Four Forecasters Had Canada's Sign Wrong by 75,000 Jobs and Its Unemployment Rate Exactly Right

Friday gave this desk two scheduled releases ninety minutes apart and a rare clean test of something we published a month ago: that asking which vendor has the right consensus is a malformed question, because the answer is per field. Canada’s employment report proved it in the hardest possible form. Four forecasters put September employment between a 5,000 and a 9,200 gain; the economy shed 68,300 jobs. Every one of them had the sign wrong and missed by between 73,300 and 77,500. And the unemployment rate printed at 6.5 per cent, exactly where three of those same four said it would. One survey, one release, one minute: a ratio forecast that was perfect and a level forecast that was inverted.

Canada: four wrong signs and one exact rate, from the same four people

The 12:30 UTC release carried both fields. Here is the scoring.

Employment change, actual minus 68,300. RBC about 5,000 higher; CIBC about 5,000 higher; one publisher’s house consensus about 7,000 higher; a newspaper’s figure 9,200 higher. Misses of 73,300, 73,300, 75,300 and 77,500 jobs respectively. Four forecasts, four wrong signs, and not one of them was within seventy thousand.
Unemployment rate, actual 6.5 per cent against a 6.4 prior. Three of the four said 6.5. RBC said unchanged at 6.4. Three exact, one wrong by a tenth.

The two fields come out of one Labour Force Survey. The level was not merely missed, it was inverted; the rate was not merely close, it was exact. The breakdown makes it worse rather than better: full-time employment fell 35,400 and part-time 32,900, a split of 51.8 to 48.2 per cent, so there is no composition story in which the headline is an artefact of one thin category. The economy shed a hundred and ten thousand jobs across August and September and the unemployment rate moved a tenth.

That last sentence is the whole point and it is not a paradox. A rate is a quotient. If the labour force contracts alongside employment, the quotient barely moves while the numerator falls off a cliff. So a forecaster who is modelling participation can be exactly right about the rate and catastrophically wrong about the level, and a surprise filter keyed to the rate will record Friday as a non-event. Average hourly pay for permanent employees rose 2.3 per cent year on year against 2.0 per cent in August — a third field, moving the other way, and nobody on our list forecast it.

What the market did: the Canadian dollar fell 0.44 per cent to 1.4287, dollar–Canada rose about fifty pips inside fifteen minutes of the release, and the two-year Canadian yield fell 9.5 basis points to 2.410 per cent. The rate field said nothing happened. The bond market priced seventy thousand jobs.

Michigan: the whole band was on the wrong side, and the nearest miss was 1.2

Ninety minutes later the October preliminary Michigan sentiment index printed 46.3. This desk spent Friday morning publishing a three-way disagreement about what the consensus was: 47.5, 47.6 and 48.0, against a September final of 48.1. We wrote that the spread straddled the sign — the lowest estimate implied a six-tenths fall and the highest a tenth, effectively unchanged — and reserved the adjudication for this slot.

There is nothing to adjudicate. The print was 1.8 points below the prior.

Actual fall 1.8 points. The largest forecast fall was 0.6; the actual is three times it. The smallest forecast fall was 0.1; the actual is eighteen times it.
The entire consensus band is 0.5 points wide. The nearest estimate missed by 1.2 — two and a half band widths. The furthest missed by 1.7.
Every one of the three was too high. There was no spread of views, only a shared level of optimism.

We have now run this test four weeks in a row and it has failed four times: a closed consensus dispute is still not the print. Settling which vendor has the right number tells you where the forecasters agreed, and on Friday they agreed in the wrong place by more than twice the distance between them. The figures are confirmed at the issuing institution, which publishes the index at 46.3 against 48.1 in September and 53.6 a year ago.

The headline forecast nearly held because both component forecasts broke in opposite directions

This is the finding we did not expect and it is sharper than the headline miss.

The index has two published sub-indices, and consensus figures existed for both:

Current economic conditions: actual 44.7 against a 50.5 consensus and a 50.9 prior. Missed by 5.8 points, low. A 12.2 per cent monthly fall.
Consumer expectations: actual 47.3 against a 45.9 consensus and a 46.3 prior. Beat by 1.4 points. A 2.2 per cent monthly rise.
Headline: actual 46.3 against the middle consensus of 47.6. Missed by 1.3.

The component misses are 5.8 in one direction and 1.4 in the other. The larger of them is four and a half times the headline miss. A forecaster scored on the headline alone looks to have been 1.3 out. The same forecaster, scored on the two things the headline is built from, was 5.8 out on one and backwards on the other.

So on one release, the composite was the forgiving field and the components were the brutal ones; on Canada, the composite ratio was the forgiving field and the level was brutal. In both cases the field that looked accurate was the one with cancellation inside it. If you score forecasters, or score yourself, on whichever single field the vendor headlines, you are measuring the amount of offsetting error in that field rather than anyone’s accuracy.

Inflation expectations went the other way again: year-ahead rose to 4.7 per cent from 4.6, long-run to 3.5 from 3.4. A sentiment index at the first percentile of its own history since 1978, with rising inflation expectations, is the configuration that makes the probability article published alongside this one difficult.

While we are scoring things: the positioning prediction, and a number that finally has a category

The CFTC positioning report landed at 19:30 UTC on Friday, measured at the Tuesday 6 October close. On 9 October this desk published two falsifiable predictions about it. Mark them.

Prediction one: the leveraged-fund net short widens. Correct. Leveraged funds went from net short 14,161 contracts to net short 25,043 — 66,123 long against 91,166 short, a widening of 10,882 contracts. A second publisher’s wire headline independently describes leveraged funds increasing their yen net short, which corroborates the direction without giving us a second number.

Prediction two: unscored, and we are not claiming it. We predicted the two publishers would stay about 2,600 contracts apart on the level while agreeing within a few hundred on the change. Only one of them has published a figure for this Tuesday. On the previous Tuesday the two agreed to the contract at minus 14,161, which is a small cross-check that passes, but the prediction itself cannot be scored until the second reader prints.

The better result is accidental. On 8 October we published that we had carried a yen net long of 55,440 contracts for nine days as though it were the position, and that it was the only one of four figures with no category written next to it. A publisher’s report on this Tuesday gives large speculators net long 62,325 contracts, change plus 6,885. Subtract: 62,325 minus 6,885 is 55,440, exactly. The uncategorised number is the legacy large-speculator net long, and it was identified not by a label but by a vendor’s own change field reconciling to it to the contract.

Which leaves the spread worse than when we complained about it. On one Tuesday, 6 October: leveraged funds net short 25,043, asset managers net long 67,795, large speculators net long 62,325. The extremes are 92,838 contracts apart — about 1.16 trillion yen, or 7.33 billion dollars at Friday’s 158.23 — against the 72,249 we published four days ago. And the two trader-category figures do not sum to the legacy one: 67,795 minus 25,043 is 42,752, which is 19,573 contracts short of 62,325, because the two breakdowns partition the same open interest differently. Three of those numbers come out of one report. None of them is wrong.

What to change in a filter before Wednesday

Score every field a release publishes, not the one the vendor headlines. Friday produced two releases in which the headline field was the one with the most internal cancellation. A surprise filter reading Canada’s rate alone saw a non-event; one reading Michigan’s headline alone saw a 1.3 miss instead of a 5.8 one.
Store a positioning figure with its trader category or do not store it. Nine days of carrying 55,440 without one cost us an article arguing about a contradiction that was partly a taxonomy. The category is not metadata; it determines the sign.
A consensus band’s width is not an error bar. Michigan’s band was 0.5 wide and the nearest estimate missed by 1.2. Sizing a news window off the spread between vendors sizes the disagreement among forecasters, which on Friday was a twentieth of the distance to the print.
Monday is a holiday in Japan and in the United States, and the US holiday is partial — the bond market is closed, the equity market is not. Dollar–yen was 158.23 to 158.35 at 22:00 UTC on Sunday against 158.226 at the weekend close, two readings four ten-thousandths apart. The next release that can move any of this is US September CPI at 12:30 UTC on Wednesday.

What this does not tell you

Statistics Canada’s own release page was not reachable this run. It returned a robots exclusion on the exact URL shape that has worked before. Every Canadian figure above is therefore a secondary reading, and the minus 68,300 is from one publisher while three others give a rounded 68,000. The full-time and part-time split, the wage figure and the market reaction all have a single reader.
The Michigan figures are primary and the consensus figures are not. The index values and inflation expectations come from the issuing institution’s own page. The three headline consensus figures are three vendors, and the two sub-index consensus figures have a single reader. If that reader’s 50.5 and 45.9 are wrong, the component argument in section three weakens considerably.
The 9,200 Canadian estimate was taken from a search-result headline and that article was never opened. It is labelled as one of four readers and nothing above depends on it alone.
We have not established why the labour force moved as it did, because that would need the participation figures and we did not get them from a reachable source. The quotient explanation in section one is arithmetic, not evidence about what Canadians did in September.
One trader category, one publisher, one week. The leveraged-fund widening is a single numeric reader with a second publisher agreeing only on direction. The 92,838-contract spread is real but it is three figures from two vendors reading one report, and we still cannot say which category a system should trade off.
Nothing here is a view on Wednesday. Four weeks of this test say a settled consensus does not predict a print, and that applies to CPI exactly as it applied to Michigan.

FX events calendar — US September CPI at 12:30 UTC on Wednesday 14 October, and the week’s speakers in UTC.
Signals — how a release with two consensus fields is scored when the fields disagree.
EA presets — news-window settings for a holiday Monday and a single-event week.