One line of arithmetic on two public series, with a script that recomputes it. It is a gauge of how much of the working-age population is not employed for any reason, not a measure of joblessness.
CNP16OV is the civilian noninstitutional population aged 16 and over; CE16OV is the employment level; EMRATIO is the employment–population ratio FRED already publishes. The script checks the identity on all 945 months since January 1948 (EMRATIO is published to one decimal, so the largest difference is 0.05).
What PEG is not. It is not an unemployment rate and it is not a measure of discouragement or slack. Its population base includes retirees, full-time students and people unable to work. This page says so rather than leaving the reader to infer it, and the arithmetic below shows how little of the gap is unemployment.
What it is for. A public, reproducible, trackable measure, updated with each BLS Employment Situation report (typically the first Friday of the month) for as long as the two FRED series exist. It is a companion to U-3 and U-6, not a replacement for either and not a “more honest” version of them. If it is ever discontinued, a dated note will say so.
| Measure | Aug 2026 | How it is built |
|---|---|---|
| U-3 | 4.1% | unemployed / labour force (FRED UNRATE; recomputed from UNEMPLOY and CLF16OV, same to 0.05) |
| U-6 | 7.7% | FRED U6RATE, as published |
| PEG | 40.91% | (275,415 − 162,746) / 275,415, in thousands |
The gap is made of exactly two things, because employment equals the labour force minus the unemployed: 38.36% of the population is outside the labour force, and 2.55% is unemployed. The unemployed are 6.2% of the gap. “Outside the labour force” is not the same as “by choice”: it also holds people who want a job and are not counted. This page does not split that block by reason (section 6).
| Measure | Aug 2026 | Sep 2026 | How it is built |
|---|---|---|---|
| Population, 16+ (thousands) | 275,415 | 275,554 | BLS Table A-1, seasonally adjusted |
| Employed (thousands) | 162,746 | 163,152 | BLS Table A-1 |
| U-3 | 4.1% | 4.2% | published; 7,109 thousand unemployed |
| U-6 | 7.7% | 7.6% | BLS Table A-15, seasonally adjusted (7.7% in August, as pinned) |
| PEG | 40.91% | 40.79% | (275,554 − 163,152) / 275,554 = 112,402 / 275,554 |
The gap fell by 0.12 points (0.118 unrounded). It splits exactly as before: people outside the labour force went from 38.36% to 38.21% of the population (−0.145) and the unemployed from 2.55% to 2.58% (+0.027). The unemployed are 6.3% of the gap, up from 6.2%. So PEG fell while U-3 rose a tenth: employment grew faster than the population, and the fall came from the part outside the labour force, not from the unemployed. One month, first release, and BLS revises; nothing more is claimed.
Provenance, stated plainly. These September figures are from Table A-1 of the BLS Employment Situation release for September 2026 (embargoed to 8:30 a.m. ET, 2 October 2026), read here through a page-reading tool, not by the script; U-6 is from Table A-15 of the same release (bls.gov/news.release/empsit.t15.htm), whose August value, 7.7%, also matches the pin. The August column in that same table matches this page’s pinned August values exactly (275,415; 162,746; 4.1%; 59.1), which is the only cross-check made. The release gives 59.2% for the employment–population ratio, and 100 − 59.2 = 40.8, which agrees with 40.79 to the published 0.1. In the table, labour force minus unemployed (163,153) differs from employed (163,152) by one thousand, a rounding in BLS’s own table; PEG uses employment directly. FRED could not be reached from where this was written, so wp128-verify.py has not been re-run against September and still pins August. CITED bls.gov/news.release/empsit.nr0.htm.
| Measure | Sep 2026 | Oct 2026 | How it is built |
|---|---|---|---|
| Population, 16+ (thousands) | 275554 | 275682 | FRED CNP16OV |
| Employed (thousands) | 163152 | 163615 | FRED CE16OV |
| U-3 | 4.2% | 4.1% | FRED UNRATE |
| U-6 | 7.6% | 7.9% | FRED U6RATE |
| PEG | 40.79% | 40.65% | (275682 − 163615) / 275682 = 112067 / 275682 |
The gap fell by 0.14 points (0.140 unrounded), the second monthly fall in a row. People outside the labour force went from 38.21% to 38.09% of the population (−0.12) and the unemployed from 2.58% to 2.56% (−0.02). The unemployed are 6.3% of the gap. PEG fell while U-6 rose three tenths: employment grew by 463 thousand against a population gain of 128 thousand, and the fall again came mainly from the part outside the labour force. One month, first reading of a series BLS revises; nothing more is claimed.
Provenance, stated plainly. The October figures are the FRED series the script reads (book6/wp32-peg.py, run on 9 November 2026), not the BLS table: the identity checks in that run pass on all 393 months, and 100 − 59.3 (EMRATIO) = 40.7 agrees with 40.65 to the published 0.1. They were not checked against Table A-1 of the October release, and the pin remains August 2026. The September column is the BLS-table reading above; FRED gives the same 40.79%.
| Month | PEG | Outside labour force | Unemployed | U-3 | U-6 |
|---|---|---|---|---|---|
| Jan 1948 | 43.41% | 41.43% | 1.98% | 3.4% | — |
| Jan 1960 | 43.95% | 40.85% | 3.10% | 5.2% | — |
| Jan 1970 | 41.95% | 39.59% | 2.36% | 3.9% | — |
| Jan 1980 | 40.03% | 36.02% | 4.01% | 6.3% | — |
| Jan 1990 | 36.80% | 33.21% | 3.58% | 5.4% | — |
| Jan 2000 | 35.41% | 32.71% | 2.70% | 4.0% | 7.0% |
| Jan 2010 | 41.55% | 35.19% | 6.35% | 9.8% | 16.6% |
| Dec 2019 | 38.95% | 36.70% | 2.25% | 3.6% | 6.8% |
| Apr 2020 | 48.75% | 39.87% | 8.88% | 14.8% | 22.9% |
| Aug 2026 | 40.91% | 38.36% | 2.55% | 4.1% | 7.7% |
| Sep 2026 | 40.79% | 38.21% | 2.58% | 4.2% | 7.6% |
| Oct 2026 | 40.65% | 38.09% | 2.56% | 4.1% | 7.9% |
Over Jan 1948 to Aug 2026 the gap has been as low as 35.26% (Apr 2000) and as high as 48.75% (Apr 2020). From January 2000 to August 2026 it rose by +5.50 points: +5.65 from people outside the labour force and -0.15 from the unemployed. So the long-run movement in PEG is almost entirely participation, and the year-to-year spikes (April 2020: 8.88% unemployed, against 2.25% in December 2019) are almost entirely unemployment. U-6 exists on FRED only from 1994.
The 1948 and 1960 values are higher than today’s. That is a fact about who was counted in the labour force, not a claim that the labour market was worse; this page does not decompose it (section 6).
The ShadowStats “Alternate Unemployment Measure” is built on what its author calls proprietary modelling of long-term discouraged workers: no published formula and no citable inputs, so a third party cannot check it. That fails the standard this project holds its pages to, that a number carries a script that produces it. PEG is deliberately a different kind of object: two public series, one subtraction.
The Phillips curve is what ties an unemployment measure to inflation, and it is where a slack measure earns its keep. No page of the corpus says “Phillips curve” (0 files at the commit the script pins), and 1 of the 37 economics pages use the word “inflation” at all. So the link is open, and this section runs it once, in the plainest form, to see whether PEG could stand in for U-3.
Which curve. A. W. Phillips’s original paper (Economica, 1958) related the rate of change of money wage rates to unemployment in the United Kingdom, 1861–1957. Samuelson and Solow (1960) recast it as unemployment against price inflation and presented it as a menu for policy choice. Phelps (1967) and Friedman (1968) added expected inflation to the price relation, before the stagflation of the 1970s; that decade then broke the simple trade-off in the data (King 2008). The test below is the price form, on United States core CPI; it is not Phillips’s wage relation, and a wage version is not run. Phillips’s paper is held here and the account of it is checked against it: the title and scope above are its own, and its fitted curve, y + 0.900 = 9.638 x−1.394 on the 1861–1913 data (x the unemployment percentage, y the annual percentage change in wage rates), is a wage relation. The account of Samuelson and Solow, Phelps and Friedman is CITED from King (2008), Economic Quarterly 94(4), a secondary source that is held here; those three primary papers are not (WANTED).
The test. For each month, take the change in 12-month core CPI inflation over the next twelve months, and compare it with how far the measure sits above its own trailing ten-year mean (a stand-in for a natural rate). The specification was chosen before its results were seen, and no other specification was tried. Correlation r, U-3 against PEG:
| Window | Months | U-3 r | U-6 r | PEG r |
|---|---|---|---|---|
| 1967-01 to 2019-12 | 636 | -0.22 | None | -0.15 |
| 1985-01 to 2019-12 | 420 | -0.08 | None | +0.02 |
| 2004-01 to 2019-12 (all three) | 192 | +0.02 | +0.02 | +0.11 |
| 2020-01 to 2025-08 (all three) | 68 | +0.56 | +0.55 | +0.69 |
U-6 exists only from January 1994, and the test needs ten years of a measure before it can form a deviation, so U-6 enters from January 2004. The rows marked “all three” put U-3, U-6 and PEG on exactly the same months. PEG itself does not need U-6 and runs from 1948; restricting it to 1994 onward would only make the comparison like for like, and the windows above already do that where it matters.
What it shows. Before 2020 both measures have the expected negative sign (more slack, inflation rising less), and PEG is the weaker of the two. From 1985 to 2019 neither is distinguishable from zero, and on the like-for-like 2004–2019 window all three are: U-3 +0.02, U-6 +0.02, PEG +0.11. From 2020 on both signs flip to positive, because inflation rose while slack shrank. So in this setup PEG is not a better Phillips-curve proxy than U-3, and nothing here rescues either one after 2020. This is a descriptive check, not an estimate of the curve: a different natural-rate definition (for example CBO’s) or a different inflation series could move it, and neither was run.
Outside evidence. Granger and Jeon (2008, a preliminary working version, held here) refit Phillips’s curve on four economies with linear models whose parameters are allowed to drift. They report that unemployment improved inflation forecasts in most periods, with the strength declining in the recent ones. In their U.S. monthly data the model that let unemployment cause inflation beat the one that did not over 1971–1989 and 1971–2007, but not over 1990–2007, where a model using inflation alone was slightly better. That is a different method (forecast error, not a correlation) on a preliminary version, so it is context for the weak recent signs above, not a confirmation of them. They also recover Phillips’s original estimates, 9.638, −0.9 and −1.394, which match the curve quoted in section 5 (CITED; the paper is held).
Nonlinearity and regimes. Gagnon and Collins (2019, Peterson Institute working paper 19-6, held here) argue that the U.S. curve may be kinked when inflation is low, not permanently flatter. They model core CPI and core PCE inflation against the gap between the unemployment rate and the Congressional Budget Office’s estimate of the natural rate, and they split the sample where core CPI inflation crosses 3 percent: above it from 1967Q4 to 1995Q1, below it afterwards. If they are right, one linear correlation over a window that spans both regimes, as the 1985–2019 row above does, could blur a real slope in the high-inflation years and show almost none in the low-inflation ones. The test in this section is linear and does not split by regime, so it cannot tell their hypothesis from a curve that has become flat for good. A regime-split version is not run (CITED; the paper is held).
Supply shocks and who is counted. Gordon (1990, National Bureau of Economic Research working paper 3393, held here as a scan and read by optical character recognition) sets out the “triangle” model: inflation depends on inertia, demand and supply shocks such as oil prices, import prices and price controls. He notes that when supply shocks are imperfectly measured, the estimated effect of unemployment or output is biased toward zero, and that the bias is likely to be small only when a full set of shock variables is included. The test in this section has no supply-shock variable. He also recalls that the raw unemployment rate cannot be dropped into a Phillips curve without adjusting for demographics and for a natural rate that moves, and he prefers an output gap or capacity utilization. The trailing ten-year mean used here is a crude stand-in for a moving natural rate, and PEG’s population base carries the same demographic drift. None of this is run here (CITED; the paper is held).
The split by reason. BLS publishes counts of people not in the labour force by reason (retired, in school, disabled, family care, want a job now). Splitting the 38.36% block that way would isolate the labour-market-relevant slice from the rest. Not run.
Foregone output. An Okun’s-law conversion of that slice into dollars, using FRED actual and potential GDP. Not run, and not attempted before the split exists.
The wage form. Phillips’s own relation, wage growth against the level of unemployment, is not tested here; FRED has average hourly earnings series from 1964 that would allow it. Not run.
Whether PEG is the right proxy for anything. Section 5 tests it once against inflation and it does no better than U-3. WP-32 uses unemployment as an urgency proxy in its section 5; whether PEG, U-3 or U-6 is the better proxy there is untested. A natural-rate version of the Phillips test is not run, and neither is a split at 1995Q1 with a Congressional Budget Office gap, as in Gagnon and Collins (2019). A supply-shock control, as in Gordon (1990), is not run either.
Every page tagged ECONOMICS in the corpus (37 of them, at the commit the script pins), with how many times its own text says “unemployment” and how many times it uses any employment word (U-3, U-6, employment, labour force, jobless, labour market, payroll, jobs). The counts are computed, not judged. 35 of the 37 never say “unemployment”; the two that do are WP-32, where it is the urgency proxy, and WP-64, once. A zero is not a defect: most of these papers are about other things. It does mean that PEG connects to the corpus by measurement and by WP-32, and this page does not claim more.
Producing script: book6/wp128-verify.py — recomputes every figure above from FRED (retrieved live; the August 2026 row is pinned, so a later revision makes the script fail loudly instead of this table going stale quietly), rebuilds the table in section 6 from docs/subjects.tsv, and checks that this page prints what it computes. CITED FRED series CNP16OV, CE16OV, CLF16OV, UNEMPLOY, EMRATIO, UNRATE, U6RATE, fred.stlouisfed.org, not held. Nothing here is investment, legal or policy advice.