← All briefs

Verdict CONFIRMED

Hired, or not fired: the bigger shock healed faster, and only the flows say why

Unemployment peaked higher in 2020 than in 2009 — 14.8% against 10.0% — and was back to normal three and a half times faster. The headline poses that puzzle and cannot answer it. The flows can: 2009 was a collapse in job-finding, slow to repair, while 98% of 2020's rise was people losing jobs, and 78% of them were on temporary layoff. People on layoff get recalled.

Published 25 September 2026 · 16 series

Reachability swept first: yes — FRED searched with a control query that had to hit (53,486 for "unemployment rate") before any result was trusted; every series run through the real connector with its title read back, which caught one guessed ID serving a different quantity. Vintage checked: yes — all series via FRED, lastUpdated 2026-09-04 (the August 2026 release). Seasonally adjusted throughout.

Segment: Hired, or not fired

  • Source theme(s): Batch 54 #9 — the only theme in the buffer the bank rates hard.
  • Spine role: main story
  • The framing people carry: that the unemployment rate is one number moving one way, and a falling rate means the same thing whenever it happens.
  • Hypothesis (from the bank), three sub-claims: (1) recessions decompose into a spike in separations followed by a much longer depression in job-finding, "with the recovery governed almost entirely by the second"; (2) 2020 was an extreme outlier — an enormous separation spike with an unusually fast job-finding recovery; (3) a long-run downward drift in both flows, a labour market "less dynamic in both directions at once".
  • Verdict: (1) QUALIFIED, (2) CONFIRMED, (3) CONFIRMED — but only when measured like-for-like. The sequencing in (1) holds in every recession, and job-finding is the majority driver of pre-2020 unemployment rises at 58–66% — a majority, not "almost entirely". (2) is sharper than the bank put it: 98% of the 2020 rise came from separations, against 34–42% in every earlier recession. (3) holds across comparable late expansions for both flows, and a naive decade average would have wrongly half-refuted it.

The numbers (every row sourced — no exceptions)

All rows: BLS Current Population Survey via FRED, lastUpdated 2026-09-04, SA. Job-finding rate f = U→E ÷ previous month's unemployed; separation rate s = E→U ÷ previous month's employed. Peaks and troughs on 3-month averages.

Claim Value Series Confidence
Identity: unemployed stock = E→U + U→U + N→U + marginal inflows within 1 thousand in all 437 months; employed within 2 thousand UNEMPLOY, LNS17400000, LNS17500000, LNS17600000, LNS17700000; CE16OV high (see Open questions)
Separations peak BEFORE job-finding troughs 1990–91: 13 mo earlier · 2001: 20 · 2008–09: 9 · 2020: 7 derived high
Job-finding recovery vs separation recovery, months from recession start 1990–91: 80 vs 51 · 2001: not before next vs 58 · 2008–09: 130 vs 83 · 2020: 22 vs 36 — reversed derived high
Peak separation rate, × pre-recession base 1990–91 1.2× · 2001 1.3× · 2008–09 1.6× · 2020 6.4× (0.92% → 5.87%) derived high
Job-finding trough 1990–91 22.5% · 2001 23.2% · 2008–09 15.9% · 2020 23.3% derived high
Share of the rise in steady-state unemployment from JOB-FINDING 1990–91 66% · 2001 61% · 2008–09 58% · 2020 2% derived, u* = s/(s+f), log decomposition high
Steady-state unemployment, pre-recession → worst year 1990–91 4.9% → 6.8% · 2001 3.6% → 5.7% · 2008–09 4.2% → 9.5% · 2020 3.2% → 10.0% — a derived measure, NOT the published rate derived high
PUBLISHED unemployment rate, peak and recovery 2008–09: 10.0% (Oct 2009), from 4.7%; back within 0.5 pp in 69 months · 2020: 14.8% (Apr 2020), from 3.5%; back within 0.5 pp in 20 months UNRATE high
Share of the unemployed who are job losers ON LAYOFF Oct 2009: 11% (1.67M of 15.35M) · Feb 2020: 13% · Apr 2020: 78% (18.05M of 23.08M) · Dec 2020: 28% LNS13023653 ("Unemployment Level – Job Losers on Layoff") ÷ UNEMPLOY high
Late expansions, like-for-like (unemp / job-finding / separation) 1997–99 4.6% / 30.6% / 1.28% · 2005–07 4.8% / 27.5% / 1.23% · 2017–19 4.0% / 27.4% / 0.99% · 2023–25 4.0% / 26.6% / 0.97% derived high
Decade averages (job-finding / separation) — the misleading view 1990s 27.4% / 1.47% · 2000s 26.0% / 1.37% · 2010s 22.4% / 1.25% · 2022–26 27.0% / 0.97% derived high as arithmetic, low as evidence

The texture (what narrative flattens)

  • The bigger shock healed faster — and only the flows say why. The published unemployment rate peaked at 14.8% in 2020 against 10.0% in 2009, yet 2020 was back within half a point of its starting level in 20 months against 69. The headline shows the puzzle and cannot explain it. The flows do: 58% of the 2008–09 rise was people not finding work, a slow thing to repair; 98% of 2020's was people losing it — and of a particular kind. At the April 2020 peak, 78% of the unemployed were job losers on layoff, against 11% at the 2009 peak. People on layoff are recalled; that is why the separation spike reversed as fast as it arrived. (A draft stated this cause before checking it; the layoff series was then fetched and confirms it.)
  • The sequence is the same every time before 2020. Separations spike first; job-finding bottoms out nine to twenty months later (seven in 2020) and takes far longer to recover — 130 months after 2008–09 against 83 for separations.
  • 2020 inverted it. Separations went to six times normal in weeks, and job-finding recovered in 22 months — faster than separations did, for the first time in the record.
  • The labour market has quietened. In 2023–25 unemployment averaged 4.0%, below 1997–99's 4.6% — yet fewer people found work each month and fewer lost it. Same tightness, less churn.

Datoid artifact (Tier-A only)

  • UNRATE is already seeded. Every flow and stock here is ingestable through the existing FRED connector, verified by running fetchFredSeries: the nine flows and two marginal-inflow series 1990-02 onward, the two stocks 1990-01 onward, all in thousands of persons — one unit, so no divisor trap.
  • The rates are ratios of a flow to a LAGGED stock — derived, not ingestable. The platform can draw a derived line only as a sum (derived-series.ts); a job-finding rate is a quotient with a one-month offset, and the steady-state decomposition is further derivation on top. Same build gap as the CPI sheet: this needs either computed series or chart support the platform does not yet have.
  • A SEMANTICS TRAP, caught by reading the title back. LNS17300000 was guessed to be "Not in Labor Force to Not in Labor Force" by position in the ID sequence. It is "Marginal Inflows to Employed." Guessed series IDs must have their titles checked before use.

Story justification (decide here — do NOT defer)

  • Does the argument have a reveal that needs sequencing? YES — BUILD.
  • Because: the reveal is a puzzle the headline poses and cannot answer. Show the published unemployment rate: 2020 spikes far higher than 2009 (14.8% against 10.0%) and recovers three and a half times faster. Let the reader sit with why — then open the flows: 2009 a job-finding collapse, 2020 a separations spike. Shown decomposed from the start, there is no question to answer.
  • A draft of this gate said the two "peaked near 10%" and "look the same". That was the DERIVED steady-state rate (9.5% and 10.0%), not the published one a reader would actually see, and on the real chart the reveal would not have happened. The corrected contrast is sharper.

The questions this raises (and whether we can answer them)

New to Stage 1, added 2026-09-23. Everything above tests a hypothesis. This section names what a reader would ask next and classifies each answer, so the choice is made BEFORE the episode is written: fetch it, attribute it, or declare it open. The point is to stop an episode alluding to context it could simply have gone and got. Fetchable rows are theme-bank candidates in their own right.

  1. Why did the 2020 layoffs reverse so fast? — ATTRIBUTABLE. The sheet establishes the fingerprint (78% of the unemployed were job losers on layoff in April 2020 against 11% in October 2009) but not the cause. Whether recall was driven by closures being time-limited, or by the Paycheck Protection Program paying firms to keep people on payroll, needs a named source. Not a series, and must not be asserted from memory.
  2. Is the quieter labour market stability, or stagnation? — FETCHABLE, and the strongest of them. JTSHIR (hires rate) and JTSLDR (layoffs and discharges rate) give churn from the employer side beside the already-seeded JTSQUR (quits); FRBATLWGT12MMUMHWGJSW and FRBATLWGT12MMUMHWGJST (Atlanta Fed, 12-month median wage growth for job switchers and stayers) price what not moving costs. If switchers hold a persistent premium, low churn is not free — which turns the sheet's flattest finding into a cost somebody bears. Fetched.
  3. In 2020, were people recalled — or did they leave? — FETCHABLE. LNS17900000 (unemployed → not in labour force) against the already-seeded CIVPART. A fast fall in the unemployment rate is a different event if the exit door did the work, and the recovery comparison at the heart of this sheet depends on which it was. Fetched.
  4. Why has churn fallen for three decades? — OPEN. Candidates include an ageing workforce, longer tenure, industry mix, non-compete agreements and better online matching. No single series settles it and separating ageing from the rest needs age-specific flows. Contested in the literature: say that, rather than picking one.
  5. Does the two-state simplification move the shares? — OPEN (method). Already under Open questions; a three-state decomposition assigns different shares, so the 58–66% and 2% figures must be named as two-state if quoted.

Open questions / where I'm guessing

  • "ALMOST ENTIRELY" OVERSTATES IT. Job-finding drove 58–66% of each pre-2020 rise — a clear majority, but separations contributed a third or more. Say "mostly", not "almost entirely". The bank's claim is also about the recovery; this sheet decomposes the rise, and supports the recovery claim only through the recovery-duration comparison (job-finding slower in every pre-2020 case). A recovery-phase decomposition was not run.
  • DECADE AVERAGES WOULD HAVE REFUTED A TRUE CLAIM — the fifth window lesson, running the other way. By decade, job-finding "rebounded" to 27.0% in 2022–26. That compares a tight post-COVID market with a 2010s decade dominated by the post-2008 slump. Compared at equivalent points of the cycle, both flows fell. The previous four sheets were caught letting a clean number through; this one nearly let a misleading comparison reject one. The rule is not "distrust clean numbers" but "compare like with like".
  • A TWO-STATE SIMPLIFICATION. The steady-state decomposition uses only employment ↔ unemployment flows. Moves to and from not-in-the-labour-force — U→N and N→U are large — are ignored. A three-state decomposition assigns different shares; the 58–66% and 2% figures are two-state and should be named as such if quoted.
  • THE IDENTITY IS NEAR-DEFINITIONAL. BLS constructs the marginal inflows so that flows reconcile with the published stocks, so a close within rounding confirms the series IDs and the method, not that the survey measures the labour market well. The same caveat as the FAO and CPI sheets.
  • ONE MISSING SURVEY MONTH COSTS TWO MONTHS OF FLOWS. October 2025 is absent from both stocks; October and November 2025 are absent from every flow, because a flow joins two consecutive months and each of those two needs October. Likeliest cause the 2025 federal shutdown — not verified here. It moves no finding (the 2023–25 average runs over 34 months rather than 36) but any chart of these rates carries a two-month gap.
  • A BUG IN MY OWN ANALYSIS, recorded for method. The first pass gave 2001 a job-finding trough of 15.9% in December 2009 — the Great Recession's number — because each recession's search window ran eight years and 2001's overran into
    1. Windows are now capped at the month before the next recession.
  • Four recessions only. CPS flows begin in 1990, so every generalisation about "recessions" rests on n = 4, one of which (2020) is the outlier the sheet describes. Treat the pre-2020 pattern as three cases.
  • Seasonal factors and population controls are revised, typically each January — so historical values move between vintages. Observed once here: the April 2020 unemployment rate, widely quoted as 14.7%, reads 14.8% in the current FRED vintage. The general practice is not verified this session; this instance is.

Briefs are research artifacts, not articles — they are the check that happens before anything else gets published. If a number here is wrong, it is wrong at the source or in our pipeline, and both are fixable: [email protected]. How this is made.