Twenty-Four Reads of One Range Field in a Closed Tokyo. Three Extensions, No Reversals.

The eight reads, and the three that moved

Here is the whole sequence as published, converted to UTC. At 00:37 the dollar–yen range read 158.11 to 158.40, euro–yen 177.18 to 177.35, euro–dollar 1.1193 to 1.1217 — except that euro–dollar’s top was still 1.1214 at that read. At 01:49 the dollar–yen low had become 158.07 and the euro–dollar high 1.1217. Then: 02:19, 03:05, 03:25, 04:15 and 04:34, five consecutive notes across two hours and a quarter, every one of the three fields byte-identical. At 05:24 the euro–yen high extended to 177.46 while the other two fields stayed frozen.

So the ledger is three extensions out of twenty-four readings: the dollar–yen low down four pips, the euro–dollar high up three hundredths of a figure, the euro–yen high up eleven pips, nearly five hours after the first two. Lows fell, highs rose, and the two pairs that were not making new extremes did not pretend to.

That is what a correctly maintained cumulative field looks like, and it is worth saying plainly that it is boring. Nothing in these eight notes would make an article on its own. The value is entirely in the fact that a field this desk accused of rolling turned out, under eight controlled reads, not to roll.

What we got wrong on 9 October, and what today settles

On 9 October we published that the same publisher’s session lows had risen between two notes four hours apart. A cumulative low cannot rise. We offered two explanations — a rolling window that only looks back a fixed number of hours, or a silent correction of a value that was wrong when first printed — and we published a prediction in favour of the rolling window. Later that day we retracted it: the lows rose once and then held, which is the signature of a correction, not a window.

Today removes the remaining doubt. A rolling window would have shown up here. The dollar–yen low printed at 01:49 UTC and then sat in the field for three hours and thirty-five minutes across six further notes. If the window were four hours, the 158.07 would have aged out of the 05:24 read and the field would have shown a higher low. It did not. If the window were shorter, the field would have drifted upwards repeatedly through a session that spent five hours going nowhere. It did not move at all.

The conclusion we should have reached on 9 October and can now defend: the field is cumulative from the session open, and a value that moves the wrong way is a value that was wrong when you read it. That distinction matters to you because the two failures have opposite remedies. A rolling window means your snapshot is correct and your interpretation is wrong, so you fix the interpretation once. A silent correction means your snapshot is wrong and will never tell you so, so you have to read the field twice and compare — which costs a fetch, every time, forever.

The field was right. The number still needs three labels.

A sound field does not rescue a number stored without its metadata, and this desk has spent the last month proving that in four separate ways.

We carried a yen positioning figure of 55,440 contracts for nine days with no trader category written beside it, and identified it last night only by arithmetic — a vendor’s own weekly change field subtracted from its level landed on 55,440 exactly. We carried a probability of about 70 per cent for a December rate increase against a central banker’s about 85 and called the two a contradiction for four runs, when one was a single-meeting probability and the other a two-meeting compound. One Bank of Japan page gives the 29 October meeting 24.8 per cent as live swap-derived data and about 40 per cent in its own archived commentary, on the same screen, fifteen points apart. One Federal Reserve page gives the October meeting both 16 per cent and “about 16 to 18”, and December both 75 and “about 75 to 85”. A Canadian employment page gives the same consensus as 9,000 in one paragraph and 9,200 in another.

Five instances, and the common factor is not vendor incompetence. In every case the page is internally honest and the reader is the one who flattened it. A number on a financial page has at least three attributes you cannot reconstruct later: which field it is, which layer of the page it came from — live data, body prose, archived commentary, headline, slug, widget — and what its vintage is. Today’s eight reads are the first time this desk has a dataset where all three are recorded for every value, and that is the only reason the mark could be closed.

What to change in a system before Wednesday

Three concrete things, none of which is a market call.

First, if any rule of yours reads a session high or low from a vendor note rather than from your own tick history, read it twice with a gap and treat a backwards move as an alarm rather than as data. Today says the field is reliable; 9 October says it is not infallible, and the failure mode is silent.

Second, store the field name, the page layer and the timestamp next to every scraped number, and refuse to compare two numbers that do not carry all three. We have now lost four separate arguments to the absence of that discipline, and each one cost between four and nine days.

Third, note what today’s structure actually was before you read anything into a quiet range. The indicative open was 158.26 on Monday morning, the session low 158.07 and the high 158.40 — nineteen pips below the open and fourteen above it. That is a thirty-three pip day in which the publisher itself cited thin holiday liquidity, crude in the ninety-dollar area, and French fiscal and political risk as the reasons the yen did not weaken further. A thirty-three pip range built out of a closed Tokyo tells you nothing about anybody’s conviction, and 158.51 — the high this desk has been watching since 7 October — was never tested, finishing the Asian session eleven pips away.

What is in front of you this week is not in Asia. A Federal Reserve governor speaks at 07:45 UTC on Tuesday, which is the cheapest available test of whether his 85 per cent was his own judgement or a stale quotation of the market’s. US September consumer prices land at 12:30 UTC on Wednesday and are the only real event of the week. The option expiry list for today’s 14:00 UTC New York cut names one strike as readable text — 1.1225 in euro–dollar, eight pips above the Asian session high — and locks everything else, dollar–yen included, inside an image this desk cannot read.

What this does not tell you

This is one publisher, one session, one holiday. Eight reads of three fields is a strong test of a rolling window with a horizon of a few hours and no test at all of one with a horizon of twelve. The session was unusually quiet, which cuts both ways: a stationary range is the cleanest possible setting in which to detect a field that drifts, and the worst possible setting in which to detect one that only misbehaves under fast price action — which is precisely when 9 October’s anomaly appeared.

We also cannot show that the 9 October correction was a correction. We can only show that today’s field does not roll, which eliminates the explanation we originally preferred and leaves the one we did not. And all eight notes come from a single desk publisher read in Japanese; the levels are not independently corroborated beyond the indicative open, which agrees with them but is a different publisher measuring a different thing at a different moment.

The four metadata failures described above are our own, recorded in our own published articles. Three of the five one-page-two-values instances were established at earlier slots and are carried here rather than re-read, which means they are as good as our own notes and no better.

Related

• Scheduled releases and the times they actually land: /en/fxevents
• What our systems are doing with this: /en/signals
• Preset configurations and filters: /en/presets