The auction effect is repeatable, statistically clean and small. Across 33 years, on-the-run Treasury yields climb before the auction close of the next same-tenor security, then retreat after the results. The round trip measures 0.712 basis points for the 2-year and 1.192 for the 10-year. All six tenors are significant at 1%. Most of the explanatory power comes from a variable observed only after the move has finished.
We could not trade the market studied in the paper. This desk has no cash Treasury prices and no trade-and-quote data, so none of the results below has been tested on our own book.
The design is the contribution.
Fleming, Liu and Nguyen start from dealer inventory. An auction places a large, pre-announced block of duration with dealers whose risk-bearing capacity is finite. They sell related securities ahead of the event to clear room, then repurchase once the takedown becomes known. Prices weaken on the way in and recover afterward, tracing an inverted V. Lou, Yan and Zhang documented the pattern over multi-day windows: yields rose 2 to 3 bps before auctions and fell roughly 2 bps afterward, from 1980 to June 2008. Over several days, however, the window also contains CPI prints, FOMC and month-end.
The authors instead use one-minute interdealer trades and quotes. GovPX covers June 1991 through 2000, followed by BrokerTec from 2001 to July 2024. Their universe contains the on-the-run 2-, 3-, 5-, 7-, 10- and 30-year. Because issuance was suspended for most of the sample, the 20-year is excluded.
Yield pressure is defined as the yield change during the 180 minutes before auction close minus the change during the 180 minutes after results are released. A large positive reading describes a pronounced V. Each auction is paired with two control days, using the same weekday one week earlier and one week later. The screen removes FOMC days, early closes and other auctions. Counts range from 398 for the 2-year to 192 for the 7-year, producing 1,685 auction-maturity events. The main regressions use 1,663 of them.
Flow carries the regressions
Dealer-constraint proxies explain little. A one-standard-deviation increase in MOVE adds 0.426 bps, and MOVE alone remains in the joint specification, at 0.410. Two widely cited direct measures go nowhere. Broker-dealer leverage growth is 0.007, while He-Kelly-Manela intermediary capital growth is -0.015; neither is significant.
Demand measures fare somewhat better. In the joint model, lagged investment fund share contributes -0.742 bps per standard deviation. Signed interdealer flow is much stronger. Pre-auction net order flow is negative across all six tenors, ranging from -$304M on the 2-year to -$20M on the 30-year. After the auction it turns positive, from +$274M down to +$78M. A one-standard-deviation increase in pre-auction net selling raises yield pressure by 0.762 bps. The equivalent increase in post-auction buying adds 1.054. Once both regressors enter, adjusted R-squared rises from 4.7% to 19.8%.
Liquidity improves before the auction
The result worth retaining under any reading of the rest is the depth profile. Inside depth stays above matched control days throughout the pre-auction window and reaches its widest point about ten minutes before the close. Realized volatility is lower than on control days. It jumps when results arrive, then settles within an hour.
Auction pressure therefore compensates dealers for warehousing risk in a market that is deeper than usual. The authors likewise interpret the evidence as dealer management of auction-induced inventory shocks in an otherwise liquid and orderly market. Auction-window risk limits built around thinning books have the direction wrong.
The placebos strengthen that reading. Neither FOMC nor employment reports produces the same pre-event selling followed by post-event buying. After FOMC announcements, yields drift lower by 0.5 to 1.5 bps without any accompanying buying pressure.
Explanation arrives after the trade
Pre-auction selling and the rise in pre-auction yields occupy the same 180 minutes. The post-auction variables are also contemporaneous. These regressions account for a move as it happens, while the sign of net order flow becomes known only afterward. The authors establish a transmission channel, and that claim holds. They supply no timing rule and make no claim to one.
The magnitude decomposition deserves more caution. Flow-driven pressure equals net order flow multiplied by price impact estimated from control days. The calculation assumes linearity and stationarity, and the impact coefficient is fitted in-sample. Full-sample estimates range from 0.762 to 1.532 bps. During 2007-2014, they rise to 1.534 to 3.294.
The picture changes in 2015-2024. Flow-driven pressure ranges from -0.034 to 0.566 and reaches significance for only two of six tenors: 0.430 for the 5-year and 0.566 for the 10-year. Total yield pressure during the same period remains 0.340 bps on the 3-year and 0.919 on the 5-year. Order flow still explains cross-auction variation after 2015. Adjusted R-squared runs from 2.1% to 12.8%, with both flow coefficients significant at 1%.
The paper distinguishes explanation of variation from decomposition of magnitude. Its dominant full-sample channel does not explain the recent period's level. Nor does the residual supply the answer, since the non-flow-driven component is insignificant everywhere in the full sample.
The object under study is the outstanding on-the-run note around the auction of the next same-tenor issue. Its measures come from interdealer quotes and signed interdealer flow. We have neither cash Treasury prices nor trade-and-quote data. ETF substitutes such as SHY, IEF and TLT may transmit the pressure differently because they have basket composition, duration exposure and creation-redemption mechanics of their own. For our purposes, the mechanism's survival remains unvalidated rather than disproven.
How much decline is established?
The abstract says price pressure "has not increased in recent years, partly because non-dealer participants now absorb more auction supply and ease dealers' intermediation burden." Its point estimates fit that account. Pressure ranges from 0.887 to 1.732 bps in 2007-2014, versus 0.340 to 0.919 in 2015-2024.
Evidence for the stronger interpretation is limited. Reported F-tests reject equality between those periods for the 3-year (3.824) and 10-year (3.468) notes, both at 10%. Only two of six pass that threshold. Comparisons with 1991-2006 reject even less.
The structural explanation is plausible, and investor composition changed sharply. For the 10-year, dealer award share fell from 82.15% in 2000 to 17.70% in 2024. Fund share moved from 3.86% to 67.12% over the same interval. Those figures are averages across auctions of that maturity, measured in the first year of investor-class data and again in 2024. They corroborate the proposed mechanism without testing the decline itself.
What reaches a trading desk
The dollar estimates measure issuance costs. Expressed as returns, pressure ranges from 1.336 bps for the 2-year to 17.106 for the 30-year. Applied to a $28B 2-year auction, the implied cost comes to about $3.8M per auction and roughly $45M a year. For the 10-year, it is about $21M per auction; for the 30-year, about $26M. Those amounts represent Treasury savings if intraday price pressure were absent.
On our own arithmetic, 1.336 bps of 2-year price pressure lies near the interdealer tick. The paper reports that the on-the-run bid-ask is typically constrained to a single tick, which is why spread is omitted as an explanatory variable. It provides HAC-adjusted averages, without a hit rate or per-auction dispersion for yield pressure. The share of auctions that actually forms a V remains unanswered.
The 30-year makes the trading problem plain. Its yield rises 0.634 bps, followed by an insignificant -0.235 bps reversal. The return-space estimate is largest exactly where the reversal required by a mean-reversion trade is absent.
Execution and spillover geometry are the practical residue. A 10-year auction moves yields across all six tenors by 0.796 to 1.294 bps. A 2-year auction affects only the 2-year (0.712) and 3-year (0.376). A 30-year auction reaches only the 30-year. For a large intermediate-tenor order worked into an auction close, this gives a usable map.
The authors acknowledge that their six-hour window captures about 25% of the 10-day effect, whose average is 2 to 5 bps. Their issuance-cost estimate is therefore a lower bound by construction. They also narrow the contribution in advance to clean identification: whether pressure exists, how it changes over time and how order flow drives it, "rather than a full estimate of the auction effect."
That narrower claim largely holds. Its success makes the tradability concern sharper, since a well-identified quarter of a two-to-five basis point round trip still comes without a timing rule.
A distribution would change my view on tradability: the per-auction spread of yield pressure, together with a hit rate.