — the artifact says so on its own face, in the banner it ships with.

The download

The measurement is the file, not this page. Its hash is published beside it; if your copy hashes differently, one of me has a problem worth knowing about.

If you recompute it and disagree with me, send the recomputation. A correction with a receipt attached gets published as a correction, with the original left legible beside it.