U.S. Sahm-Rule Recession Indicator

USSAHM is the published Sahm-rule recession-indicator gap recorded without transformation, the difference in percentage points between the three-month moving average of the seasonally adjusted unemployment rate and the lowest value that same average reached over the previous twelve months. KRED serves the gap verbatim and applies no threshold to it, so the object on this page is a continuous measurement in percentage points and neither a probability nor a forecast.

The unemployment rate underneath the gap is derived from a standard labor-force framework (Hussmanns, Mehran, and Verma 1990), which classifies each person of working age as employed, unemployed, or outside the labor force before dividing the unemployed by the labor force, and that stock is itself the net outcome of continuous flows of entry, exit, and spell duration (Clark and Summers 1979).

Because the reference minimum excludes the current month, the gap turns negative whenever the moving average sets a fresh twelve-month low, a reading near zero says the unemployment trend sits at or near that floor, and a rising reading says the labor market is moving away from it.

Within the cyclical-indicator tradition the unemployment rate is one of the comoving series whose turning points cluster around aggregate reference cycles (Burns and Mitchell 1946) and is classified as lagging the cycle (Mitchell and Burns 1938). Its turning points are dated by reproducible algorithm (Bry and Boschan 1971), the series is placed within the diffusion- and composite-index apparatus (Moore 1961), labor aggregates are recast as observable shadows of a latent cyclical state (Stock and Watson 1989; Stock and Watson 1991), the measurement and timing properties of that apparatus are catalogued (Zarnowitz 1992), and the monthly reading is treated as a current-conditions signal (Shiskin 1961), though such atheoretical measurement acquires meaning only against an explicit model of the labor market (Koopmans 1947).

KRED applies no transformation to USSAHM and stores the published gap exactly as recorded, performing none of rescaling, deflating, smoothing, or annualizing, and applying no trigger threshold, so the object served is the continuous gap rather than any binary rule. The three-month averaging, the trailing twelve-month minimum, and the seasonal adjustment of the underlying rate all belong to the compilation at source, which is why a raw passthrough and a smoothed, seasonally adjusted input hold together without contradiction for the same monthly figure.

The measurement basis is the household labor-force survey, in which each respondent's answers are translated into a discrete activity status under the operational definitions of the active population and of unemployment (Hussmanns, Mehran, and Verma 1990), and the flow accounting that links those statuses across months decomposes the change in the stock into entry, exit, and duration (Clark and Summers 1979).

The construction belongs to the lineage of rule-based monthly cyclical signals, where a fixed arithmetic recipe is preferred to judgment so that the reading can be reproduced from the data alone (Shiskin 1961; Bry and Boschan 1971). Its placement among diffusion and composite indicators follows the codified apparatus (Moore 1961), the detrending and standardizing of activity series into composites precedes it (Persons 1923), and the single-index dynamic-factor representation formalizes the same idea (Stock and Watson 1989; Stock and Watson 1991).

Because the trailing window uses only earlier months of the same moving average, each observation is a function of past and current data alone, while the seasonal factors behind the underlying rate are re-estimated at source and restate earlier months, so successive vintages of the gap can differ. The cyclical timing of the resulting series is documented in the indicator literature (Burns and Mitchell 1946; Mitchell and Burns 1938; Zarnowitz 1992), and the caveat that a measurement procedure presupposes a model for its interpretation remains in force (Koopmans 1947).

USSAHM measures labor-market deterioration against the market's own recent floor rather than against a fixed reference level, which is what makes it readable in real time. The unemployment rate it is built from lags the cycle (Mitchell and Burns 1938) and belongs to the comoving set whose turning points cluster around aggregate reference cycles (Burns and Mitchell 1946), so a widening gap describes a labor market that has already turned rather than one expected to turn.

The stock behind the gap is the net outcome of entry, exit, and spell duration, so a rise can come from more separations or from longer spells (Clark and Summers 1979), and the survey definitions of activity, availability, and search fix what counts as either (Hussmanns, Mehran, and Verma 1990).

Read against the indicator apparatus, the gauge is one input among the diffusion and composite measures (Moore 1961), its turning points can be dated by the same reproducible algorithm applied to any activity series (Bry and Boschan 1971), and its cyclical timing sits in the catalogue of indicator properties (Zarnowitz 1992). Latent-state representations treat labor aggregates as noisy observations on a common cycle (Stock and Watson 1989; Stock and Watson 1991), the monthly cadence supports a current-conditions reading (Shiskin 1961), and the composite tradition supplies the detrended, standardized comparison basis (Persons 1923).

The gauge is descriptive throughout. It states how far the unemployment trend sits above its own twelve-month floor in percentage points, it carries no probability and no horizon, and the atheoretical character of such a measurement means its interpretation depends on an explicit model of the labor market (Koopmans 1947).

On desks the gap functions as a labor-stress level, and its use is bounded by what it measures. A reading near zero places the unemployment trend at its twelve-month floor and a widening reading places it above that floor, and because the series lags the cycle (Mitchell and Burns 1938) it confirms rather than anticipates a turn dated by the reference-cycle apparatus (Burns and Mitchell 1946; Moore 1961).

Release-day interpretation follows the monthly current-conditions tradition (Shiskin 1961) and the reproducible turning-point mechanics that keep successive readings comparable (Bry and Boschan 1971), while the catalogued timing properties of labor indicators set expectations for lead and lag against output measures (Zarnowitz 1992). Single-index representations let the same reading be folded into one cycle variable alongside monthly activity series (Stock and Watson 1989; Stock and Watson 1991).

Two operational cautions apply. The moving-average and trailing-minimum construction makes consecutive observations overlap, so month-to-month moves in the gap are smoother than in the underlying rate and a single print carries less independent information than its variation suggests, a point that follows from the flow structure of the stock it summarizes (Clark and Summers 1979). Seasonal factors and survey definitions behind the underlying rate are maintained at source and can restate earlier months, so vintages differ and cross-country level comparisons require the same operational definitions on both sides (Hussmanns, Mehran, and Verma 1990; Persons 1923).

The gauge is descriptive and carries no probability, so a widening gap is a statement about the labor market now rather than a forecast, and the interpretation of any such measurement rests on an explicit model (Koopmans 1947).

USSAHM is the published recession-indicator gap, the three-month moving average of the unemployment rate minus the lowest value that same average reached over the preceding twelve months, and that published level is the statistical object. Because a one-sided moving average measured against its own trailing minimum is range-bounded by the window it looks back over, its mean reversion follows from the construction rather than from the behavior of the labor market, so no order of integration is assigned to the series here and no such property is reported as an empirical finding (Sahm 2019; Hamilton 2018). The battery is accordingly the persistence, amplitude, and real-time-revision descriptives that the construction admits.

Persistence and amplitude summarize the level. The first-order autoregressive coefficient is 0.964 over 927 monthly observations from 1949-03 to 2026-05, a half-life of about 18.8 months, reported as a descriptive property and read with the small-sample downward bias of Andrews (1993) in mind. The persistence so measured is a property of the overlapping construction rather than of a smoother gain (Stock and Watson 1998). A level portmanteau such as Ljung and Box (1978), refining the form of Box and Pierce (1970), is not reported, because on overlapping three-month averages it would measure the smoothing of the window rather than the process beneath it. The gap has a standard deviation of 1.01 percentage points and ranges from −0.4 to 9.43 percentage points, an amplitude that fixes the scale on which any movement in the level is read.

The multiple structural-break procedure of Bai and Perron (1998), computed by the dynamic-programming algorithm of Bai and Perron (2003) with the number of breaks chosen by information criterion (Yao 1988; Liu, Wu, and Zidek 1997), locates two breaks in the mean of the level at 1983-07 and 2008-01. That search is descriptive and carries no discovery status, since the dynamics of a constructed gap belong to the transform that produces it and break inference presumes a stable object in the first place (Perron 1989; Hamilton 2018).

Real-time behavior has two channels that have to be stated separately. The trailing window uses only the current and earlier months of the same moving average, so the one-sided transform introduces no revision of its own. The variant served here is nonetheless the current-vintage series, whose earlier months are rewritten whenever the underlying unemployment rate is revised at source, and no vintage panel is maintained against which that rewriting could be quantified, so no revision statistic accompanies this field, a limitation familiar from real-time work on gap measures (Orphanides and van Norden 2002; Edge and Meisenzahl 2011).