Week of 1 October 2026 · rebuilt from Federal Reserve data every weekday
The U.S. Treasury yield curve, taken apart every week since 1990
On 1 October 2026 the 10-year Treasury zero-coupon yield was 5.25%. About 4.06% of that is where short-term rates are expected to be over the decade, and 1.19% is term premium, the extra yield investors ask for to tie money up for ten years. TIPS prices imply inflation averaging 2.37% a year over the same ten years. The slope of the curve puts the odds of a recession in the next twelve months at 7%.
Latest reading
| Measure | Latest | 1-year change | 5 years |
|---|---|---|---|
| 10-year Treasury yieldpar yield on the fitted curve | 5.27% | +1.05 pp | |
| Curve slope, 10Y − 3Mnormal since 20 March 2026 | +1.12 pp | +0.92 pp | |
| 10-year term premiumanchored to surveys of forecasters; plain ACM +2.06% | +1.19% | +0.50 pp | |
| Expected short rate, next 10 yearsthe rest of the 10-year zero yield | 4.06% | +0.59 pp | |
| 10-year breakeven inflationreal 10-year yield 2.88% | 2.37% | -0.04 pp | |
| 5y5y forward breakevenexpected inflation 5 to 10 years ahead | 2.39% | -0.02 pp | |
| Recession odds, next 12 monthsprobit on the 10Y − 3M spread; forward-spread model 4% | 7% | -7 pts |
Changes are over the past year; hover a sparkline for its range. Term premium and expected rate are monthly, the rest weekly.
Since 1990
How it is checked
Real markets have no answer key, so each model is first run on a simulated market where the true curve, term premium and inflation are known, then compared with independent estimates from the Federal Reserve. From the latest run:
- 1,918 weekly curves since 1990; the median curve misses its quotes by 3.8 basis points.
- 10.6 bp from the Federal Reserve's GSW curve on zero rates from 1 to 30 years.
- Survey-anchored term premium, estimated in real time, against the Fed Board's Kim-Wright estimate: correlation 0.94, typical gap 44 bp.
- Plain ACM term premium, estimated in real time, against the Fed Board's Kim-Wright estimate: correlation 0.08, typical gap 116 bp.
What didn't work
These stay in the documentation, because they are as useful to know as the results that held up.
- No forecasting model beat "yields stay where they are" one month ahead. Mixing a model half-and-half with that guess only ties it at six to twelve months.
- The textbook term premium model (ACM), re-estimated each month on the data available then, jumps around too much to use. Anchoring it to surveys of forecasters fixed most of that; a bias-corrected version made it worse.
- Splitting the slope into expected rates and term premium predicted recessions worse than the plain slope since 2005. The 2022-24 inversion pushed the real-time model to 90% and no recession has followed so far.
- The survey-anchored term premium, the one closest to the Fed Board's, did not predict what bonds went on to earn: forecasters kept expecting rate rises that never came, so it expected bonds to lose to bills while they beat them. A famous bond-return predictor (Cochrane-Piazzesi) did far worse than the plain historical average once re-estimated in real time.
- Version 2.2 said the arbitrage-free model had better-calibrated forecast intervals. A proper coverage test showed they are too wide, and the claim was withdrawn.
The numbers and tests behind each are on the results page.
Where this came from
It started as a sophomore-year script: a Nelson-Siegel curve fitted to FRED data and a 3-D Plotly surface. Each version since has fixed something the previous one got wrong. The biggest was that FRED's yields are par yields, not zero rates, which most Nelson-Siegel code ignores. The changelog lists every mistake and its fix.
Everything runs from a GitHub Action every weekday evening. To run it yourself:
pip install nss-engine
nss-engine run # FRED data since 1990 -> output/dashboard.html, report, CSVsNo API key needed. The README shows the Python API.