All Three Legs of Our Auction Pre-Registration Won. Then the Baseline Flipped the Sign of the Answer.

The pre-registration, scored

Taking the legs in the order they were written:

• High yield above 5.216 per cent. Won. The auction cleared at 5.441 per cent, which is 22.5 basis points above the 13 August comparable. This was the leg we were most confident about and therefore the least informative: ten-year yields had gone from the 4.80s to 5.00 over the period and Tuesday’s ten-year cleared at 5.020 per cent, so the direction was close to a free bet. A leg you could not realistically lose is not evidence about your method, and we said at the time that we wanted all three to be losable. This one was the weakest of the three on that test.
• Bid-to-cover between 2.30 and 2.65. Won, on both numbers. Two publishers print 2.36, a third prints 2.401. Both sit inside the band, which is the one genuinely satisfying feature of the result: the band was wide enough to be robust to a vendor disagreement we did not know was coming. It was also wide enough that a 0.041 discrepancy is invisible inside it, which is the same property seen from the other side.
• Indirect share between 64 and 76 per cent. Won. Indirects took 68.1 per cent, almost dead centre of a twelve-point band. Direct bidders took 20.5 per cent and primary dealers 11.4 per cent; the three sum to 100.0, which is the first thing to check on any allotment set and it closes.

So: three for three, and the honest grade is a pass with the mark mostly going to the band widths rather than to the forecasts. We pre-registered ranges broad enough to survive, which is the right trade when you are scoring a method rather than showing off, but nobody should read three wins as three insights.

The tail closes for the first time — and only because two publishers each printed half of it

This desk has chased a when-issued level for 30-year auctions three times and published, repeatedly, that the field is not public — that it is a dealer quote the Treasury never releases, which makes every tail figure you read somebody else’s subtraction that you cannot check. Last night that finally broke, and the way it broke is more interesting than the number.

One publisher printed a when-issued yield of 5.437 per cent and no tail. A different publisher printed a tail of 0.4 basis points and no when-issued level. Neither page is auditable on its own. Put them together:

• High yield 5.441 per cent minus when-issued 5.437 per cent equals 0.004 percentage points, which is 0.4 basis points.
• That reproduces the published tail exactly, to the stated precision.

So the auction tailed by four tenths of a basis point, and for the first time in this file we can say that having done the arithmetic rather than having been told. Two caveats that matter. The when-issued figure is single-sourced, from a publisher whose own page carries a published time with no zone label and metadata with a conflicting offset, so we treat its clock as unresolved even while we use its level. And a 0.4 basis point tail is small enough that it is inside the rounding of several ways of quoting these things — it is a real tail in the strict sense, meaning the high yield came in above the when-issued level, and it is not a tail anyone should build an argument on.

The reusable part: when a derived field is missing, look for its inputs at a different publisher rather than looking for the field itself somewhere else. We spent three sessions asking for a tail. The tail arrived because someone finally printed a when-issued level, and the subtraction was ours.

The baseline flips the sign, and then the vendor flips it again

Here is the whole problem in one table. Same auction, two candidate priors.

• Against 8 October — $22 billion, 5.618 per cent, cover 2.54 (one publisher says 2.542), indirects 72.3 per cent, directs 20.9 per cent: the yield fell 17.7 basis points, the cover fell 0.180 (or 0.141), indirects fell 4.2 points and directs fell 0.4. Demand deteriorated and the Treasury paid less.
• Against 13 August — $25 billion, 5.216 per cent, cover 2.39, indirects 66.8 per cent: the yield rose 22.5 basis points, the cover fell 0.030 — or rose 0.011 on the other publisher’s figure — and indirects rose 1.3 points. Demand was flat to slightly better, at materially more yield.

The indirect share changes sign on the choice of baseline: down 4.2 points against October, up 1.3 points against August. The cover changes sign on the choice of publisher, once you have picked August: down 0.030 on one number, up 0.011 on the other. A reader handed the October comparison concludes the thirty-year is losing its sponsorship. A reader handed the August comparison concludes the opposite and is looking at the same auction.

August is the right prior and we are not pretending this is a close call. Yesterday’s auction was a $25 billion new issue; 13 August was a $25 billion new issue of the same tenor; 8 October was $22 billion. The rule we wrote for ourselves after getting this wrong is that the unit is same tenor, same size, same type, and that “the last auction” and “the last auction of this tenor and size” are different fields of which the first always looks complete. Every publisher we read last night compared to October, none named a date for the prior it used, and one of them — this is the part we enjoy least — carried October’s 5.618 per cent inside today’s auction row on its calendar before the event, which is the figure we ourselves discarded yesterday morning as the wrong baseline. The vendor was serving our rejected number as today’s answer.

What a system should store is therefore not the comparison but the components, keyed by tenor and size, so the baseline is a query rather than a stored verdict. Store a delta and you have stored somebody’s choice of prior permanently and invisibly.

Where the paper went, and an unlabelled average we can finally name

Strip the adjectives and the distribution is the story. Primary dealers took 11.4 per cent against an average of 9.5 per cent — 1.9 points more, which is 20.0 per cent above that average — and one publisher calls it their largest share since June. Dealers taking more is the residual: it is what is left when the bidders who wanted the paper did not want all of it. Directs took 20.5 per cent against an average of 20.0, essentially unchanged. Indirects took 68.1 against an average of 70.5, down 2.4 points.

So the shortfall is small and it went to dealers. That is consistent with a 0.4 basis point tail and it is consistent with “in line”; it is not consistent with any of the stronger adjectives, and it is the reason we would not have written “weakest since June” even if we could second-source it, which we cannot.

Now the averages, which are where this desk has had a standing complaint. One publisher prints four averages — cover 2.44, indirect 70.5, direct 20.0, dealer 9.5 — and states no window for any of them, which has made them unusable for two months. A second publisher prints a cover average of 2.437 and labels it explicitly as a six-auction average. Round 2.437 to two places and you get 2.44. The collision is 0.003 wide.

That identifies the unlabelled window: the first publisher’s averages are the six-auction window, which means its indirect 70.5, direct 20.0 and dealer 9.5 are all now usable with a stated denominator for the first time. The method is the one we have been using on pivot ladders and probability aggregates — an unlabelled figure is identifiable by collision with a labelled one — and this is the first time it has recovered four fields from one match.

It also hands you a live example of why the window matters. Measured against 2.44 the cover is 0.080 light; measured against 2.437 it is 0.036 light. Those are the same auction against the same window, and they differ by a factor of 2.22, entirely because the two publishers disagree about the cover itself. That is the 2.36-versus-2.401 dispute cashed out: it is invisible inside our pre-registration band and it roughly halves or doubles how light the auction looks against its own average. We are publishing both values and naming both publishers rather than picking one, and we are logging it as a genuine vendor disagreement on one field rather than two readings at two times, because an auction cover is a single settled quantity computed from one set of tenders. Unlike an intraday high, it cannot drift.

What this does not tell you

• We do not know which cover is right. 2.36 appears at two publishers and 2.401 at one. We have not been to the Treasury’s own results page, which publishes the authoritative figure, and our file says not to try it on the auction day itself. The right way to close this is to read the issuer this morning, and the next slot should.
• The when-issued level is single-sourced. The tail reproduces beautifully from it, which is reassuring and is not independent confirmation — a wrong when-issued level and a tail computed from it would also agree.
• The August baseline figures — $25 billion, 5.216 per cent, cover 2.39, indirects 66.8 per cent — come from our own file and were not re-sourced last night. If August is wrong, the sign flip above is wrong with it.
• We have no primary dealer or direct share for the August auction, so the dealer comparison runs against the six-auction average only. The most interesting number in the piece is the one with the shallowest baseline.
• “Largest dealer share since June” is one publisher’s claim and we could not check it. We are reporting that it was claimed, not that it is true.
• The averages are now windowed by a collision, not by a statement. It is strong evidence and it is an inference; if either publisher revises a figure the identification goes with it.
• No settlement-day effect is modelled here. The paper settles Monday 16 November, four days after the auction, and we say nothing about what that does to the front end.
• This is not a view on thirty-year yields. Three pre-registration legs won and we still think the bidder distribution is the only part of the result with information in it.

Related

• FX events calendar · auction, release and settlement times in UTC
• Signals · how we size around scheduled supply
• Presets · news-window and filter configuration