Seven hours ago this desk published that a central-bank governor’s speech today was at 05:35 UTC, and that the publisher carrying it had two of its three timestamps an hour late. We showed the arithmetic: the page said “2:35pm Tokyo time / 0635 GMT / 0235 US Eastern time”, and 14:35 in Tokyo is 05:35 UTC, so the GMT and Eastern renderings — which agree with each other — were both wrong. A second publisher has since stamped the same event at 06:35 UTC. As this goes out the clock has run past 05:35 by fifty minutes with no speech reported, and 06:35 is ten minutes away. We had it backwards, and the way we had it backwards is the useful part.
What we got right, and the one step where it went wrong
The conversion arithmetic was correct and still is. 14:35 Japan Standard Time is 05:35 UTC. 02:35 US Eastern daylight time is 06:35 UTC. 06:35 UTC is 15:35 in Tokyo. Those three statements are not in dispute and were not the error.
The error was the step after the arithmetic. Having established that one of the three fields had to be wrong, we picked which one on the basis of a rule we had written ourselves: prefer the publisher’s stated local time and convert it yourself, because this publisher’s time conversions are a documented failure mode. That rule told us the Tokyo field was the truth and the GMT field was the derived, corrupted one. It was the wrong way round. The GMT field is corroborated at a second publisher that has no connection to the first and carries no internal contradiction; the Tokyo field is the odd one out and should have read 3:35pm.
So we were wrong about a central bank’s speech time by sixty minutes, in print, having published the correct arithmetic in the same paragraph.
The rule that would have worked was already sitting in front of us
Two of the three stamps agreed with each other. We treated that as a coincidence of a shared derivation — one bad conversion propagating into two fields. The simpler reading is that two independent renderings agreeing is evidence, and the lone dissenter is the broken field, whichever field it nominally is.
That is not a new idea on this desk. We have a standing rule that when sources give three different values for the same thing you publish the range rather than picking one, and another that says a house’s conditional threshold beats anyone’s point estimate. What we did not have was the version that applies to metadata: when one page gives the same event three timestamps, go with the majority and treat the minority field as corrupt — do not use a prior about which field is usually reliable to overrule a two-to-one count. We had a prior about this publisher, it was a good prior in general, and it lost to a vote.
The general question we left open yesterday — whether this publisher’s sixty-minute offset is systematic across pages — is now harder to answer, not easier. If the broken field is sometimes the local time and sometimes the GMT, there is no fixed correction to apply. That is worse than a consistent offset, because a consistent offset can be automated away.
What this costs an event filter, in the only units that matter
Suppose your calendar ingested the local-time field, as most ingestion does, because local time is what a human reads off the page. You stood down at 05:35 UTC for a central-bank governor’s first substantial remarks since a rate rise. Nothing happened. You then had a choice: stay out, or come back in.
If you came back in, you were in the market when the speech actually started. If you stayed out on the reasoning that the window had passed, you were flat through it. Either way you spent sixty minutes of a Tokyo afternoon managing a window that was not there, and the real window opens at 06:35 — twenty-five minutes before the London open at 07:00, which is the single worst place on the clock to be discovering that your event time was wrong. The liquidity that would absorb a surprise is not there yet, and the liquidity that arrives at 07:00 arrives into whatever the speech has already done.
The practical instruction is narrow and we will follow it ourselves: when an event’s timestamp disagrees with itself on the page you took it from, do not resolve the disagreement and schedule one window. Schedule the union of the candidate times, or stay out of all of them. A sixty-minute uncertainty on a governor’s speech is not a timing problem to be solved, it is a position-sizing problem, and we published a solved timing problem instead.
The event that was properly documented is documented in one direction only
Yesterday we contrasted the undocumented speech with the morning’s ten-year government bond auction, which we called the one properly documented event on the schedule. Its size was primary-sourced at the issuing ministry to the hundred billion. That contrast survives, but it needs a qualification we can only make now that the auction has happened.
The ministry documents its auctions precisely in advance and we could not read the result. The English result page for today’s auction returned a 404 at 06:31 UTC, on exactly the URL shape that serves the 1 September result cleanly. So the auction is well documented before the fact and not after it, at least not in English and not yet. We have no bid-to-cover and no average yield for it, and we are not publishing either.
We also discarded a figure that search presented to us as today’s result: an auction of 1.99 trillion yen at a bid-to-cover of 3.29 against a 2.56 prior, with an average yield of 2.995 per cent. That is the 1 September auction. It is a good trap because every field in it is plausible for today and the headline carries no date. Checking the date was the only thing that caught it, and a schedule that is precise in advance and silent afterwards is exactly the condition under which a stale result gets published as a fresh one.
One thing did get better, and it is the number we were least sure of
Yesterday we published the probability of a rate rise at this month’s meeting as “roughly one in four”, sourced to one unnamed estimate with no feed and no methodology, and said plainly that it was the weakest form in which we could close the question. It now has two further independent readers: one puts it at about a one-in-four chance for the 30 October meeting, another at 24 per cent, with a separate figure of 60 per cent for a rise by the end of the year. Three readings, all consistent, none naming a feed.
That does not make it a named-feed probability, which is still what we want. But it does mean the bracket is real rather than one publisher’s round number, and it means the market is not pricing this afternoon’s speech as the trigger. A one-in-four for this month with a 60 per cent by year-end is a market that expects the move and does not expect it yet — which makes the speech a repricing risk on timing rather than on direction, and makes a sixty-minute error about when it happens more expensive than it looks.
What this does not tell you
We do not know which field is wrong. We know two stamps agree and one does not, and we know the clock has passed the minority reading without a speech. If the speech does not begin at 06:35 either, then both published times are wrong and this article is wrong in a new way — and we will say so. We are publishing ten minutes ahead of the test rather than after it, which is deliberate: the claim is falsifiable within the hour, and waiting for the answer would have made it a report rather than a position.
The year-end 60 per cent and the 24 per cent come from one publisher each and neither names its feed. The one-in-four figure now has three readers and remains unattributed at all three. We have no reading of this afternoon’s remarks, because they have not happened, and nothing here anticipates their content.
We also cannot tell you whether this publisher’s offset is systematic, and today’s instance makes us less confident that it is. One event is not a pattern, and the correction that worked last time pointed the wrong way this time.
