Margin calibration study
The published margin, checked against 176 month-to-month comparisons since 2012: how often conditions held at each distance from the line where they change.
Across 176 month-to-month comparisons, conditions were the same the following month 169 times, or 96%. Every band below held more often than the percent published for it, and the two closest bands run in the wrong order. The margin is therefore a distance, not a probability. It says how much room there is before the line where conditions change; the count beside it says how often that room mattered.
| Points from the line | Published % | Held | Months |
|---|---|---|---|
| 0 to 5 | 55% | 100% | 11 |
| 5 to 10 | 60% | 70% | 10 |
| 10 to 20 | 65% | 100% | 12 |
| 20 to 37 | 75% | 94% | 54 |
| 37 or more | 92% | 99% | 89 |
The finding holds in the recent year too. Across the most recent twelve month-end readings, every band with a comparison still held more often than the percent said: 2 of 3 in the 5 to 10 band against a published 60%, and 7 of 8 in the 20 to 37 band against a published 75%.
Derived from the month-end archive each time this page is built, so it improves as the record grows. A held month means conditions did not change. Conditions change only after a score has crossed its line and stayed there for a second reading.
Re-scored on Oct 9, 2026, when the Chicago Fed financial conditions index joined the inputs the record is scored on, the same inputs that now set the guidance. No month's conditions changed, so the count above is the one published before it.
How it was measured
The readings counted, the outcome scored and the distance bands.
- Sample. Every month-end reading in the archive that has a following month to compare against — 176 of them. The record holds 177 readings, Jan 2012 to Sep 2026; the newest has no following month yet.
- Outcome. Whether conditions were the same one month later. No other outcome is scored, and none is weighted.
- Bands. Distance to the nearest line where conditions change, in points of funding pressure, with a risk-appetite point converted the same way the published margin converts it. The cut points were chosen before the outcomes were looked at.
What it does not show
The limits are the reason the number was not simply re-fitted.
- 7 changes in conditions since Jan 2012 is not enough to estimate a probability per band. A curve fitted to that base rate would report about 97% every week and discriminate nothing. That is worse than a number that at least measures something real.
- One month is the only horizon tested. Nothing here says what a reading is worth over a quarter.
- The archive is one model over one country’s rates. It is evidence about this number, not about forecasting scores generally.
So the margin keeps measuring distance, and distance is an honest thing to measure. This record is published beside it rather than folded into it.
Sources
- Rates: St. Louis Fed (FRED) Treasury series (DGS1MO, DGS3MO, DGS2, DGS10).
- Archive and method: how the scores are computed.