A Reopened Ten-Year Is Recorded as 9-Year 11-Month, and Comparisons Break There

Between May and July 2026 one Treasury note went through three auctions. The security did not change: CUSIP 91282CQQ7, a 4.375 percent note in series C-2036 maturing on 15 May 2036. The bid-to-cover ratio printed 2.40, then 2.57, then 2.59.

What changed in the record is the label. Treasury's auction file describes the first sale as a 10-Year, the second as a 9-Year 11-Month, and the third as a 9-Year 10-Month. A screen, a scraper or a spreadsheet that filters the term field on 10-Year keeps the 2.40 and discards the 2.57 and the 2.59. Whatever comes out of that filter is not a demand series for the ten-year note. It is partly a calendar artefact.

The distinction matters at a specific moment. Auction results follow the bidding deadline, and the ratio is read against whatever a desk holds as the previous comparable auction. That comparison is made from a stored series, and the stored series is usually built by a filter that somebody wrote once. If the filter keys on the term label, the baseline it returns is not the one the reader thinks they are looking at.

The term field records what is left, not what was sold

Treasury sells a new ten-year note at the quarterly refunding, then sells more of that same note in the months that follow. The governing regulation defines the second kind of sale in one sentence. Under 31 CFR 356.2, “Reopening means the auction of an additional amount of an outstanding security.” The reopened security carries the original coupon, the original maturity date and the original CUSIP.

The refundings fall in February, May, August and November, so one ten-year note reaches the market at three separate auctions, at three separate yields, in front of three separate sets of month-end and index-rebalancing flows. Demand at the third sale is a statement about a security that has already been distributed twice, which is a different question from demand at the first.

Because the maturity date is fixed and the calendar keeps moving, the remaining term shrinks between auctions. The published dataset carries both readings. The field security_term holds the remaining term at the moment of the auction, and the field original_security_term holds the term the security was issued at. For the three sales of 91282CQQ7 the first field reads 10-Year, 9-Year 11-Month and 9-Year 10-Month in turn, while the second field reads 10-Year each time and the reopening flag reads No, Yes, Yes.

There is a second collision, and it runs the other way. The term label is not reserved for nominal notes. The auction held on 23 July 2026 also carries security_term of 10-Year, but its inflation_index_security flag reads Yes: it is an inflation-indexed security with its own bidder base and its own pricing conventions. A term filter drops two of the note's own auctions and admits an instrument that trades on a different basis.

What a term filter actually collects

As of 3 September 2026 the dataset holds ten auctions dated 1 March or later whose original_security_term is 10-Year. Nine had been held; the tenth, dated 9 September, was announced but not yet auctioned and carries no ratio. Of the nine completed sales, six were nominal notes and three were inflation-indexed.

The three inflation-indexed sales in that window were held on 19 March, 21 May and 23 July, printing 2.47, 2.52 and 2.30. Two of them carry a shortened term label, so a nominal-note filter keyed on 10-Year excludes them by accident rather than by design. The third, on 23 July, was a new issue and therefore carries the full 10-Year label, which is how it lands inside a series that was meant to hold nominal paper only.

Three of those nine records carry a security_term that literally reads 10-Year: the nominal new issue of 12 May at 2.40, the inflation-indexed sale of 23 July at 2.30, and the nominal new issue of 12 August at 2.53. Their mean is 2.41, a figure calculated for this article rather than published by Treasury.

Two of those three are nominal, but they are not the same note. The May sale was CUSIP 91282CQQ7, a 4.375 percent security maturing on 15 May 2036. The August sale was CUSIP 91282CRF0, a 4.625 percent security maturing on 15 August 2036. Even the records that survive the filter and pass a nominal-only screen describe two different securities, three months and a quarter-point of coupon apart.

The six nominal note auctions, in date order, printed 2.45, 2.43, 2.40, 2.57, 2.59 and 2.53. Their mean is 2.495 and their sample standard deviation is 0.079, both calculated for this article from the same published ratios. The spread between the weakest and the strongest of the six is 0.19.

The gap between the two means is 0.085. Set against the 0.079 standard deviation of the nominal series, that is roughly 1.08 standard deviations. It comes from the label rather than from anything bidders did. A trader comparing the August 2.53 against a filtered baseline of 2.41 sees a strong auction; against the series the note actually generated, 2.53 sits a little above the middle and slightly below the June and July reopenings.

The denominator is not the accepted column

The same section of the regulation defines the ratio itself. Under 31 CFR 356.2, “Bid-to-cover ratio means the total par amount of securities bid for in an auction divided by the total par amount of securities awarded. It excludes bids by, and awards to, the Federal Reserve for its own account.” The second sentence is the one that breaks a naive recomputation.

Take the 12 August 2026 ten-year note. The published fields give total tendered of 116,957,209,500, total accepted of 52,623,557,100 and an offering amount of 42,000,000,000. Dividing the first by the second gives 2.223. The published ratio is 2.53.

The accepted column exceeds the offering by 10,623,557,100, which is 20.19 percent of the accepted total. An award to the Federal Reserve for its own account is the category the regulation removes from the ratio, and an excess of the accepted total over the offering is where such an award would sit in these columns. Removing the same amount from both columns gives 106,333,652,400 divided by 42,000,000,000, or 2.5318, which rounds to the published 2.53. That reconstruction is this article's arithmetic; the dataset carries no field that names the add-on directly.

The size of that excess is not fixed, and it is not a demand variable. The two auctions examined here sit far apart on it: 10,623,557,100 in August against effectively nothing in June. Because the offering amount and the accepted total are both published, the gap between them can be measured auction by auction, and checking it is the first step before any ratio is recomputed from raw columns.

The June reopening supports the reading from the other side. On 10 June the accepted total of 39,000,001,000 sat on an offering of 39,000,000,000, so there was effectively no excess to remove, and total tendered of 100,044,669,600 divided by that accepted total gives 2.5652, which rounds to the published 2.57. The simple division works when the add-on is absent and fails by 0.31 when it is present. That is a large error in a statistic quoted to two decimal places.

Bidder shares have to be built from the dollar columns

The same file reports who took the paper, but in dollars rather than percentages, so the shares have to be constructed. For the 10 June reopening, competitive accepted was 38,914,680,400, split into 3,683,175,000 to primary dealers, 4,796,000,000 to direct bidders and 30,435,505,400 to indirect bidders. As shares of competitive accepted, calculated for this article, that is 9.46 percent, 12.32 percent and 78.21 percent.

For the 12 August new issue, competitive accepted was 41,821,613,600, split into 3,597,810,000, 6,135,867,600 and 32,087,936,000. The same calculation gives 8.60 percent to dealers, 14.67 percent to direct bidders and 76.73 percent to indirect bidders. The dealer share was 0.86 percentage points higher at the reopening; the direct share was 2.35 percentage points lower.

Two auctions do not establish a pattern, and the denominator matters here too: shares of competitive accepted are not the same as shares of total accepted, because the noncompetitive tranche and any add-on sit outside the competitive column. A series of dealer shares is only comparable if the denominator is held fixed across every observation in it.

The file also carries an allocation percentage, which reports the share of bids at the stop that were filled: 71.34 at the June reopening against 65.27 at the August new issue. It answers a narrower question than the bid-to-cover ratio does, about crowding at one price rather than about the depth of the whole book, and the two statistics can move in opposite directions at the same auction.

What Would Invalidate This

The six-auction sample is small. A standard deviation estimated from six points is itself imprecise, so the statement that the labelling gap is about one standard deviation should be read as an order of magnitude rather than a measured quantity. Widening the window to two or three years would tighten that estimate and could easily move the multiple below one.

The 0.085 gap is arithmetic on one specific window. A window that happens to place the quarterly new issues on the strong side of the range would produce a gap with the opposite sign, and a window containing no inflation-indexed sale labelled 10-Year would shrink it. The direction of the distortion is not fixed; only its presence is.

A vendor that normalises the term field, or that exposes the original term and the reopening flag alongside it, hands over a series with none of this in it, and a desk working from such a series gains nothing from the correction described here. The problem belongs to series built directly from the raw file, which is the common case for a small research stack.

The reconstruction of the August ratio is an inference. It is consistent with the definition in 31 CFR 356.2 and it reproduces the published figure to two decimals, but the dataset does not label the excess, and another decomposition that happens to reproduce 2.53 cannot be ruled out from these fields alone.

Finally, none of this makes bid-to-cover informative. Building a clean series removes a measurement error; it does not establish that the cleaned series carries signal about the next session. Auction statistics compete with dealer positioning, the concession built into the days before the sale and the macro calendar around it.

Concrete Framework

1. Key the series on the security, not the label. Use original_security_term together with the reopening flag, or key on CUSIP. A term filter alone silently changes the population it collects.
2. Separate nominal from inflation-indexed before averaging. The inflation_index_security flag does this in one column. In the March-to-August window it moves three of nine observations out of the nominal series.
3. Recompute the published ratio before trusting a self-built one. If total tendered divided by total accepted does not reproduce the printed figure, the accepted column contains an add-on that 31 CFR 356.2 excludes. Compare the accepted total with the offering amount to size it.
4. Fix one denominator for bidder shares and keep it. Competitive accepted and total accepted give different percentages for the same auction, and mixing them across a series creates drift that looks like a change in demand.
5. Write down the comparison set before the print lands. For a ten-year auction that means the prior nominal ten-year sales, new issues and reopenings together, with the count stated. A baseline chosen after the number appears will tend to flatter whatever the number was.
6. Size the reaction against the dispersion of the series, not the last print. With a sample standard deviation near 0.079, a 0.03 difference from the previous auction sits inside the normal spread of these six observations, and a plan built on it is reading noise.

This article describes how a public auction dataset is structured and how figures inside it are calculated. It is a study of data construction, not investment advice, and it recommends no security or position. Every figure above is drawn from the U.S. Treasury Fiscal Data auction file and from 31 CFR 356.2; the derived means, shares and differences are identified in the text as calculations made for this article.