The 10-year minus 2-year spread since 1976: every inversion, dated

The question is narrow on purpose: since the 2-year constant-maturity Treasury series began in June 1976, how many times has the 10-year yield closed below the 2-year yield, for how long, how deep, and what happened next. The answer is a table of fourteen dated episodes, computed by the script attached to this page from a CSV you can download. Nothing here forecasts anything; the point is to have the record in one place with its definitions written down.

The data

Three daily series and one monthly indicator, all pulled from FRED on 2026-09-05 with the public CSV endpoint, no API key:

• DGS10 and DGS2: market yields on 10-year and 2-year constant-maturity Treasuries, from the Federal Reserve's H.15 release, in percent.
• T10Y2Y: the spread FRED computes from the Treasury Department's own constant-maturity data (the series notes say the underlying data has come directly from Treasury since June 2019). It is DGS10 minus DGS2 in percentage points, and we use it rather than recomputing the difference so that the figures match what anyone sees on the FRED chart.
• USREC: a 0/1 monthly flag that interprets the NBER's peak and trough dates. The series notes are explicit that the choice of which day inside a month the turn happens is FRED's convention, not the NBER's.

The merged file has 13,396 rows. DGS2 starts on 1976-06-01, so that is where the spread starts, and the last observation is 2026-09-04 with a spread of 0.41 percentage points.

The method

An "inversion episode" needs a definition, and the count depends on it. Ours:

1. A negative close is a day where T10Y2Y is below zero.
2. Consecutive negative closes form an episode. A gap of up to five non-negative closes does not end the episode; a sixth does.
3. Episodes with fewer than five negative closes are counted separately as blips and not tabulated.
4. For each episode, the script records the first and last negative close, the number of negative closes, the deepest value and its date, whether the episode began during an NBER recession, and the months from the first negative close to the next recession start in USREC.

The bridging rule matters. Without it, the 1978 to 1980 inversion would split into a dozen fragments every time the spread ticked to zero for a day. With a much longer bridge, the 2006 and 2007 fragments would merge into one. Five days is a judgment; the parameter sits at the top of the script and you can change it.

neg = spread < 0
episodes, start, last_neg, gap = [], None, None, 0
for d, is_neg in neg.items():
    if is_neg:
        if start is None:
            start = d
        last_neg, gap = d, 0
    elif start is not None:
        gap += 1
        if gap > MAX_GAP:          # MAX_GAP = 5 non-negative closes
            episodes.append((start, last_neg))
            start = None

The table

Output of python code/yield-curve-inversions-since-1976.py on 2026-09-05:

Six blips of fewer than five negative closes were not tabulated: 1981-12-28 (1 close), 1989-11-03 (2), 1998-05-26 (1), 2005-12-27 (3), 2019-08-27 (3), and 2022-04-01 (2). Across the whole sample, 2,049 of 12,563 closes were negative, 16.3 percent. NBER recession starts inside the window: 1980-02, 1981-08, 1990-08, 2001-04, 2008-01, 2020-03.

What the table shows

The 2022 to 2024 episode is the longest in the record by a wide margin: 539 negative closes against 423 for the 1978 to 1980 inversion, and its trough of -1.08 points on 2023-07-03 is the deepest since December 1980. As of the last USREC observation in the file, no recession start has been recorded after it. That is a statement about the data on 2026-09-05, not about the future; the NBER dates turning points with a lag of many months, and the flag can change.

Every recession that began between 1980 and 2008 was preceded by a tabulated episode, and the lead from the first negative close of the nearest preceding episode ranged from 5 months (episode 6, before 1990-08) to 20 months (episode 4, before the same recession). Episode 3 is the odd one: it began in January 1982, inside the 1981 to 1982 recession, and the "next" recession start in the table is 1990, which is why its lead reads 103 months. It is a within-recession inversion, not a warning of anything.

The 2020 recession is the counter-example in the other direction. On the 10-year minus 2-year measure, 2019 produced only a three-close blip on 2019-08-27, below our five-close floor. The recession that followed in March 2020 was pandemic-driven, and reading it as a confirmation of the 2019 inversion asks more of the data than the data can carry.

Depth and duration do not line up with what followed. The 1998 episodes were shallow (troughs of -0.07 and -0.03) and the next recession was 33 months away. The 2006 to 2007 episode was also shallow, never deeper than -0.19, and a severe recession began 19 months after it started. The 1980s episodes were the deepest and preceded the two recessions of 1980 and 1981 to 1982 by 18 and 11 months.

What it does not show

Fourteen episodes and six recessions are a small sample, and several episodes are fragments of one macro period (1988 to 1990 contributes three rows; 2006 to 2007 contributes four). Counting "successes" from a table like this is close to meaningless statistically, and we have not done it.

The definition drives the count. Lower MIN_DAYS to one and the six blips join the table; raise MAX_GAP to twenty and rows 4 to 6 collapse into a single 1988 to 1990 episode. Anyone quoting "the yield curve has inverted N times" should say which N and which rule.

This is one spread. The 10-year minus 3-month spread (T10Y3M on FRED) inverts on a different schedule and is the measure used in much of the recession-probability literature. A 2018 FEDS Note from the Federal Reserve Board, "(Don't Fear) The Yield Curve", argues that a near-term forward spread carries the information and that the long-end spread adds little once it is accounted for. That is an argument about which series to look at, and it is not tested here.

Nothing in this table is a causal claim. An inversion is a market price, and the mechanism connecting it to later output is contested. The dataset and the script are attached so the rule can be changed and the table rerun.