On Friday morning we published four American consensus figures and said the day’s releases were the only thing on the board: industrial production +0.3% month on month, capacity utilisation 76.4%, manufacturing output +0.3%, and the Conference Board’s leading index +0.1%. The prints were 0.0%, 76.3%, −0.3% and −0.1%. Four for four missed, and two of the four had the sign wrong. Both of the issuing institutions settle their own numbers in one fetch each, so there is nothing to argue about this time — which makes it a better test of what a consensus is worth.
The scoreboard, settled at the issuer
The Federal Reserve’s G.17 release of 18 September says industrial production was unchanged in August after increasing 0.2% in July, that capacity utilisation was unchanged at 76.3%, that manufacturing output fell 0.3%, that mining ticked up 0.1% and that utilities rose 1.8%. The Conference Board says its leading index decreased 0.1% in August to 99.5 on a 2016 base, after a 0.2% increase in July, while the coincident index rose 0.1% to 114.9 and the lagging index rose 0.2% to 120.6.
Against the consensus we published:
- Industrial production: expected +0.3%, printed 0.0%. Missed by 30 basis points.
- Capacity utilisation: expected 76.4%, printed 76.3%. Missed by a tenth — and note that it did not fall, it simply failed to rise.
- Manufacturing output: expected +0.3%, printed −0.3%. Missed by 60 basis points and went the other way.
- Leading index: expected +0.1%, printed −0.1%. Missed by 20 basis points and went the other way.
A second reader gives the same three actual-versus-expected pairs for the Fed release and timestamps it 13:15 GMT, which is the time the Federal Reserve’s own release page gives. That is a small thing and it closes something: our Friday piece argued that one widely-read calendar had every US time an hour early. An independent timestamp on the release itself lands on the institutional time, not the calendar’s. The calendar was wrong, not the institutions.
The manufacturing line is the one that matters
Three of these four are soft-ish misses. The manufacturing number is not. A consensus of +0.3% against a print of −0.3% is a six-tenths error on a series that moves in tenths, and it is the component that carries the industrial story — the headline was rescued by a 1.8% jump in utilities, which is weather, not demand. Strip utilities out and August was a down month for American factories.
The same reader puts manufacturing capacity utilisation at 75.7%, down three-tenths, and 2.5 points below its average since 1972. We have not settled that one at the Fed and we are flagging it as one reader’s figure.
The operational point is not the direction. It is that if you size a window on the assumption that a consensus is a centred estimate, this release just showed you it can be a coin flip on sign for the component that carries the series. A filter keyed to “beat or miss” would have got the headline roughly right and the real economy exactly backwards.
We asked whether the leading index on the cut did anything. Here is the honest answer.
On Friday we published, out loud, the question of whether a low-impact release landing exactly on the 14:00 New York option cut — with a USD/JPY 157.00 strike sitting close to spot — would move anything. It was the second consecutive day a print had landed on the cut, and we said a low-impact event is the case that would actually discriminate.
The answer is that we cannot tell, and the reason is instructive. Thirty minutes after that cut, the Bank of Japan telephoned dealers to check where the yen was trading. USD/JPY made its session high somewhere in that half-hour window and then gave back more than a hundred pips. Two events, thirty minutes apart, one of them far larger than the other, and no minute-by-minute record we trust that separates them.
So the discriminating test we set up on Friday was destroyed by an unrelated event on Friday night. That is worth recording rather than fudging: the experiment was well designed and the week ran over it. Our Asia Desk piece this morning covers the rate check and the three-day Tokyo closure that follows it; this is the same Friday evening seen from the calendar rather than from the order book.
Every yen-relevant number this week prints while Tokyo is shut
Japan is closed Monday, Tuesday and Wednesday and reopens on Thursday 24 September. Here is the week in UTC. Times are converted from two Japanese publishers’ JST schedules; where they agree we have said so, and the conversion is a flat subtraction of nine hours because Japan does not observe daylight saving.
- Monday 21 Sep, Japan closed. 10:30 — Chicago Fed president speaks. 12:30 — Chicago Fed National Activity Index.
- Tuesday 22 Sep, Japan closed. 03:10 — an RBA speech. 11:00 — ECB president speaks. 14:00 — Richmond Fed manufacturing index. 14:05 and 14:20 — two Federal Reserve speakers back to back. A Japan–US summit is also listed for the day.
- Wednesday 23 Sep, Japan closed. 08:00 — euro area flash PMIs. 08:30 — UK flash PMIs. 13:45 — US flash PMIs. 14:30 — weekly US crude inventories.
- Thursday 24 Sep — Tokyo reopens and the Bank of Japan’s 1.25% takes effect on the same day. 01:30 — Australian employment. 07:30 — Swiss National Bank decision. 12:30 — US jobless claims. 14:00 — new home sales.
- Friday 25 Sep. 09:15 — Bank of England governor speaks. 12:30 — US durable goods orders. 14:00 — University of Michigan sentiment, final. China is on holiday.
Read that list as a scheduling problem rather than a forecast. The three days on which the yen has no domestic market are the three days carrying a Japan–US summit, two ECB appearances, three Federal Reserve speakers and the global flash PMI round. Everything that can reprice the yen this week happens while the market that trades it most is on holiday, and the first Tokyo price that sees any of it is Thursday morning — the same session in which the new policy rate starts accruing.
What this does not tell you
Whose consensus we scored. The four figures we marked are the ones we published on Friday, from two vendors that agreed on all four. Another vendor may have had different numbers and would score differently. A consensus is a publication, not a fact, and we have now said so often enough that it should be assumed.
One number we cannot reconcile. On Friday we read the Conference Board’s July index as 99.5. Today the same page says August decreased 0.1% to 99.5. Both cannot be right unless July was revised, and the page does not show us the revision. We are not building anything on the index level, and we are flagging it rather than quietly using whichever version suits.
The week’s times are converted, not primary. Two publishers give the JST schedule and they agree on the shape of the week; we have converted. Every converted time lands on a standard institutional slot — 12:30 UTC is 8:30 Eastern, 13:45 is 9:45 Eastern, 14:00 is 10:00 Eastern — which is the cheapest sanity check available on a converted calendar and it passes. That is not the same as reading each release page, and before you put size on any single one of these, read its issuer.
The Swiss decision is one publisher’s time. It appears twice, but in two articles from the same publisher, which is one reader, not two.
A speaker’s name we cut. The RBA event on Tuesday is real in both schedules. The name attached to it in the source we read did not survive a second look, so we have published the event without it, as we have before.
