olivlaw

From books

The spread between the 10-year and 3-month yields is the best-studied indicator of US recessions, but eight hits remain a small sample, and the 2022–2024 inversion showed that the signal is a probability, not a date.

Paul A. Samuelson · Science and Stocks: Sense and Nonsense · 1966 · Paul A. Samuelson, «Science and Stocks», rubrica din Newsweek, 19 septembrie 1966, p. 92 (redarea română: traducerea noastră)3 minutes read
“Wall Street indexes predicted nine out of the last five recessions! And its mistakes were beauties.”Paul A. Samuelson · Science and Stocks: Sense and Nonsense · 1966 · Paul A. Samuelson, «Science and Stocks», rubrica din Newsweek, 19 septembrie 1966, p. 92 (redarea română: traducerea noastră)

An inverted curve raises the probability of a recession, but an indicator that has hit eight times is not yet a rule, only a small sample.

In 1996 Arturo Estrella and Frederic Mishkin showed, in an issue of the New York Fed’s ‘Current Issues in Economics and Finance’, that the spread between the 10-year and the 3-month US Treasury yields predicts recessions better than other financial or macroeconomic indicators, at horizons of two to six quarters. The spread goes into a probit model, which turns a number into a probability of recession four quarters ahead. The mechanism: the 10-year yield is the average of expected short rates plus a term premium; when the market thinks the central bank will cut rates because the economy is weakening, the spread narrows or turns negative. At the same time, high short rates squeeze the margin of banks, which borrow short and lend long, so credit tightens. The inversions of 2000, 2006–2007 and 2019 were followed by recessions that began in March 2001, December 2007 and February 2020. Recessions are dated by the NBER, with a delay: the December 2007 peak was announced in December 2008, the February 2020 one in June 2020. The limits: eight NBER recessions since 1968 are a small sample, and the indicator has false alarms (1966 is the one most often cited; some authors add 1998). The 2022–2024 inversion, lasting about two years, was not followed, by the end of 2024, by any NBER-dated recession; central-bank asset purchases can distort the term premium. For a forecaster: the spread gives a probability with a wide lag, not a date.

Worked example: how to read a spread, and how strong an indicator with eight hits really is. Estrella and Mishkin (1996, data 1960–1995) publish a table linking the 10-year–3-month spread to the probability of a recession four quarters ahead.

10-year – 3-month spread (percentage points)   Probability of recession 4 quarters ahead
+1.21                                           5%
−0.82                                           50%
−2.18 (average for Q1 1981)                     86.5%

Reading it: at a spread of +1.21 the model gives 5%; at −0.82, 50%; and the −2.18 average of the first quarter of 1981 announced 86.5% for the first quarter of 1982, which was indeed in recession (July 1981 – November 1982). The spread is computed every day from two public yields: the example works at any date.

Now the sample. NBER recessions since 1968 number eight: 1969–1970, 1973–1975, 1980, 1981–1982, 1990–1991, 2001, 2007–2009 and 2020. Suppose, generously, that the indicator caught all of them. If the true hit rate were p, the probability of seeing eight out of eight is p to the power 8. How low can p fall before eight out of eight becomes improbable (below 5%)? From p^8 = 0.05 we get p = 0.05^(1/8): the natural logarithm of 0.05 is −2.996, divided by 8 gives −0.3745, and its exponential is 0.688. Conclusion: eight hits out of eight are compatible with a true rate of about 69%, and the false alarms (1966, 1998 according to some authors, 2022–2024 on data to the end of 2024) are not in this calculation. A small sample does not become large because it got lucky eight times.

For the forecaster: base rate and indicators. The reference class is the list of NBER-dated recessions: between 1948 and 2020 there are 12 recession starts (1948, 1953, 1957, 1960, 1969, 1973, 1980, 1981, 1990, 2001, 2007, 2020), about one every six years on average. Over a 12-month window the crude base rate is about 12 / 73 ≈ 16%; it is a starting point that the spread moves up or down, not a current forecast.

Indicators to watch: the 10-year–3-month spread, computed from US Treasury yields published by the Federal Reserve (the H.15 series, redistributed by FRED); the recession probability published monthly by the New York Fed on the basis of the spread; initial claims for unemployment insurance (Department of Labor); announcements of the NBER Business Cycle Dating Committee.

The trap: confusing inversion with the start of the recession. The interval between the first inversion and the cycle peak has ranged from a few months to almost two years, and inversions have also been followed by periods with no recession. A second trap: NBER dating comes late, sometimes half a year after the peak, so a question tied to it must have a wide resolution window.

A scorable question (illustrative): ‘Will the daily 10-year–3-month spread (the T10Y3M series on FRED) be negative on at least one business day between 1 January and 31 December 2027?’ Resolution source: the T10Y3M series on FRED, values available on 15 January 2028; days with no value are not counted.

Why it matters The 10-year–3-month spread stayed negative for about two years, from autumn 2022 into 2024, the longest inversion in decades, with no NBER recession by the end of 2024: an indicator with eight hits was no longer a verdict.

Positive spread:normal curveNegative spread:inverted curve
The further the spread falls below zero, the higher the 12-month recession probability.

← Back to the feed