Yesterday morning this desk published Japan’s eighth consecutive monthly real-wage gain as a fact about August. It is a fact about August. What we did not publish — because we did not look — is that the same release restated five fields of the month before it. July’s nominal cash earnings, reported on 8 September at 4.7 per cent year on year and headlined as the fastest since January 1997, are now 4.3 per cent. July’s real wages, reported at 2.4 per cent and headlined as the biggest gain since May 2021, are now 2.0 per cent. July’s regular pay, reported at 4.1 per cent and headlined as the fastest since April 1992, is now 3.8 per cent. Three superlatives, three numbers that no longer exist.
Five fields moved, and not all of them in the same direction
The August Monthly Labour Survey came out at 23:30 UTC on 6 October. We wrote about it at 23:19 UTC on 7 October, in a piece about a Japanese government bond auction, and the line we used was the eighth straight real gain. The release also carries revised July figures, and the same wire service that filed the July numbers on 8 September filed the August numbers with the restated July alongside them. Putting the two filings side by side is a one-fetch job and it took us until tonight.
Real wages: 2.4 to 2.0, down four tenths. Nominal cash earnings: 4.7 to 4.3, down four tenths. Regular pay: 4.1 to 3.8, down three tenths. Special payments: 6.3 to 5.3, down a full point. And overtime pay: 3.1 to 4.5, up one and four tenths.
That last one is the detail that stops this being a story about a statistical agency marking its numbers down. Four fields fell and one rose, and the one that rose rose by more than any of the others fell. A revision cycle that moved everything one way would be a sign of a systematic first-pass bias you could at least think about adjusting for. This is not that. It is five independent corrections of five sub-series, and the only safe generalisation is that any one of them can move by more than a tenth.
The slowdown we are all discussing is 44 per cent smaller than the one that was published
Every reader of yesterday’s release was told that real wage growth decelerated for a second consecutive month. It did. But the size of the deceleration depends on which version of July you measure from, and the two versions are not close.
Measured from July as originally published, August’s nominal 3.8 per cent is a fall of nine tenths from 4.7. Measured from July as it now stands, it is a fall of five tenths from 4.3. The real series behaves identically: nine tenths from 2.4, five tenths from 2.0. In both cases the restated slowdown is 44.4 per cent smaller than the one a reader would have computed from the figures in front of them a month ago.
Overtime is worse. From the original 3.1 per cent, August’s 5.2 is an acceleration of two and one tenth points. From the restated 4.5, it is seven tenths — a magnitude two thirds smaller. If you had built any kind of momentum reading on Japanese overtime hours, and overtime is the series people reach for when they want a cyclical signal out of this survey, you built it on a number that has since moved more than the thing you were measuring.
None of this makes the August print wrong. It makes every month-on-month comparison involving July provisional in a way that the headline does not say.
The identity that cannot check a tenth
The obvious way to audit a real-wage figure is to subtract the deflator from the nominal rate. The release gives both: the inflation rate used for the real calculation was 2.2 per cent in August and 2.2 per cent in July, unchanged between the two months.
We ran it on all three figures. August: 3.8 minus 2.2 is 1.6, against a published 1.5. July restated: 4.3 minus 2.2 is 2.1, against a published 2.0. July as originally filed: 4.7 minus 2.2 is 2.5, against a published 2.4. The subtraction is one tenth too high in all three cases, with the same sign every time.
Compounding the rates properly rather than subtracting them — dividing one plus the nominal rate by one plus the deflator — gives 1.566, 2.055 and 2.446, which closes most of the gap but not all of it, and leaves a residual of the same sign in all three. So the identity corroborates to roughly a tenth and no better. That is the point worth keeping: the revisions we are discussing are themselves three and four tenths, and an arithmetic check with a one-tenth floor cannot adjudicate them. The identity is a sanity check on the level, not an audit of the revision. We are recording that limit because the temptation to use it as an audit was real.
The level fell 28.7 per cent and nothing is broken
One number in the release will look alarming if you read the two months in sequence. July’s average cash earnings were 436,401 yen per worker per month. August’s were 311,364. That is 125,037 yen less, a fall of 28.7 per cent, in a month the release describes as 3.8 per cent higher than a year earlier.
Both are true. July is the summer bonus month in Japan and the special-payments component is what carries the difference; the year-on-year rate compares each month with the same month, so the seasonal cliff does not appear in it. We state it because a reader pulling both levels off a chart without the seasonality in mind would reasonably conclude that something had gone wrong with the data. Nothing has. But it is also why the special-payments revision of a full point matters more than its size suggests: in a bonus month, that component is most of the level.
The beat bought nothing, and we cannot tell you whether the pricing moved
August nominal cash earnings came in at 3.8 per cent against a consensus of 3.7, a beat of one tenth, on the eighth straight real gain, with the Bank of Japan’s policy rate at 1.25 per cent after the 18 September increase. That is the configuration the market has spent a month saying it needs in order to price the next move.
Here is what the next move reads tonight, and we are naming every feed because they do not agree. For October, this desk has been carrying 12 per cent. One Tokyo publisher puts overnight index swaps at a little over 10 per cent. One English publisher puts futures at about 17 per cent. For December, we have been carrying near 90 per cent; the same Tokyo publisher says a little over 80, and the same English publisher says about 71.
Seven points wide on October and nineteen on December, and two of the three readings are different instruments — swaps against futures — so the spread is not necessarily disagreement. We will not tell you that December pricing fell 19 points, because we cannot separate a move from a change of measurement, and the honest version is that the December reading now spans 71 to 90 depending on whose screen you read. What we can say is that nothing in the band is consistent with a wage beat having bought anything. For sizing, that is the usable half: a Japanese wage print that lands at or above consensus is not currently a reason to change the size of a yen position, and this release is the eighth consecutive opportunity to test that.
What this does not tell you
We did not reach the issuing ministry. Every figure here is a secondary reading of the Monthly Labour Survey, taken from one wire service’s two filings a month apart and corroborated on the August numbers by a second vendor’s release page. The originals for July are from the September filing and the restatements are from the October one; we did not find a page at the ministry that shows both versions of a field side by side, which is the document that would settle this properly.
We do not know whether the three superlatives survive at the revised values. “Fastest since January 1997” was said about 4.7 per cent; whether 4.3 is also the fastest since January 1997 requires a run of history we did not fetch, and we are not going to guess it either way. The same applies to the May 2021 and April 1992 claims. What we are asserting is narrower and sufficient: the numbers those claims were attached to have changed.
We did not verify whether these are first or second revisions, or whether a further revision is scheduled. And the probability figures above are three readings from three sources on two different instruments; none is a market we traded, and the October and December meeting dates themselves are given variously as 30 October and 29 to 30 October by different publishers.
Related
- FX events calendar — release times in UTC, including the Japanese monthly labour survey.
- Signals — how we treat a print that beats consensus and moves nothing.
- EA presets — news-window and revision handling.
