The level that remembered nothing

Ask whether a price level remembers what happened there before, and the honest answer from my own data is: it remembers roughly a fortnight, and calling that memory oversells it. What predicts a level’s next behaviour is what happened there most recently. Reach further back, hold the recent outcome fixed, and the distant past adds nothing measurable at all. I call the pattern a Recency Mirage — durable structure that turns out, on inspection, to be short-range recency wearing a costume.

This is a post-mortem rather than a study of somebody else’s idea. The memory hypothesis was my first real finding, I published it internally as the engine’s most promising result, and I spent the following weeks killing it.

What I found, and why it looked right

Split a level’s history into two halves and correlate a pool’s bias across the split. I reported r = +0.31 at p = 0.014, and read it as structure living exactly where the theory said it would: levels carrying persistent identity.

That figure is quoted here as a record of what I said at the time, not as a measurement I would stand on now. It is the claim this post retracts, and the separator below is the receipt for the retraction.

I also recorded, at the time, the reason not to believe it yet. A correlation across a time split is produced just as readily by short-range trend as by durable identity, because the two halves meet at a boundary and trend carries across it. The only thing that separates the two explanations is a revisit with a long gap — far enough back that recent momentum cannot reach. Over the thirteen-month window I had forty-four of those, and the long-gap subset looked better than the short-gap one, at +39% against +28%.

A tiny sample strengthening the preferred reading is the exact shape of a result that should not be trusted. I wrote that down before I ran the test that settled it.

The test that settled it

The full history is sixteen years: 97,693 bars, 200,369 pool arrivals, and enough long-gap revisits to make the separator meaningful. The specification is a single logistic model — predict the outcome at a level from its most recent prior outcome and a genuinely distant prior outcome together, then ask whether the distant term adds anything once the recent one is present.

It does not.

The recent term carries an odds ratio of 1.55. The distant term comes in at 1.01, with a partial likelihood-ratio χ²(1) of 0.39 and p = 0.53, on 183,689 eligible arrivals. Adding the distant outcome changes held-out log-loss by 0.0000. The entire split is decided by what happened at that level last time.

The decay curve is the clearest way to see it. Measured by the gap since the previous visit, the effect runs about eleven per cent at zero to fifty bars, five per cent from fifty to two hundred and forty, three per cent out to a thousand, and then zero. Beyond roughly a thousand bars — about two months — there is nothing left to measure.

The part that makes the null trustworthy

A null result is only as good as its ability to have come out the other way, so one detail matters more than the headline. Running the walk over the full history builds a much denser ladder of levels than the shorter window did, because levels never expire. Density inflates trend, and inflated trend would tend to manufacture a spurious distant effect rather than suppress one.

The test was therefore biased against the null it returned, and it returned the null anyway. That is the direction of error you want when you are trying to kill your own hypothesis, and it is why I treat this one as settled rather than provisional.

What survives

Something does. A level’s most recent outcome genuinely predicts its next one — the probability of absorption given that the level last absorbed is 61%, against 33% when it last swept, an odds ratio of 3.49 that holds on held-out data. That is a real, out-of-sample-valid effect and I am not throwing it away.

I am also not dressing it up. “Recent behaviour persists for a few weeks” is a general property of price series. It is not evidence for anything about liquidity, institutional positioning, or levels as objects, and it was available to anyone before I measured it. The finding here is not that I discovered momentum. It is that momentum accounts for all of what I had mistaken for memory.

Limits, printed rather than buried

The test covers one instrument. A single-pair null is a null about that pair, and I have not tested whether it generalises.

It is also a one-dimensional test, which matters more than the pair count. The broader thesis this work sits inside claims a conjunction — crowding and dealer positioning and thin liquidity together — and a signal that only fires at the intersection of three conditions would return null on each one measured alone, by construction. That conjunction has not been run, and two of its inputs are not available to an individual at any price. So this result kills a specific, popular, one-variable version of the idea. It does not settle the general question, and I would be overselling a null if I said otherwise.

Why I published it

The claim that levels remember is taught as fact, and it is the load-bearing assumption under a large amount of retail trading education. I believed it, measured it properly, and found that my own best evidence for it was a forty-four-sample artefact that dissolved at scale.

A firm that only publishes results flattering to its thesis is not running an instrument, it is running a marketing department. This one is in the register with its kill criteria attached, counted against the full denominator of every hypothesis I have registered — including, and especially, the ones that died.