Methodology
Treasury Noise Methodology
Treasury noise measures how far Treasury yields stray from a smooth curve. It rises when arbitrage capital is scarce, which makes it a market-wide illiquidity indicator. Safe Rate rebuilds it every business day from Treasury's end-of-day prices. From 2008 to 2025 our series correlates 98.9% with the authors' own, about as closely as their own alternative curve methods matched it.
Where it comes from
The measure is Grace Xing Hu, Jun Pan and Jiang Wang's, from "Noise as Information for Illiquidity," The Journal of Finance 68(6), December 2013, pages 2341 to 2382. The paper and the authors' daily series are published by Jun Pan at en.saif.sjtu.edu.cn/junpan. The series we serve as source=authors is theirs, unchanged, and is credited to that page wherever it is shown.
Their reasoning: in normal times arbitrage capital keeps Treasury yields close to a smooth curve. In a crisis that capital is withdrawn, and yields scatter. The scatter, measured across the whole curve, is information about liquidity in the broader market, not just in Treasuries.
How it is built
- Each business day, take every bill, note and bond with one month to ten years left. Leave out callable, inflation-indexed and floating-rate securities.
- Fit a Svensson zero-coupon curve to their prices, with price errors weighted by inverse duration so that they behave like yield errors. We publish two versions of this curve, defined below.
- For every note and bond with one to ten years left, take its market yield minus the yield of the same cashflows priced on the curve.
- Leave out any bond more than four standard deviations from the curve.
- Noise is the root mean square of what remains, in basis points.
Each day is fitted on its own, so one day's value never depends on another's. Prices are Treasury's own end-of-day marks; see Our prices, below.
The two curves
Noise is the scatter of bond yields around a fitted curve, so the curve is part of the measure. Both curves are Svensson zero-coupon curves: a six-parameter forward curve, turned into discount factors and fitted so that each bond's price comes out as close as possible to the market. The same bonds and the same weights go into both. They differ only in whether the fit is constrained:
svensson(the default): the fit with no constraints. The curve bends wherever the prices pull it. This is exactly Hu, Pan & Wang's construction, and the authors' own series uses it.safe-rate: the same form regularised, as Safe Rate's published Treasury curve is. Penalties on roughness and on large coefficients keep the curve sensible between and beyond the securities, which is what a published curve is read for.
Where this page says "our series" without naming a curve, it means svensson.
Our prices
We price the measure on Treasury's own end-of-day prices, the FedInvest file the Bureau of the Fiscal Service publishes on TreasuryDirect each business day for every marketable Treasury. It is a public, U.S. government source, which is why we can publish a series built from it and extend it every day. The authors used daily dealer quotes, which are licensed.
Treasury publishes one end-of-day price per security and no methodology for it. We checked it against an independent dealer-quote source, held as a referee, over 1.6 million matched quotes from 2008 to 2025. On average it sits within half a cent of the dealers' bid for bills and notes, about half a spread below the mid, and its typical disagreement with the dealers is inside their own quoted bid-ask spread. The paper uses the mid, but a shift common to every bond moves the fitted curve, not the scatter around it, so the noise is unaffected.
To isolate the price source, we built the measure both ways, with the same method on the same days: once on Treasury's prices and once on dealer quotes. Svensson curve, September 2008 to December 2025:
| Treasury end of day against dealer quotes | Value |
|---|---|
| Level correlation | 99.2% |
| Month-end change correlation | 89.7% |
| Monthly-average change correlation | 97.2% |
| Daily change correlation | 31.9% |
| Mean noise, bp | 2.27 against 2.34 |
Over months the two price sources give nearly the same measure. Day to day they part: daily changes correlate only 32%. Treasury's marks move more smoothly than dealer quotes, but their daily moves do not track the dealers' daily moves closely. That is the main cost of the public source, and the reason to read the series monthly or smoothed.
How closely it matches the authors' series
Correlation with the authors' published series over the 4,332 business days both cover, September 2, 2008 to December 31, 2025, measured November 17, 2026. The authors' own mean over the window is 2.47 basis points.
| Prices | Curve | Level | Month-end change | Monthly-average change | Daily change | Mean, bp |
|---|---|---|---|---|---|---|
| Dealer quotes | svensson | 99.3% | 88.1% | 96.5% | 59.6% | 2.34 |
| Dealer quotes | safe-rate | 98.9% | 79.1% | 95.3% | 53.7% | 2.57 |
| Treasury end of day | svensson (default) | 98.9% | 85.6% | 94.5% | 24.6% | 2.27 |
| Treasury end of day | safe-rate | 98.2% | 82.8% | 91.8% | 17.5% | 2.52 |
The yardstick is the authors' own. In their appendix they rebuild the measure on cubic splines instead of Svensson, and the rebuilds correlate 97.8% to 99.0% in levels and 83.3% to 91.9% in monthly changes with their original. Our default series sits inside both ranges.
Read down the table for the price source and across the pairs for the curve. On either price source, the svensson curve tracks the authors more closely than safe-rate. On either curve, dealer quotes track them more closely day to day, which is the authors' own kind of data; over months the two sources are close.
Use it monthly or smoothed, not day to day. Daily changes on our series agree with the authors' only 25% of the time, against 60% on dealer quotes. Day to day, much of the movement is the prices, not the market. The 20-day mean in the API is there for this reason.
Where ours is better, and where it is worse
Better: it has no false spike on June 3, 2024. The authors flag that day themselves: their outlier filter missed two bonds, four Treasuries issued in 1995 to 1997, and the measure jumped to 4.31 basis points. On dealer quotes we reproduce the spike exactly, from the same bonds. Treasury's end-of-day prices put those bonds closer to the curve, and our series reads 1.52. Z-scored on its own past year, the authors' June 3 reads 2.38, enough to trip an alert at 2. Ours reads −1.40.
It is also the cleanest of the five series we scored. Its moves persist rather than reverse, and it has the fewest one-day blips (a blip is a day far above both its neighbors):
| Prices | Curve | Change autocorrelation | 20-day variance ratio | Blips a year |
|---|---|---|---|---|
| Authors' published series | svensson | −0.285 | 0.439 | 3.4 |
| Dealer quotes | svensson | −0.309 | 0.476 | 5.1 |
| Dealer quotes | safe-rate | −0.313 | 0.416 | 4.5 |
| Treasury end of day | svensson | −0.284 | 0.663 | 3.3 |
| Treasury end of day | safe-rate | −0.367 | 0.338 | 4.9 |
An autocorrelation nearer zero and a higher variance ratio both mean that less of the day-to-day movement is jitter that reverses the next day.
Worse: June 14 to 16, 2022, the week of the 75 basis point rate hike, is a spike in our series only: 5.77 basis points, against 3.03 for the authors and 3.11 on dealer quotes. It is not a few stray bonds. Dozens sit off the curve in a pattern across maturities, and we have not determined why. Our series also reads about 20% lower than the authors' through 2024, mostly because of the same old bonds.
The same as the authors': in quiet years, 2011 to 2021, both our series and the dealer-quote rebuild read 5% to 17% below the authors'. In stress years they match within 2% to 3%. We have not explained the quiet-year gap.
Which curve to use
svensson, the default, unless you are testing the choice. The safe-rate series matches the authors' average level more closely, but it matches their monthly changes less well and is the least clean of the five series. In 2022 and 2023 it reads 15% to 27% too high. The likely reason is that its penalties stop the curve bending into a sharply inverted shape, so curve shape shows up as noise. The constraints that make a published curve trustworthy between securities cost accuracy at the securities, and accuracy at the securities is all this measure reads.
What it signals, and what it does not
Noise rises with market stress rather than ahead of it, in every series we tested, the authors' included. We z-scored each series against its own past and checked when it crossed 2 against the equity drawdowns since 2009. In every case the median alarm came after the drawdown had begun. That matches the paper, which presents noise as a measure of illiquidity, not a forecast.
Coincident is still useful. A rule that cuts risk as noise rises can avoid much of a long drawdown, such as 2008, 2020 or 2022. It cannot avoid a short, sharp one. Several alarms that look false against equity drawdowns fall on real Treasury stress: the 2013 taper tantrum, the October 2014 flash rally, December 2016 and the 2019 repo spike. The samples are small, 4 to 19 alarms per test, and the differences between series are within chance.
What the API serves
GET /v1/signals/noise, one entry per business day. Each entry carries the noise in basis points, its trailing 20-day mean, and z-scores against the previous one, two and five years. The z-scores exclude the day itself and are empty until their window fills. Every indicator uses only days up to its own, so it shows what a reader could have known that day. Our entries also carry the noise before the outlier filter and the number of securities behind it.
source=fedinvest(default): ours, September 2, 2008 to November 16, 2026, oncurve=svensson(default) orcurve=safe-rate.source=authors: the authors' series, January 2, 1987 to December 31, 2025,curve=svenssononly. They publish the noise alone, so its other fields are empty. It has no entries from September 11 to 20, 2001, when the bond market was shut. The authors revise their series: their current file gives 13.31 for October 19, 1987, where their 2011 slides show 13.93. Compare against the series, not the paper's figures.
The series is loaded in batches, not updated daily, and ends on the last date loaded. Full field descriptions are in the API reference.
Sources
- Method and the authors' series: Hu, G. X., J. Pan and J. Wang (2013), "Noise as Information for Illiquidity," The Journal of Finance 68(6), 2341 to 2382, doi:10.1111/jofi.12083. Paper and data: https://en.saif.sjtu.edu.cn/junpan/.
- Prices: U.S. Department of the Treasury, FedInvest end-of-day prices, a U.S. government publication.
- Reference: an independent dealer-quote source for U.S. Treasuries, used for comparison only.
- Curve model: Svensson (1994).