license: apache-2.0
language:
- en
pretty_name: US Treasury Events (point-in-time, auctions and official curves)
size_categories:
- 100K<n<1M
task_categories:
- tabular-regression
- time-series-forecasting
tags:
- finance
- treasury
- auctions
- yield-curve
- interest-rates
- point-in-time
- tips
- ziplime
configs:
- config_name: pit
data_files:
- split: train
path: data/pit/*.parquet
- config_name: auction_results
data_files:
- split: train
path: data/auction_results/*.parquet
- config_name: auction_calendar
data_files:
- split: train
path: data/auction_calendar/*.parquet
- config_name: auction_announcements
data_files:
- split: train
path: data/auction_announcements/*.parquet
- config_name: nominal_yield_curve
data_files:
- split: train
path: data/nominal_yield_curve/*.parquet
- config_name: real_yield_curve
data_files:
- split: train
path: data/real_yield_curve/*.parquet
- config_name: bill_rates
data_files:
- split: train
path: data/bill_rates/*.parquet
- config_name: curve_features
data_files:
- split: train
path: data/curve_features/*.parquet
US Treasury Events
Every Treasury auction since 1979 and every official daily yield curve since 1990 — and, for each of them, the instant it became public.
11 114 auctions · 11 108 results · 113 836 nominal curve points · 34 141 real · 26 453 bill rates · 268 960 point-in-time rows · October 1979 to today
The pipeline lives in recipe/ at the same revision as the data.
See PIPELINE.md for the method.
Two timestamps, and the reason they are in one dataset
An auction result and a yield curve are dated in opposite ways, and a strategy that reads both has to respect both.
The auction states its own minute
The Treasury publishes a small artifact for every auction since March 2008, and it settles the question outright:
<CompetitiveClosingTime>13:00</CompetitiveClosingTime>
<ReleaseTime>13:02</ReleaseTime>
Two minutes after the close, stated by the Treasury. That is knowledge_at,
and knowledge_time_precision is minute.
4 872 results carry a stated release minute, the first on 7 April 2008. The other 6 236 do not, and say so.
Before March 2008 there is no artifact. Those results are dated to the end of the auction day, because the alternative — assuming the result was public at the close — would hand a backtest the number minutes before anyone had it, on every auction for twenty-nine years.
The curve does not
This is the line most Treasury datasets get wrong, so it is worth quoting the Treasury directly. Its methodology page says the inputs are
indicative, bid-side market price quotations (not actual transactions) for the most recently auctioned securities obtained by the Federal Reserve Bank of New York at or near 3:30 PM each trading day
and, further down the same page,
Yield curve rates are usually available at Treasury's interest rate website by 6:00 PM Eastern Time each trading day, but may be delayed due to system problems or other issues.
Those are two different instants. 3:30 p.m. is what the number measures; 6:00 p.m. is when anyone could read it. So every curve row carries both:
| Column | |
|---|---|
market_observation_at |
15:30 Eastern — what the number describes |
knowledge_at |
18:00 Eastern — when it could be read |
knowledge_time_precision |
modeled_conservative |
availability_time_basis |
treasury_usually_available_by_1800_et |
Dating a curve to 3:30 p.m. would hand a backtest two and a half hours of the
afternoon it did not have, every trading day for thirty-five years. The gate
fails the build if any curve row is knowable before 18:00 Eastern, and
tests/test_quality.py proves the check by breaking it.
What is deliberately missing
wi_yield and auction_tail are null on every row, and the gate fails the
build if either is ever populated.
The tail is the auction's high yield minus the when-issued market yield. No
official point-in-time when-issued series exists. The daily par yield curve is
not one — it is struck from indicative quotations at 3:30 p.m., hours after most
auctions close — and subtracting it from the high yield produces a plausible
number measuring something else. A column called auction_tail that does not
contain an auction tail is worse than no column at all.
A null is not a zero
The Treasury did not start publishing all of this at once, and the differences are large enough to change an analysis.
| Begins | |
|---|---|
| auction query | October 1979 |
| bidder categories — indirect, direct, primary dealer | 2008 |
| the artifact that states the release minute | April 2008 |
| nominal par yield curve | January 1990 |
| bill rates | January 2002 |
| TIPS real yield curve | January 2003 |
Read a 1995 auction expecting an indirect bid share and you get null, because the Treasury did not publish one. It is not zero, and this dataset never turns it into zero.
metadata/field_coverage.parquet records the first date every numeric field is
populated, so a reader can tell "the Treasury had not started publishing this"
apart from "nobody bid". Nowhere in this dataset does an absent field become a
zero.
The methodology changed, and the rates are not recomputed
The par yield curve is the output of a procedure, and the procedure changed: the monotone convex method replaced the quasi-cubic hermite spline as of 6 December 2021, which the Treasury states on its methodology page.
A rate from 2015 and a rate from 2022 are therefore not the same kind of
number. Neither is recomputed under the other's method — the published values
stand as published. methodology_id on every nominal curve row says which
procedure produced it, and metadata/yield_curve_methodology.parquet says what
each one was, with the Treasury's own description and the source URL.
Maturities come and go, and nothing is filled in
The 20-year bond returned in May 2020 after seventeen years away. The 4-month bill was added in October 2022. The set of maturities on the curve is not constant, and this dataset stores the curve in long form so that a maturity the Treasury did not publish is an absent row, not a null that looks like missing data and invites interpolation.
| Maturity | First published | |
|---|---|---|
| 3M, 6M, 1Y, 2Y, 3Y, 5Y, 7Y, 10Y | 2 January 1990 | the original set |
| 20Y | 1 October 1993 | |
| 30Y | 2 January 1990 | absent 15 Feb 2002 – 9 Feb 2006, four years |
| 1M | 31 July 2001 | |
| 2M | 16 October 2018 | |
| 4M | 19 October 2022 | |
| 1.5M | 18 February 2025 |
The 30-year gap is the one that catches people: 1 455 days with no long bond on the curve, in the middle of a series that otherwise runs unbroken. Here those days simply have no 30-year row.
Nothing is interpolated. A curve published without a 20-year point is published here without one.
Using it
import polars as pl
pit = pl.read_parquet("data/pit/*.parquet")
# The 10-year constant maturity, as it was actually knowable
ten = pit.filter(pl.col("entity_id") == "UST_CMT_10Y").select(
"event_date", "knowledge_date", "actual", "change"
)
# Every 10-year note auction, at the minute its result was released
auctions = pit.filter(pl.col("entity_id") == "UST_AUCTION_10Y_BID_TO_COVER")
# Everything a strategy could have known at 4pm Eastern on a given day
known = pit.filter(
pl.col("knowledge_date") <= pl.lit("2026-09-10T20:00:00").str.to_datetime(time_zone="UTC")
)
That last filter is the one that matters. Run it against 4:00 p.m. Eastern and the day's auction results are there — they came out at 1:02 p.m. — but the day's yield curve is correctly absent, because it was not on the Treasury's site until six.
What is in it
| Config | Rows | What one row is |
|---|---|---|
pit |
268 960 | one number becoming public — 1 287 series, plus a Delta table |
auction_results |
11 108 | one completed auction, at the minute its result was released |
auction_calendar |
11 114 | one auctioned or announced security, dated from its announcement |
auction_announcements |
11 114 | one announcement, with its competitive and noncompetitive closing times |
nominal_yield_curve |
113 836 | one maturity of one day's par yield curve, 1990→ |
real_yield_curve |
34 141 | one maturity of one day's TIPS par real curve, 2003→ |
bill_rates |
26 453 | one bill maturity of one day, on both quote conventions |
curve_features |
9 180 | derived slopes and one spread, per trading day |
Auctions split 7 396 bills, 1 967 notes, 934 cash management bills, 394 bonds, 269 TIPS and 154 floating rate notes. Six are announced but not yet held.
pit keys auction metrics by term rather than by CUSIP, so
UST_AUCTION_10Y_BID_TO_COVER runs across issues instead of stopping when one
bond is auctioned. Curve points are UST_CMT_10Y, UST_REAL_CMT_10Y,
UST_BILL_13W_CE; a bill becomes two series because it is quoted two ways.
One limitation worth knowing: in 1984-85 the Treasury occasionally auctioned two
securities of the same odd term on the same day, and with no artifact to give
them different minutes, 12 pit rows share an entity, a date and an instant.
record_id keeps them apart, and the gate reports them rather than hiding them.
Derived, and labelled as such
curve_features is arithmetic on two numbers that were public at the same
instant, so it inherits that instant. Where a leg is missing — no 20-year point
between 2002 and 2020, no real curve before 2003 — the feature is null rather
than interpolated.
One caution on nominal_10y_minus_real_10y_bp: it is a spread between two
official curves published for the same day. It is not a market breakeven.
Both legs are par yields struck from indicative quotations, and neither is a
traded inflation instrument.
In auction_results, dealer_accepted_pct, direct_accepted_pct,
indirect_accepted_pct, noncompetitive_accepted_pct and
tendered_to_offering are computed only from numbers in the same result, and
are null wherever either side is absent. In pit they carry
value_method = "derived"; everything else is "reported".
Provenance
Every row carries source_url and, where an artifact was read, raw_sha256 —
the digest of the exact bytes the Treasury served. Nothing here comes from FRED,
a vendor, or a mirror: a secondary copy of a Treasury rate cannot carry the one
thing this dataset exists to record.
Source: the United States Department of the Treasury, a US Government work in the public domain. The Department does not endorse, certify or verify this dataset. The additions here — the release instants, the long-form curves, the validation — are Apache-2.0.
Rebuilding it
The first build reads 6 229 auction artifacts and 185 rate files, and takes about thirty-five minutes. Every run after that reuses them — a finished auction's result cannot change, and neither can a year of rates that is over — and finishes in under three seconds, having asked the Treasury for seven documents: the auction query, the announced list, and the current year of each of the five rate feeds.
python -m recipe.cli build # incremental
python -m recipe.cli build --rebuild # read everything again; what a parser change needs
The two produce identical tables; tests/test_auctions.py covers which rows
are eligible for reuse and why a date-precision row is not.