When a Yield Curve Un-Inverts, the Reason Matters More Than the Sign

A Sign Change Is Not a State Change

The spread between the 10-year and 2-year Treasury yields stood at +0.50 percentage points on 20 August 2026 — the 10-year at 4.69%, the 2-year at 4.19%, per the Federal Reserve's H.15 release. That is a normally sloped curve, and every monthly average from August 2025 onward is positive as well. Somewhere between April 2024, when the monthly average sat at −0.33 points, and September 2024, when it printed +0.10, the sign flipped.

The narrative attached to that fact runs like this: the inversion was the warning, the un-inversion is the all-clear, and the recession the inversion advertised either arrived quietly or was called off. The problem with that reading is not that it is optimistic. The problem is that it treats one number as one piece of information. The spread is a difference between two independently traded prices, and two prices moving in opposite directions, or in the same direction at different speeds, can produce identical spread changes for opposite reasons.

A separate piece on this site argued that an inversion does not guarantee a recession in the following quarter. This is the mirror-image claim, and it needs its own argument rather than a restatement of that one. The claim here is narrow and mechanical: un-inversion carries almost no information until the move is decomposed into which leg did the work. The same +0.30 point widening can mean the front end collapsed or the long end sold off. Those are not variations on a theme. They are different macro states that share an arithmetic signature.

Four Ways to Get the Same Number

The spread is 10-year minus 2-year. That leaves exactly four ways for it to change, and market convention has a name for each:

• Bull steepening — both yields fall, the 2-year faster, the spread widens. Usually read as the front end pricing a lower future policy path.
• Bear steepening — both yields rise, the 10-year faster, the spread widens. Usually read as term premium, issuance absorption or long-horizon inflation compensation rising.
• Bull flattening — both yields fall, the 10-year faster, the spread narrows. Often a duration bid or a downgrade to long-run growth.
• Bear flattening — both yields rise, the 2-year faster, the spread narrows. The front end repricing the near-term path higher.

Two of those widen the curve and two flatten it, and the two that widen it point in nearly opposite directions. An analyst tracking only the spread series sees one line going up in both cases. That is the argument of this piece in a sentence; the rest is the arithmetic behind it.

The recent record, monthly averages of the constant-maturity series:

Month	10-year	2-year	Spread
2025-08	4.26	3.70	+0.56
2025-09	4.12	3.57	+0.55
2025-10	4.06	3.52	+0.54
2025-11	4.09	3.55	+0.54
2025-12	4.14	3.50	+0.64
2026-01	4.21	3.54	+0.67
2026-02	4.13	3.47	+0.66
2026-03	4.25	3.71	+0.54
2026-04	4.32	3.80	+0.52
2026-05	4.48	4.00	+0.48
2026-06	4.47	4.11	+0.36

The spread never left a 31 basis point band across those eleven months — from +0.36 at the low to +0.67 at the high. Anyone monitoring only that column would conclude nothing much happened. Underneath, the 2-year travelled from 3.47% to 4.11%, 64 basis points in four months, while the 10-year moved from 4.06% to 4.48%. The stability of the difference concealed a great deal of motion in both terms.

The decomposition, four episodes at a time

The diagram below takes four windows from the last two years and splits each spread move into its components. The bars are changes in yield, not levels; bars left of the axis are yields falling.

Rows one and four are the point. Both widened the spread. In the first, both yields fell and the front end fell 43 basis points faster. In the fourth, both rose and the long end rose 14 faster. A spread chart records these as one event.

What the 2024 Un-Inversion Actually Was

Between April 2024 and September 2024 the 2-year fell from 4.87% to 3.62% — 125 basis points — while the 10-year fell from 4.54% to 3.72%, or 82 basis points. The spread gained 43 basis points and crossed zero. But the mechanism was the front end collapsing, not the long end rising.

That distinction matters because the front end is, to a first approximation, a discounted path of expected short rates over the next two years. A 125 basis point decline in that estimate is the market marking down where it expects policy to sit over eight quarters. That kind of repricing is not a signal that stress has passed; it is a signal that the market expects a response to something. Un-inversion by bull steepening is mechanically the same event as an easing cycle being priced in, and easing cycles are priced in when the near-term outlook deteriorates.

None of which forecasts anything. The point is that un-inversion through a front-end collapse and un-inversion through a long-end selloff describe different worlds, and "the curve normalized" covers both.

The 2026 Round Trip, and Where It Left the Curve

The recent record inverts the 2024 mechanism. From December 2025 to June 2026 the 2-year rose 61 basis points, 3.50% to 4.11%, while the 10-year rose 33, 4.14% to 4.47%. The spread narrowed 28 basis points — +0.64 to +0.36 — and the narrowing was entirely a front-end story. The January-to-June window is starker: the 2-year up 57 against the 10-year's 26, spread down 31.

Then the mechanism reversed again. Between the June average and the daily H.15 print for 20 August 2026, the 2-year added 8 basis points to 4.19% while the 10-year added 22 to 4.69%. The spread widened from +0.36 back to +0.50, and this time the long end did the work. The chart below plots both series over the last twelve months; the divergence between the legs is easier to see there than in any table.

Read the chart from the two lines rather than the gap between them. The spread ends at +0.50, close to the +0.56 of a year earlier, which suggests a market that went nowhere. The levels say otherwise: both yields sit higher than twelve months ago, and the path between contained one widening driven from the front, one narrowing driven from the front, and one widening driven from the back. Three macro stories, one nearly flat spread series.

Scenario A — Front End Falls, Curve Widens From Below

Here the 2-year retraces its 2026 rise while the 10-year holds or falls more slowly. It is the 2024 pattern repeating, and the second half of 2025 in miniature: the 2-year fell 20 basis points against the 10-year's 12, widening the spread by 8 with both yields lower.

What would have to show up. A 2-year back below roughly 4.00% while the 10-year stays above 4.50% puts the spread at +0.50 or wider, with both legs below the 20 August prints. Confirming evidence: a soft sequence in labour market data — the monthly employment report, the JOLTS openings series — with the front end moving on the release rather than drifting.

What falsifies it. A 2-year that stays pinned near 4.15%–4.25% through several data releases. If the front end will not move on soft data, the market is not repricing the near-term path, and any widening is coming from somewhere else.

Weight. This branch matches the pattern that produced the original un-inversion, which is why it deserves scepticism rather than deference. The front end spent the first half of 2026 moving the other way — up 61 basis points — and reversing that takes new information, not the absence of it.

Scenario B — Long End Rises, Curve Widens From Above

Here the 10-year continues the move it has made since June and the widening happens with both yields higher — the most recent leg, +22 basis points on the 10-year against +8 on the 2-year.

What would have to show up. A 10-year clearing roughly 4.80% while the 2-year stays under 4.25% puts the spread above +0.55 with the long end doing all of it. The supporting evidence is structural rather than cyclical: absorption of coupon supply at auction, the tail between the stop and the pre-auction level, the composition of bidding. Quarterly refunding announcements, published in the opening week of February, May, August and November, set the issuance profile the long end has to digest.

What falsifies it. The 10-year stalling below 4.70% while the 2-year grinds higher. That converts the picture into Scenario C without the spread ever moving much.

Weight. This branch separates "the curve normalized" from "conditions improved" most cleanly. A curve steepening from the long end while the front end sits still is not the market pricing relief. It is the market demanding more compensation to hold duration, which can coexist with a deteriorating outlook rather than contradicting it.

Scenario C — Front End Rises Again, Curve Re-Flattens

The third branch is the December-2025-to-June-2026 pattern resuming: the 2-year rising faster than the 10-year, the spread compressing toward zero, and at the limit re-inversion.

What would have to show up. A 2-year above roughly 4.40% with the 10-year near 4.70% puts the spread around +0.30, below the +0.36 June low. If the 2-year crossed the 10-year again, the curve would have run inversion to normalization to inversion inside two years — itself a statement about how little the sign carries. The inputs are the monthly CPI release and the PCE price index, both feeding the near-term path the 2-year discounts.

What falsifies it. Two consecutive months in which the 2-year is flat or lower while the 10-year rises. That is Scenario B by another route.

Weight. This branch has the most recent precedent, being what actually happened in the first half of 2026. Precedent is worth something, though not much on a six-month sample.

Reading the branch from the data rather than the narrative

The test is mechanical and runs on two numbers from any H.15 release. The decision path below is the entire method.

The Two-Leg Decomposition Breaks When Both Ends Move for the Same Reason

This frame has real limits. Four are worth stating plainly.

The clean separation between "policy path" and "term premium" is an assumption, not an observation. The 10-year embeds roughly the same near-term path expectations the 2-year does, plus eight more years of them, plus compensation for uncertainty. When the near-term path shifts hard enough, both legs move together and the decomposition credits the long end with something that originated at the front. The Dec-2025-to-Jun-2026 window is a candidate: the 10-year's 33 basis point rise may be mostly the front-end move propagating outward rather than an independent term premium story.

Monthly averages smooth away the moment a mechanism changed. Every number in the table is an average of daily observations, so a 10-basis-point intramonth reversal and a flat month look identical. The June-to-August comparison is worse, setting a monthly average beside a single daily print. Its +0.14 widening is real at both endpoints, but it is not a like-for-like measurement.

Two points is a thin description of a curve. The 2s10s spread ignores the 3-month bill, the 5-year and the 30-year, and there are configurations in which 2s10s and 3m10y disagree about the sign. Nothing here establishes that the 2-year is the right front-end reference; it is the conventional one, which is a different claim.

And the honest answer on the direction is that it is not yet known. A long-end widening of roughly two months stands against a front-end flattening that ran five, February through June. Two months is not enough to show the mechanism changed rather than paused.

What to Watch Next Week

• The two legs separately, not the spread. The H.15 release publishes the constant-maturity 2-year and 10-year each business day. Record both levels, not the difference. Reference point, 20 August: 4.19% and 4.69%.
• The end-of-month coupon cycle. Treasury normally auctions 2-year, 5-year and 7-year notes in the final week of the month. Those price the front and belly, so weak demand there is a front-end signal, not a duration signal. The long-end read comes from the mid-month 10-year and 30-year operations.
• The monthly PCE price index, normally published toward month end. Watch which leg reacts. A front-end-only reaction is Scenario A or C depending on direction; a long-end reaction to an inflation print is closer to Scenario B.
• The +0.36 and +0.67 boundaries — the June low and January high on monthly averages. A daily spread outside that band is the first evidence the eleven-month range has resolved.
• Whether the 2-year moves at all. The most informative observation next week is a 2-year holding inside a 5 basis point range across several releases. That is evidence the front end has stopped repricing, which pushes weight toward Scenario B by elimination.

Concrete Framework

1. Log both yields weekly from H.15. Two columns, one row per week. The spread is derived, never recorded directly.
2. Compute two changes, not one. Per window, the change in the 2-year and the change in the 10-year, separately. The spread change is their difference — the third number, not the first.
3. Assign the direction label. Both lower and front faster is bull steepening. Both higher and long faster is bear steepening. Both higher and front faster is bear flattening. Both lower and long faster is bull flattening.
4. Check the magnitude against the noise floor. A move under roughly 10 basis points on either leg over a month sits inside normal variation. In the table above, seven of the ten month-to-month steps moved the spread less than 5 basis points.
5. Mark the scenario only when the same label repeats. One month is a print. Two consecutive months with the same label is a pattern. Measured end to end, December 2025 to June 2026 was a bear flattening; measured step by step, the same label ran from February through May.
6. Write down what falsifies the current label before the next release. Scenario B: a 10-year stalling under 4.70% while the 2-year rises. Scenario A: a 2-year pinned near 4.20% through soft data. Scenario C: two months of a flat-or-lower 2-year.
7. Refuse the summary sentence. "The curve normalized" describes two different underlying states. Any note reporting the spread without reporting which leg moved has discarded most of its information.

The sign of the 2s10s spread was never the signal. It compressed the signal into one bit, and one bit cannot distinguish a front end pricing a downturn from a long end pricing a supply problem. Which of the three mechanisms is running now is answerable from public data, weekly, with subtraction — and it is a different question from whether the curve is positively sloped.