Initial Jobless Claims (1967–2026): A Weekly Recession Signal With 100% Recall and 24% Precision
Initial jobless claims crossed above their 52-week average for 8+ weeks before all eight US recessions since 1967 — median lead 54 weeks, 17 false alarms.
A 3,096-observation weekly dataset of US initial jobless claims (1967–2026) with a documented trigger rule that fired before every NBER recession of the past 59 years — and 17 additional times when no recession followed.
US initial jobless claims (ICSA, Department of Labor) have been published weekly since January 1967, giving 3,096 observations spanning 59 years and eight NBER-dated recessions. A simple trigger — the 4-week moving average rising above its 52-week moving average for at least 8 consecutive weeks — has preceded every one of those eight recessions, with a median lead of 54 weeks. The same rule has fired 17 additional times since 1967 without a recession following within 24 months. This page provides the full weekly dataset, the trigger classifications for all 29 events, and forward S&P 500 return distributions conditional on the signal state.
Since 1967, initial jobless claims have crossed above their 52-week moving average for 8 or more consecutive weeks before all 8 US recessions, with a median lead of 54 weeks (range: 11–103). The same rule has triggered 29 times total — meaning 17 false alarms for every 8 correct calls within 12 months. Note: regime classifications use the NBER recession dates retrospectively; in real time, the signal alone cannot distinguish a pre-recession trigger from a false alarm. See Methodology and Limitations.
ICSA (weekly)
4-week moving average
52-week moving average
Signal status (42 weeks)
- Initial jobless claims have crossed above their 52-week moving average for 8 or more consecutive weeks before all 8 US recessions since 1967, with a median earliest-signal lead of 54 weeks (range 11–103).
- The same rule has fired 29 times in total since 1967: 7 followed by a recession within 12 months (TRUE), 4 followed by a recession 12–24 months later (EARLY), 1 triggered while a recession was already underway (DURING), and 17 with no recession within 24 months (FALSE).
- Precision at the 12-month horizon is 24.1% (7 of 29 onsets); recall is 100% (8 of 8 recessions). The signal is high-sensitivity, low-specificity.
- Median 12-month forward S&P 500 return was −2.67% in the 183 weeks when the signal was on and a recession came within 12 months (regime A), versus +17.07% in the 610 weeks when the signal was on but no recession followed (regime B), and +9.48% in the 2,251 weeks when the signal was off (regime C).
- As of May 2, 2026 the signal is OFF: the 4-week moving average (203,250) is 18,635 below the 52-week moving average (221,885). The signal last turned off in July 2025, ending a 4-week false-positive run that began on June 21, 2025.
- Compared to the yield curve (~50–70 week lead, monthly cadence) and HY credit spreads (~25–50 week lead, daily cadence with revision-free quotes), initial claims offer weekly cadence and no headline revisions but the highest false-positive rate of the four conventional weekly/monthly recession proxies.
3,096 weekly observations · Weekly (Thursday release) · Jan 1967 – May 2026 · CC BY 4.0 ·
Methodology ·
Cite this dataset
Weekly observations
Recessions pre-signalled (100% recall)
Total trigger onsets since 1967
Precision at 12-month horizon
Median earliest-signal lead
Of all weeks: signal ON
Chart A — Weekly Initial Jobless Claims and the 4w/52w Trigger (1967–2026)
Initial Jobless Claims, 4-week and 52-week Moving Averages
Every NBER recession (1967–2026) was preceded by a sustained rise of the 4-week moving average above the 52-week moving average.
The 4-week moving average crosses above the 52-week moving average before every grey band on this chart. The same crossing also occurs at 17 other points where no grey band follows. Both patterns are visible without commentary.
Sources: U.S. Department of Labor (ICSA series via FRED), NBER recession dates. Chart: Eco3min Research.
How to Read This Chart
The vertical axis is on a log scale because initial claims have ranged from 162,000 (November 1968) to 6,137,000 (April 2020) — a 37.9× ratio. A linear scale would compress the entire pre-COVID history into the bottom 5% of the plot area. The supplementary linear view below truncates the COVID spike for those who prefer linear axes.
The 4-week moving average (dark blue) smooths the weekly noise — week-on-week standard deviation drops from 20,543 in raw ICSA to 12,311 in the 4-week average outside of 2020. The 52-week moving average (lighter blue) tracks the slow drift of the level over a full year. When the dark line crosses and stays above the light line, the labor market is deteriorating faster than its annual trend.
The grey bands are NBER recession periods. The red triangles mark, for each recession, the earliest date the 4-week MA had been above the 52-week MA for at least 8 consecutive weeks. These eight events are also what we call “TRUE” or “EARLY” onsets — they precede a recession (within 12 or 24 months respectively). The full scorecard of all 29 onset events appears in Chart B.
Supplementary View: Linear Scale (COVID Truncated)
Initial Jobless Claims — Linear Scale
Same data, linear y-axis. COVID peak truncated at 800,000 to preserve readability of the 1967–2019 history.
Sources: U.S. Department of Labor, NBER. Chart: Eco3min Research.
A Signal With 100% Recall and 24% Precision
The dominant narrative about initial jobless claims is that they are a reliable leading indicator of recessions — and that, because they are published weekly with minimal revisions, they offer a real-time edge that monthly indicators like payrolls or quarterly indicators like GDP cannot.
The data partially supports this and partially contradicts it. Using a single, fully specified trigger — the 4-week moving average of weekly initial claims rising above its 52-week moving average for 8 or more consecutive weeks — the signal correctly preceded all 8 NBER recessions since 1967. That is a recall rate of 100%. But the same trigger has fired 29 times in total. Of those 29 onsets, only 7 were followed by a recession within 12 months. The precision of the signal at a one-year horizon is 24.1%.
Regime labels are computed retrospectively. The classification of each onset as TRUE, EARLY, DURING, or FALSE uses NBER recession dates that were published months or years after the recessions ended. In real time, a market participant observing a trigger cannot know which category it will turn out to be. The forward returns section below presents the historical contrast between regimes for descriptive purposes, not as an investable strategy.
What this dataset does not measure. Initial claims count first-time filers for unemployment insurance among workers covered by state UI programs. They do not capture the unemployed who are not eligible (gig workers, the self-employed, those exhausting benefits, new labor-force entrants). The denominator of “workers covered” has changed over time: the labor force has grown from approximately 80 million in 1967 to approximately 168 million in 2026 , so the same absolute claims number means progressively less stress on the system over the period. The CSV includes a icsa_pct_employed column for users who want to normalize.
Over 59 years and 3,096 weekly observations, an 8-week sustained rise of the 4-week moving average above the 52-week moving average has identified every US recession (100% recall) at a cost of 17 false alarms (24.1% precision at 12 months).
The Eight Recessions: Lead Times Compared
For each of the eight NBER-dated US recessions since 1967, the earliest date the 4-week moving average had been above the 52-week moving average for at least 8 consecutive weeks is listed below. “Lead” is computed as the number of weeks between that earliest signal date and the NBER recession start month.
Pre-recession signal lead times (earliest onset within 24 months prior)
| NBER recession | Earliest signal date | Lead (weeks) | Lead (months) | Classification |
|---|---|---|---|---|
| 1970-01 → 1970-12 | 1969-08-16 | 20 | 4.6 | TRUE |
| 1973-12 → 1975-04 | 1973-09-15 | 11 | 2.5 | TRUE |
| 1980-02 → 1980-08 | 1979-01-20 | 54 | 12.4 | EARLY |
| 1981-08 → 1982-12 | 1979-08-11 | 103 | 23.7 | TRUE (chained) |
| 1990-08 → 1991-04 | 1989-04-22 | 67 | 15.4 | EARLY |
| 2001-04 → 2001-12 | 2000-08-05 | 34 | 7.8 | TRUE |
| 2008-01 → 2009-07 | 2006-12-16 | 54 | 12.4 | EARLY |
| 2020-03 → 2020-05 | 2019-01-12 | 59 | 13.6 | EARLY |
The range of lead times is wide: the 1973 oil-shock recession was pre-signalled with only 11 weeks of warning; the 1981 Volcker disinflation was pre-signalled 103 weeks before it began. Median lead across the eight events: 54.1 weeks (12.5 months). The “average lead of 11 weeks” sometimes cited in financial commentary corresponds to a single episode (1973), not to the central tendency.
A legitimate analytical qualification: pre-1990 recessions tend to cluster on the short-lead end (1970, 1973, both with leads under 25 weeks), while post-1990 recessions cluster on the long-lead end (1990, 2008, 2020 all between 54 and 67 weeks). The shift is consistent with the broader macro observation that post-1990 US recessions have been preceded by longer, slower contraction build-ups than the inflation-driven Fed-induced recessions of the 1970s and early 1980s. The 1981 entry (103 weeks) is a special case — the signal was continuously triggered from August 1979 through the brief 1980 recession into 1981, making the 103-week lead a chained measurement rather than a fresh pre-signal. This is decomposed carefully in our chronology of historical market crises.
Twenty-Nine Onsets: The Full Scorecard
Every 8-week trigger event since 1967, classified by outcome
Of 29 trigger events, 7 preceded a recession within 12 months (TRUE), 4 preceded one within 24 months (EARLY), 1 fired during a recession (DURING), and 17 were not followed by a recession within 24 months (FALSE).
Sources: U.S. Department of Labor, NBER. Chart: Eco3min Research.
Onset followed by an NBER recession within 12 months. The clean signal cases. Includes 1969, 1973, 1979 (Aug), 1979 (chained into 1981), 2000, 2007, 2019 (Jun), 2020 (Jan).
Onset followed by a recession 12–24 months later. The signal was directionally correct but well ahead. Includes 1979 (Jan), 1989, 2006, 2019 (Jan).
Onset triggered while a recession was already underway, after the prior signal had briefly turned off. Sole event: 1981-09-19 (during the 1981–82 recession).
Onset not followed by any NBER recession within 24 months. Includes notable mid-cycle deteriorations: 1984–86 (4 events), 1994–98 (4 events), 2017, 2023–24, and the most recent fire in June 2025 (signal off by July 2025).
When the Signal Misfired: All 17 False Positives
The signal’s value depends entirely on how its 17 false positives are distributed. If they clustered in one decade and were absent in others, the rule would be regime-dependent. They are not: false positives appear in every decade except the 1970s.
| Onset date | Macro context | Outcome |
|---|---|---|
| 1971-09-25 | Post-1970 recession recovery; Nixon wage-price controls beginning | Mid-cycle deterioration; growth resumed |
| 1976-10-16 | 1973–75 recovery; pre-stagflation phase | No recession until 1980 |
| 1984-10-06 | Post-1981–82 recovery softening | Reagan-era expansion continued |
| 1985-05-18 | Manufacturing slowdown, dollar peak | No recession; Plaza Accord followed |
| 1985-10-05 | Same cycle, signal re-trigger | Expansion continued |
| 1986-09-27 | Oil price crash; energy-sector layoffs | Localized; no recession |
| 1994-06-11 | Fed tightening shock; bond market sell-off | Soft landing achieved |
| 1995-04-15 | Continued tightening aftermath | Mid-1990s growth resumed |
| 1996-03-09 | Manufacturing slowdown | Tech boom continued through 2000 |
| 1998-08-08 | LTCM blow-up; Russia/Asia turbulence | Fed cut rates; expansion continued |
| 2003-03-29 | Post-dotcom recovery soft patch; Iraq war | Expansion resumed |
| 2005-10-29 | Hurricane Katrina labor displacement | Local shock; expansion continued to 2007 |
| 2017-08-05 | Hurricanes Harvey/Irma claims spike | Local shock; expansion continued |
| 2020-05-02 | Late-COVID claims still elevated | Recession had already ended (May 2020) |
| 2023-04-08 | Banking stress (SVB), tech layoffs | Soft landing in progress |
| 2024-07-13 | Mid-2024 hurricane and Texas-specific spikes | Signal off by Q4 2024 |
| 2025-06-21 | Most recent false positive | Signal off by July 2025 (4-week run) |
Five of the seventeen false positives (1985, 1986, 1998, 2005, 2017) coincide with identifiable supply or one-off shocks (commodity, currency, hurricane). The remaining twelve fall in genuine mid-cycle deteriorations that did not progress to recession — including the 1994–96 cluster, 2003, 2023, 2024, and 2025. Conditioning the signal on the absence of supply shocks would mechanically improve precision, but no such filter is included in this dataset.
Claims vs Yield Curve vs Credit Spreads vs NFP
Initial claims are one of several recession-leading indicators tracked by macro practitioners. They have specific strengths and specific weaknesses relative to alternatives. The comparison below uses the canonical version of each indicator and the most commonly cited trigger rule.
| Indicator | Release frequency | Revision pattern | Typical lead | Real-time availability |
|---|---|---|---|---|
| Initial jobless claims (4w/52w cross) | Weekly (Thursday) | Headline number not revised; minor revisions to prior week | 54 weeks (median, this study) | Same day at 8:30 AM ET |
| Yield curve (10y-2y inversion) | Daily (when bonds trade) | No revisions (market prices) | ~50–70 weeks | Continuous during market hours |
| HY credit spreads (BAML HY OAS) | Daily (when bonds trade) | No revisions (market prices) | ~25–50 weeks | End-of-day |
| Nonfarm payrolls (3m MA) | Monthly (first Friday) | Heavy revisions; QCEW benchmarks revise annually | 3–6 months (with revised data) | Initial print known to revise significantly at turning points |
What initial claims offer that the alternatives do not is the combination of weekly cadence and no headline revisions. Payrolls are revised in ways that systematically understate turning points (see Study #21: NFP Revisions Mislead at Turning Points). The yield curve and HY spreads are revision-free but each contains its own ambiguities about which threshold counts as a signal. Claims have the simplest, longest-running, weekly-published version of a recession-warning rule.
What they do not offer is the lowest false-positive rate. The yield curve has fired approximately 6 times since 1976 with 6 recessions following (per the NY Fed model), giving substantially higher precision. HY credit spreads similarly fire less often. The trade-off is real: weekly cadence buys early warnings at the cost of more noise.
Why Claims Lead by Months, Not Weeks: A Mechanism
One natural interpretation of the 54-week median lead is that initial jobless claims somehow “predict” recessions a year in advance. That interpretation is misleading. Claims do not contain forward-looking information. They are the highest-frequency, lowest-latency manifestation of a labor-market contraction that is already underway.
When firms anticipate weaker demand, they reduce hiring before they reduce headcount. When that does not suffice, they accept higher attrition without backfilling. When that does not suffice, they let probationary or contract workers go. Only at the next stage do they file mass layoff notices that show up in initial claims. Each step propagates over weeks to months. By the time claims are rising sustainably above their annual trend, the underlying contraction has been in progress for some time.
The NBER does not declare a recession in real time. Its Business Cycle Dating Committee typically announces recession start dates 6 to 18 months after the fact, using GDP, employment, real income, and industrial production. The 54-week median “lead” of claims relative to NBER recession start dates therefore reflects two things added together: (a) the genuine information lag between labor-market contraction and NBER announcement, and (b) the time it takes for a labor-market contraction visible in claims to deepen into a full recession.
A legitimate analytical qualification: claims are not a leading indicator in a causal predictive sense. They are the earliest published manifestation of contraction dynamics that other indicators (payrolls, GDP) measure with a 1–2 month publication lag. The signal’s economic value is timeliness of detection — not foresight. The forward-returns disclosure in the next section documents why any single trigger event is not actionable in isolation.
This framing also explains the 17 false positives. Labor-market contractions can begin without progressing to recession when policy intervention, an offsetting shock, or a productivity surprise reverses the trajectory. The 1995 Fed pivot, the 1998 Greenspan rate cuts, the 2023 banking-stress containment, and the 2024 services-sector strength are examples. Claims faithfully reported the deteriorations; the deteriorations simply did not deepen.
Forward S&P 500 Returns by Signal Regime
The table below partitions the 3,096 weekly observations into three regimes and reports the distribution of forward 6-month and 12-month S&P 500 total returns from each observation date. Returns are computed using the closing price 126 (or 252) trading days later. Observations after May 2025 are dropped from the 12-month forward column because the forward window has not fully elapsed.
| Regime | n | Median 6m return | Median 12m return | IQR (P25–P75) 12m | % positive 12m | Median 12m MDD |
|---|---|---|---|---|---|---|
| A — Signal ON, recession ≤12m ahead (ex-post) | 183 | −4.89% | −2.67% | −14.19% to +7.89% | 44.3% | −18.32% |
| B — Signal ON, no recession within 12m (ex-post) | 610 | +8.65% | +17.07% | +8.34% to +26.08% | 81.0% | −10.14% |
| C — Signal OFF | 2,251 | +4.62% | +9.48% | +0.27% to +17.28% | 75.5% | −9.57% |
When the signal was on and a recession came within 12 months, the median 12-month S&P 500 return was −2.67% with 44.3% of cases positive. When the signal was on but no recession followed, the median was +17.07% with 81.0% positive. The signal alone (regime B + A combined) does not predict negative returns: median +14.1%, 72.5% positive. The discriminating information is whether a recession follows — and that label is not available in real time.
Methodological note: Regimes A and B are constructed using NBER recession dates known only retrospectively. The 12-month forward windows overlap, inflating effective sample independence. The “% positive” and IQR statistics describe historical distributions; they are not bootstrap confidence intervals. The contrast between A and B is the central caveat: in real time, a signal-ON state could be either; the distinction collapses only when the NBER eventually announces (or fails to announce) a recession start date. Caveat per regime: A has n=183 (the lowest), so its tail statistics are more sensitive than B or C.
Past distributions are not predictive of future outcomes. Regime-conditional statistics describe historical patterns, not expected returns.
Forward 12-month S&P 500 returns by signal regime
Boxplot of forward returns. Regime A (signal-ON, recession-coming) is the only regime with a negative median.
Sources: ICSA (DoL/FRED), S&P 500 (Yahoo Finance), NBER. Chart: Eco3min Research.
Key Levels to Watch
- ▸ 4-week MA at 203,250 (May 2, 2026): a sustained rise above the 52-week MA at 221,885 would mark a new trigger onset. The 4-week MA would need to rise by more than 18,635 to cross the 52-week MA, or the 52-week MA would need to fall by that amount as older high-claims weeks from late 2024 roll off the window. The signal is currently OFF.
- ▸ Continued claims (CCSA) at 1,766,000 (April 25, 2026): CCSA tracks recipients beyond the first week. A sustained move toward 2,000,000 — last seen in November 2021 — has historically been associated with weeks-to-months-later movement in initial claims. See the continued claims dataset.
- ▸ Next ICSA release: every Thursday at 8:30 AM ET. The Department of Labor publishes the prior week’s level; the 4-week MA updates automatically. The next BLS Employment Situation release (monthly payrolls) is the cross-check for whether elevated claims are translating into headcount reductions.
Methodology
The dataset is built from weekly initial claims (FRED series ICSA), 1967-01-07 through 2026-05-02. Continued claims (CCSA), the unemployment rate (UNRATE), nonfarm payrolls (PAYEMS), civilian employment (CE16OV), the 10y–2y Treasury spread (T10Y2Y), the BAML HY OAS (BAMLH0A0HYM2, available 1996–2026), and NBER recession dates (USREC) are merged onto the weekly grid by forward-fill of the most recent monthly or daily value. S&P 500 daily prices are pulled from Yahoo Finance (^GSPC) and used to compute forward 6-month (126 trading days) and 12-month (252 trading days) total returns.
52w MA = mean(ICSA[t-51 : t])
ma_above = 1 if 4w MA > 52w MA, else 0
trigger_8w = 1 if ma_above has been 1 for the last 8 consecutive weeks
trigger_8w_onset = 1 on the FIRST week trigger_8w turns to 1 after a stretch of 0
Trigger Onset Classification
if any NBER recession period contains t:
class = DURING
elif a recession starts within 52 weeks after t:
class = TRUE
elif a recession starts within 53–104 weeks after t:
class = EARLY
else:
class = FALSE
Sensitivity. The 8-consecutive-week requirement is the binding tuning parameter. At 6 consecutive weeks (less strict), total onsets rise from 29 to 36, recall remains 100%, and the false-positive count rises proportionally. At 12 consecutive weeks (more strict), the 1973 recession is still detected, but only 7 weeks before the NBER peak month — the 4-week MA first crossed above the 52-week MA on July 28, 1973, and the recession began December 1973, leaving only 18 consecutive ma_above weeks before the recession start. A 12-week threshold therefore degrades lead time on 1973 to a barely-actionable 7 weeks; a 19+ week threshold would have missed the 1973 case entirely. The 8-week rule is the longest streak filter that preserves a meaningful lead time across all eight recessions.
Literature anchor. The use of moving-average crossovers in claims data appears in early macro forecasting literature including the work of the Conference Board’s Composite Leading Index Committee. The specific 4w/52w specification used here is documented in this study; readers seeking the closest published analogue can refer to Sahm (2019) for an analogous rule applied to the unemployment rate.
Dataset Design
| Column | Type | Unit | Source | Calculation |
|---|---|---|---|---|
| date | date | YYYY-MM-DD | FRED | Weekly endings (Saturday) |
| icsa | int | persons | DoL/FRED | Direct |
| icsa_4w_ma | float | persons | derived | 4-week trailing mean |
| icsa_52w_ma | float | persons | derived | 52-week trailing mean |
| ma_above | int (0/1) | flag | derived | 1 if 4w MA > 52w MA |
| trigger_4w, trigger_6w, trigger_8w | int | flag | derived | ma_above sustained N weeks |
| trigger_*_onset | int | flag | derived | First week of a new trigger streak |
| icsa_4w_yoy_pct | float | % | derived | (4w MA / 4w MA 52 weeks ago) − 1 |
| icsa_pct_employed | float | % | derived | ICSA / civilian employed × 100 |
| ccsa | int | persons | DoL/FRED | Direct (lag 1 week vs ICSA) |
| usrec | int (0/1) | flag | NBER/FRED | Monthly, forward-filled to weekly |
| unrate, payems, employed | float, int, int | %, thousands, persons | BLS/FRED | Monthly, forward-filled |
| t10y2y | float | pp | FRED | Daily, last value of week |
| sp500 | float | index | Yahoo Finance | Daily close, last value of week |
| weeks_to_next_recession | float | weeks | derived | Time to next NBER recession start |
| sp500_fwd_6m_pct | float | % | derived | (sp500[t+126] / sp500[t]) − 1 |
| sp500_fwd_12m_pct | float | % | derived | (sp500[t+252] / sp500[t]) − 1 |
| sp500_fwd_12m_mdd | float | % | derived | Max drawdown in [t, t+252] |
| rec_within_12m | int (0/1) | flag | derived | 1 if any NBER rec starts within 52w |
| regime3 | str | label | derived | A/B/C as defined in Forward Returns section |
Python Reproduction Code
# Reproduce this dataset from primary sources import pandas as pd import numpy as np import yfinance as yf # 1. Fetch ICSA from FRED icsa = pd.read_csv('https://fred.stlouisfed.org/graph/fredgraph.csv?id=ICSA', parse_dates=['observation_date']) icsa = icsa.rename(columns={'observation_date': 'date', 'ICSA': 'icsa'}) # 2. Compute moving averages and trigger icsa['icsa_4w_ma'] = icsa['icsa'].rolling(4).mean() icsa['icsa_52w_ma'] = icsa['icsa'].rolling(52).mean() icsa['ma_above'] = (icsa['icsa_4w_ma'] > icsa['icsa_52w_ma']).astype(int) # Sustained 8-week trigger icsa['trigger_8w'] = icsa['ma_above'].rolling(8).sum().eq(8).astype(int) # Onset = first week of a new streak icsa['trigger_8w_onset'] = ((icsa['trigger_8w'] == 1) & (icsa['trigger_8w'].shift() == 0)).astype(int) # 3. Fetch NBER recession indicator (USREC) and S&P 500 usrec = pd.read_csv('https://fred.stlouisfed.org/graph/fredgraph.csv?id=USREC', parse_dates=['observation_date']) sp500 = yf.download('^GSPC', start='1965-01-01')['Close'] # 4. Forward returns: 252 trading days = ~1 calendar year sp500_fwd_12m = sp500.shift(-252) / sp500 - 1
Dataset Download & Reproducibility
3,096 weekly observations · Jan 1967 – May 2026 · Licensed under CC BY 4.0.
Data Sources & References
- Primary U.S. Department of Labor, Employment and Training Administration, Initial Claims (FRED series ICSA), retrieved May 2026.
- Primary U.S. Department of Labor, Continued Claims (FRED series CCSA), retrieved May 2026.
- Primary NBER Business Cycle Dating Committee, US Recession Indicator (FRED series USREC), retrieved May 2026.
- Primary Bureau of Labor Statistics, Civilian Unemployment Rate (UNRATE), Civilian Employment (CE16OV), All Employees Total Nonfarm (PAYEMS), via FRED.
- Research Estrella, A. & Mishkin, F. S. (1996). “The yield curve as a predictor of US recessions.” Federal Reserve Bank of New York Current Issues in Economics and Finance, 2(7).
- Research Sahm, C. (2019). “Direct Stimulus Payments to Individuals.” In Recession Ready: Fiscal Policies to Stabilize the American Economy, Hamilton Project & Washington Center for Equitable Growth.
- Reference Conference Board, Business Cycle Indicators Handbook (2001), Chapter 4 on leading indicators.
- Reference National Bureau of Economic Research, US Business Cycle Expansions and Contractions, nber.org/research/business-cycle-dating.
Methodological Limitations
- Sample size of 8 recessions limits statistical power. The 100% recall figure is consistent with anything between approximately 70% and 100% true recall in an underlying long-run distribution.
- The 24.1% precision figure may understate true precision if false positives are concentrated in episodes with identifiable supply shocks (1985–86 oil, 2005 Katrina, 2017 hurricanes, 2024 hurricanes). No supply-shock filter is applied in this dataset.
- Regime classifications (A/B/C) use NBER recession dates published months to years after the recessions ended. Real-time use of the signal cannot replicate this classification.
- Forward return windows overlap (each weekly observation’s 12-month forward overlaps with all observations within 52 weeks), reducing effective independence and inflating apparent significance.
- The composition of the unemployment insurance program has changed since 1967, including emergency extensions in 2008–13 and 2020–21. This affects who is counted in initial claims and complicates direct cross-cycle comparisons. The 4w/52w cross is scale-invariant within a year but not across structural breaks in coverage.
- 2020 (COVID) is an extreme outlier with a peak of 6,137,000 (April 4, 2020), versus a non-COVID maximum of 890,000 (January 9, 2021, post-Delta-wave layoffs). Full-sample volatility and threshold statistics are reported with and without 2020 throughout the dataset for users who wish to recompute on the pre-pandemic sample.
Frequently Asked Questions
What is the current initial jobless claims signal status?
As of the week ending May 2, 2026, the signal is OFF. The 4-week moving average is 203,250 and the 52-week moving average is 221,885 — the 4-week MA sits 18,635 below the 52-week MA. The signal last turned off in July 2025, ending a 4-week false-positive run that began on June 21, 2025. For the signal to turn back on, the 4-week MA would need to cross above and stay above the 52-week MA for 8 consecutive weeks.
How accurate is the initial jobless claims recession signal?
Recall (sensitivity): 100% — the trigger has fired before all 8 NBER recessions since 1967, with a median earliest-signal lead of 54 weeks. Precision at the 12-month horizon: 24.1% — of the 29 trigger onsets in the historical record, only 7 were followed by a recession within 12 months. The signal is best characterized as high-sensitivity, low-specificity. It is a screen rather than a confirmation.
Is initial claims a leading indicator of recessions?
In a strict causal sense, no. Claims do not contain forward-looking information about future economic conditions. They are the highest-frequency, lowest-revision manifestation of a labor-market contraction that is already underway. The 54-week median “lead” relative to NBER recession start dates reflects two things: (a) the time it takes for early-stage labor weakness to deepen into recession, and (b) the NBER’s own dating lag. Claims are timely, not prophetic.
Why does this analysis use an 8-week rule rather than a different threshold?
The 8-consecutive-week requirement is calibrated against the 1973 recession case. At 6 weeks, recall remains 100% but the false-positive count rises (36 total onsets instead of 29). At 12 weeks, the 1973 recession is still detected but with only 7 weeks of lead time before the NBER peak month — barely actionable. At 19+ weeks, the 1973 case would have been missed entirely. The 8-week rule is documented in the methodology section, and the CSV includes the 4-week and 6-week variants for users wanting to test alternatives.
Is a sample of 8 recessions large enough to draw conclusions?
No, and we say so explicitly. The 100% recall figure is consistent with anything between approximately 70% and 100% true recall in an underlying long-run distribution. The 24.1% precision figure is more robust because it draws on 29 events rather than 8. The trigger is informative when weighted alongside other recession indicators (yield curve, credit spreads, payrolls trend) and uninformative in isolation.
What does this dataset not measure?
Initial claims count first-time filers for unemployment insurance among workers covered by state UI programs. They exclude gig workers, the self-employed, those who have exhausted benefits, and new labor-force entrants who haven’t yet worked. They do not capture hiring rates, quits, or job openings — for those, see JOLTS data. They are not a measure of the unemployment rate (which uses the household survey), nor of total employment (which uses the establishment survey).
How does this compare to the Sahm Rule?
The Sahm Rule (Sahm, 2019) uses a 3-month moving average of the unemployment rate, triggering when it rises 0.5 percentage points above its trailing 12-month minimum. Both rules use a short MA crossing a longer MA. The Sahm Rule operates on monthly data with a lower signal frequency and historically tighter precision. The claims-based 4w/52w rule operates on weekly data with higher false-positive rates. They are complementary: claims fire first, the Sahm Rule confirms.
Source
Related Eco3min Research
Last updated — 31 May 2026
Disclaimer – Financial Information: The analyses, commentary, and content published on eco3min.fr are provided for informational and educational purposes only. They do not constitute investment advice or a solicitation to buy or sell financial instruments. Past performance is not indicative of future results. All investment decisions involve risk and are the sole responsibility of the reader.