About
Every trading session, this site publishes one auction read before the cash open: what the market has done, what it is trying to do, how good a job it is doing, and what it is likely to do next.
An ongoing experiment in whether a disciplined analytical framework, applied to real market data without discretion or mood, produces market context worth having.
The reads follow Auction Market Theory: markets exist to facilitate trade, they advertise price to find activity, and the useful question is never “what will it do” but who currently holds territory, and who is offside. Each session gets read at four zoom levels — monthly, weekly, daily, intraday — because a level only means something relative to the timeframe that built it.
The framework does not change between sessions. Same four questions, same order, whether the read is interesting or dull — and when it is revised, that is done deliberately and never in reaction to a single bad read. That constancy is the point: it makes a run of reads comparable, and it makes the experiment falsifiable.
Every figure is computed, not estimated. Value areas, points of control, order-flow delta, single prints, initial balance, range extension — all computed from the market’s own trade record rather than recalled or approximated. Prices are back-adjusted continuous, so windows spanning a contract roll stay comparable.
The synthesis is machine-written. A language model composes the narrative from those computed inputs, working through a fixed prompt framework. The framework is human-designed and each read is human-reviewed before publication, but the prose is not hand-written and it would be dishonest to imply otherwise.
Occasionally a read is published late. The job runs unattended at 08:05 ET and sometimes it breaks. When that happens the read is still written, but from data cut at 08:05 — the same figures the on-time run would have had, and nothing from the session already underway. Those reads carry a notice at the top saying when they were written, and they are graded against the full session like any other. The alternative is a silent gap in the archive on exactly the days something went wrong, which is the least honest option available.
This is disclosed prominently because it is the interesting part, not a caveat to bury. The claim being tested is narrow and specific: not that a model has insight into markets, but that a rigid framework over real data beats improvisation — and that the discipline of answering the same questions every session is worth more than the cleverness of any individual answer.
A read is not a prediction. It is a set of scenarios — a lean, and the branches that would mean the lean is wrong — each with the price that makes it live. It is graded on whether the set captured the session and whether a reader could tell early which branch was running.
So a read can lean the wrong way and still be a good read. On one session the lean was short, never triggered, and the alternate — a reclaim that would leave sellers trapped — is what happened, on the level the read had named that morning. Scoring only the lean would have called that a miss.
Coverage bought by vagueness is not coverage. Branches have to be separable: two triggers within a quarter of an average day’s range describe the same world twice. Breadth is graded down, not rewarded.
A level breaks when price closes beyond it on two consecutive one-minute bars — a wick through and back confirms a level rather than breaking it. That rule was fixed before the first read was published and does not move. Distances are judged in average true range, so a two-point miss on a fifty-point day is noise, not a failed call.
Grades come from a different model than the one that wrote the read, and every read keeps its grade — including the bad ones, which stay up unedited.
It is not advice, and it is not a signal service. There are no entries, no exits, no targets to act on, and no track record being advertised — because no trades are being published.
Mechanical entry triggers built on order flow do not survive honest out-of-sample testing. The value of a read like this is subtractive: which levels matter, which side is already committed, and when the honest answer is to stand down. It is climate, not a forecast.
Where a read is weak, it says so. Where the evidence cuts both ways, it says that too. A read that only ever sounds confident is not being written carefully.
Nobody’s name is on them, and that is deliberate. A read should stand on its numbers and its method — if it is only worth reading because of who wrote it, the experiment has already failed. Publishing anonymously keeps the claim where it belongs.
So accountability lives in the method instead. The framework is described rather than hidden, and the archive is the only track record on offer — every read, dated, in order, right or wrong.
It is early. The immediate goal is a body of reads long enough to judge honestly — whether the framework holds up across regimes, and whether a machine-written read stays useful when the market stops being obvious.
If it does hold up, the framework itself is the more interesting artifact than any single read, and there will likely be something to offer beyond the free archive. Nothing behind a paywall today, and everything published so far stays free.
The best way to follow along is the RSS feed, or just check in before the open.
Published pre-open, one per session. The archive has everything, oldest reads included.
Read the latestAuction Reads are published pre-open. Same structure every time: what the auction did, what it is trying to do, how well it is doing it, and what it is likely to do next.
Read the archive