A GDP-Weighted Multi-Signal System for Real-Time Recession Detection — the plain-English companion

the paper (DOI) · code & data repository

Companion to the research paper of the same title. This is education, not investment advice. Nothing here tells you what to buy, sell, or predict. It explains what the paper found, how the checking worked, and what the results do and do not mean.

The claim in one sentence

A recession detector built from the economy’s earliest paper trail — weekly unemployment filings, state by state, with two independent alarms required to agree — can speak months before a recession is confirmed, at a price the paper measures rather than hides.

The referee who calls the game a year late

Start with a strange fact about how the United States handles recessions: nobody tells you one has started until long after it has.

A recession, roughly, is a stretch of time when the whole economy shrinks — businesses sell less, factories slow down, and many people lose their jobs at once. There is an official referee for this. A committee of economists at an organization called the National Bureau of Economic Research studies the data and declares, officially, when each recession began. The committee is careful and its calls are trusted. But careful takes time. The recession that began in December 2007 — the one that became the worst downturn in generations — was not officially declared until December 2008, a full year in. Even the 2020 recession, the most obvious one in living memory, took four months to announce.

Think about what that means. If you waited for the referee, you would learn a recession had started long after it had already reached your job, your family, your town. The whistle blows after the play is over.

So the useful question is not “are we in a recession?” — by the time that gets answered officially, the answer is old news. The useful question is: what is the earliest reliable sign that this kind of trouble has begun, and how do you read that sign without getting fooled by every random wobble in the data?

This paper builds a machine for that. And then it does something unusual. The machine was built to test a bold claim someone else had made — that a design like this catches every modern recession early, with almost no false alarms. The paper ran the test, the test came back negative, and the paper printed the failure in its own results, in bold. What survived the failure turned out to be more specific, and more useful, than the claim that died. Watching that happen — the test, the failure, and what got kept — is the story this companion walks through.

The paper trail layoffs leave

To detect a recession early, you need data that shows up early. Most economic numbers do not. The size of the economy is reported months late and then corrected repeatedly. The famous monthly jobs report comes from surveys, gets revised, and covers a month that is already over.

But there is one number that arrives almost immediately, because it is not a survey at all. When a person is laid off in the United States, they can file for unemployment insurance — a government program that pays part of their lost wages while they look for new work. Every filing is a form, counted by the state that received it, and the weekly counts are published by the Department of Labor within days.

That count is a paper trail. It is not somebody’s estimate of how many layoffs happened; it is the layoffs themselves, leaving receipts. And since recessions produce layoffs in volume, the receipts start piling up near the very beginning of the trouble — weeks or months before slower numbers catch it, and long before the referee speaks.

The raw material for this whole paper, then, is simple to say: the weekly count of new unemployment filings, from every state, going back to 1967. More than half a century of receipts, all of it public data that anyone can download.

Two habits that keep the clock honest

Raw weekly data has two traps in it, and the paper deals with both in ways worth understanding, because each choice was made for a reason.

The first trap is the calendar. Layoffs have seasons. Retail workers are let go after the holidays every January; construction slows every winter; school-year jobs end every June. If you looked at raw filings, you would see a scary spike every single January that means nothing. The usual fix is statistical smoothing — mathematically sanding the seasonal bumps out of the data. But smoothing has a cost: it blurs sharp turns, and sharp turns are exactly what a recession detector is hunting. So the paper uses a plainer fix: compare each week to the exact same week one year earlier. This January’s filings against last January’s. If this January is much worse than last January, that is not the season talking — something changed. The seasons cancel out because both sides of the comparison contain the same season.

The second trap is gaps. Real-world data has missing weeks and irregular reporting. If a program measures time by counting rows in a spreadsheet — “eight rows back” — a missing row silently shifts every date. So the paper’s rule is that all time arithmetic happens in actual calendar days, never in row counts. A gap in the data cannot quietly bend the clock.

Neither choice is glamorous. Both exist so that when the detector fires, the reason is the economy and not a bookkeeping accident.

Sealing the data before anyone looks

Here is a problem every data researcher faces, and most readers never hear about: government data changes after it is published. Agencies correct errors, refine estimates, and quietly revise history. The unemployment filings for some week in 2001 are not necessarily the same number today that they were in 2002, or in 2015. Each snapshot of the data at a moment in time is called a vintage — like a wine labeled by year.

Why does that matter? Because if your data can shift under you, your results can shift too, and nobody can check your work, since they may be holding a different vintage than you were.

The paper’s answer is to freeze everything. Every input file was downloaded once, at a stated date, and then sealed. Sealing here means something concrete: each file is run through a standard mathematical fingerprinting method that produces a unique code from the file’s exact contents — change even one digit inside the file, and the fingerprint changes completely. All the fingerprints are recorded. Anyone who later downloads the paper’s data can re-run the fingerprints and confirm, to the digit, that they are holding the exact files the paper used.

Think of it as an exam sealed in an envelope before the students arrive. Everything downstream — every test, every number, every conclusion — runs on that one sealed vintage. Keep the sealed envelope in mind; it comes back at the worst possible moment in this story, and the moment is the point.

A smoke detector in every room, sized to the room

Now the machine itself. It has three layers, and the first two do the detecting.

The first layer is a national tripwire. Add up new filings across the whole country, compare to the same week last year, and trip when the increase is large and lasts. “Large and lasts” is spelled out exactly: the default setting trips when national filings run at least 10% above last year for at least 4 consecutive weeks. Consecutive is literal — a missing week breaks the streak and the count starts over. The streak requirement exists because one bad week is often just noise; four in a row is a pattern.

The tripwire is fast, but it has a blind spot: it cannot tell trouble that is spreading from trouble that is contained. An oil bust can hammer Texas. A hurricane can spike one state’s filings. Neither is a national recession. A recession is trouble that propagates — that spreads across states and industries instead of staying put.

So the second layer treats each state as its own room, with its own smoke detector: is this state’s filing count worse than a year ago, right now? Then it asks the key question — not “how many rooms have smoke?” but “how big are the rooms with smoke?” Each state is weighted by its share of the national economy. The measuring stick is a standard government figure called GDP (gross domestic product — the total value of what an economy produces).

A small example shows why the weighting matters. Imagine a country with only two states: one produces 90% of the economy’s output, the other 10%. If the small state’s detector fires, 10% of the economy is in a deteriorating room — concerning for that state, but not a national emergency. If the big state fires, 90% of the economy is in trouble. Same number of alarms; completely different meaning. Counting states would treat those two situations as identical. Weighting by economic size tells them apart.

The result is the paper’s propagation index: at any moment, the share of the national economy located in states that are deteriorating. When that share climbs past a set level and stays there — the default asks for at least 25% of the economy in deteriorating states over an 8-week test — the second detector fires. Smoke is not just present; it is spreading through rooms that matter.

The two-key rule

Here is where the layers combine, and the combination is the heart of the design.

Neither detector alone is trustworthy enough. The national tripwire is fast but jumpy. The propagation index is steadier but can drift. A detector that cries wolf gets ignored — which makes it worse than no detector at all, because it burns the trust an alarm needs to be useful.

So the system uses a rule you have met elsewhere in life: two keys. Some consequential actions require two people to agree — two signatures on a big check, two officers turning keys. The system’s version: an alert only stands when the national tripwire fires and the propagation index confirms it, in the right order, within a set window of time. One alarm proposes; the other must second the motion.

Why does requiring agreement help so much? Because the two detectors are fooled by different things. A data quirk that jolts the national count usually does not simultaneously push a quarter of the economy’s states into year-over-year decline; a slow regional drift that raises the propagation index does not usually coincide with a sharp national spike. When the causes of false alarms are mostly independent like this, the chance of both alarms being wrong at the same time is far smaller than either being wrong alone. Errors have to multiply, and multiplying two small chances gives a much smaller one. That is the mathematical reason confirmation works, and one of the paper’s companion research papers works it out in general.

The price is paid in the open: waiting for the second key always costs some speed. The system chooses to pay it. How much cleanliness that purchase actually buys is a question the paper answers with a number, a few sections from now.

Calling the shot before the swing

One more piece of machinery, and it is the one that makes everything else believable.

Suppose a researcher builds a detector, tries a hundred different settings, and publishes the one that happens to look best on past data. Is that detector good, or just lucky? You cannot tell — and neither can the researcher. With enough knob-turning, some setting will always fit the past beautifully and mean nothing about the future. This is the quiet way most too-good-to-be-true results are manufactured, often with no bad intent at all.

The defense is called pre-registration: write down, before looking at the results, exactly what will be tested, what settings will be used, and what would count as success or failure. Then run it once and report what happened. It is calling your shot before the swing. A called shot that lands means something; a shot called after the ball is already in the seats means nothing.

This paper pre-registered its design. That means the signal definitions, the default settings, and the pass/fail rule for the big claim were all written down first. So was the full menu of alternative settings to be examined: a grid of trigger levels from 10% to 18% and waiting periods of 3 to 5 weeks — 54 combinations in all, chosen in advance. The handful of small definitional choices made during the work are disclosed in the paper as dated amendments, each fixed before the test it affected was run. The point of all this ceremony is single: when the results arrive, no one — including the author — was able to steer them.

The bold claim under test

The system descends from an earlier, unpublished construction. That earlier manuscript claimed a remarkable record: its design had detected every one of the eight U.S. recessions since 1967, ahead of the Sahm Rule, with one false alarm to the Sahm Rule’s seven.

Every recession. Earlier. Cleaner. If true as stated, that is close to the best of both worlds, because — as the rest of this companion will show — earliness and cleanliness usually pull against each other.

Claims like that deserve re-testing, not repeating. So the paper fixed its pass/fail rule in advance: the rebuilt system must detect all eight recessions, with fewer false alarms than the Sahm Rule, at the source’s own settings. Then it sealed a fresh vintage of all the data and ran the test once, committed to printing whatever came out.

The test says no

The default settings detect seven of the eight recessions. The 2001 recession is missed. On false alarms, scored by identical rules for both sides, the system produces 4 false-alarm episodes to the Sahm Rule’s 2. Both halves of the claim — all eight, and cleaner — fail. The paper prints the word REFUTE in its results.

The autopsy of the 2001 miss is where the sealed envelope comes back. Both of the system’s detectors actually fired around 2001 — the tripwire tripped, the propagation index rose. But the two-key rule requires the confirmations to land within a set window of each other, and on the freshly sealed data vintage, the gap between the two firings fell just outside that window. On the source’s older vintage, the same gap had fallen just inside it. Nothing about the design changed. The history itself had been revised — those quiet government corrections — by just enough to move a borderline case across the line.

That is what the paper calls a knife edge: a result that flips depending on which snapshot of the data you happen to hold. And the plain reading cuts against the original claim in a deeper way than one miss. If “detects all eight” was only ever true on one vintage of the data and false on the next, then the record was never as solid as it looked. It was balanced on an edge. The paper’s re-test is simply the first time anyone checked the balance.

Counting false alarms fairly

Before going further, it is worth slowing down on how false alarms get counted, because sloppy counting is where detector claims usually hide.

A false alarm, here, is an alert that fires when no recession follows. But “follows” needs a deadline — how long do you wait before declaring the alarm wrong? The paper’s pre-set rule: each alert gets six months. If the referee eventually dates a recession as having begun within that window, the alert was a catch. If not, the alert matures into a false alarm and is counted against the system, permanently.

Why insist on this? Because the tempting alternative — only counting the alerts that happened to line up with known recessions — smuggles the answer key into the test. You would be using the referee’s final calls, which arrive a year late, to decide which alarms “count,” which is knowledge no real-time user of the detector could have had. The six-month maturation rule forces the scoring to live in the same fog the detector lives in.

Scored this way, the national tripwire alone carries 6 matured false alarms across six decades. And here is a detail that changes how you should feel about them: every single one corresponds to a nameable real-world event — an actual surge of layoffs that genuinely happened but did not spread into recession. The detectors are not seeing ghosts; they are seeing real fires that stayed contained. The same is true of the Sahm Rule’s two false alarms. One was a strange echo of layoffs during the slow recovery after 2001. The other was a trigger in mid-2024 that many people discussed at the time — and which matured, by the rule, as a false alarm.

Fifty-four ways to set the dials

One configuration failed. The next question is the one that decides whether anything here is worth keeping: did it fail because the design is broken, or because those particular dial settings are?

This is where pre-registering the whole grid pays off. The paper had committed, in advance, to examining all 54 combinations of trigger level and waiting period. Nothing about running the full grid after the failure is a do-over — checking the grid was always part of the plan, which is exactly why its answer can be trusted.

The answer: 36 of the 54 combinations detect all eight recessions. Two-thirds of the settings space does what the default could not. The design can do it; those settings don’t.

And one more pattern, stronger than it first sounds: across the entire grid — every combination, every decade — every miss anywhere is 2001. No setting misses any other recession. Fifty-four different ways of turning the dials, and the only recession any of them ever fails to catch is the same one. That one is the dot-com bust — a downturn whose early damage was unusually concentrated in one industry and unusually fast, which is the specific shape of trouble that a spread-across-states detector handles worst. When 54 differently-tuned versions of a machine all stumble in exactly one place, you have learned something real about the machine and something real about that place.

Rare, but real: clean and complete together

The grid answers a second pre-registered question: how often does a single configuration achieve everything at once — all eight recessions caught and zero matured false alarms?

In 13% of the space. That is far below the 76% the source had claimed — another correction the re-test forced into the open. But it is not zero, and the clean-and-complete settings are scattered across different eras and dial positions rather than clustered in one lucky corner. A property that shows up across a pre-committed grid is structural. A property that shows up in one hand-picked cell might be luck. The 13% is the first kind.

Where does the cleanliness actually come from? The paper isolates the mechanism with a comparison. Take the same two detectors and connect them with a simple OR rule — either alarm alone counts. Result: 7 false alarms. Now require the second key. The confirmed system’s false-alarm record drops to near zero. Same detectors, same data; the only difference is the requirement of agreement. That is the two-key rule caught in the act of doing its job, with the price and the purchase both on the receipt.

And the grid draws one boundary that deserves to be stated exactly as the paper states it. Across all 54 configurations, no setting achieves zero false alarms and a positive average head start at the same time. You can be early on average, or perfectly clean — never both, anywhere in the tested space. Any detector advertised as both is describing settings this grid does not contain. That boundary is the true shape of the problem, and everything in the next section lives on it.

The confirmation step

The Sahm Rule deserves its own explanation, because it plays a specific role here — and because it was never built to be what it is often used as.

It was proposed in 2019 by the economist Claudia Sahm as the trigger for automatic stimulus payments to households: a policy tool, meant to send help fast once a recession has begun. In her own words, “The star was always the stimulus check, not the indicator that other people named after me.” Others later adopted it as a recession indicator, and she has cautioned that it is an empirical regularity, not a law of nature.

It is built from the monthly unemployment rate — the share of people who want work and cannot find it — averaged over three months. It fires when that average has risen half a percentage point above its low point of the past year. By Sahm’s own definition, when it fires, a recession has already begun.

Its record across the same eight recessions: all eight detected, only 2 false alarms in six decades — and an average firing date 99 days after the recession had begun, ranging from about one month to about eight, recession by recession.

That lateness is not a flaw. A rule that releases government payments has to be reliable before it is early, and confirming that a recession has begun is the job it was designed to do. So this system does not try to beat it: judging an early-warning tool as “faster than the Sahm Rule” would measure the Sahm Rule against a job it was never built for. The claims-based system is a tripwire — built from faster, rawer material, and jumpier for it. In the ladder described below, the two work in sequence. When the tripwire catches a recession, it speaks months before the Sahm Rule confirms it — and the Sahm Rule’s confirmation is what an expensive decision should wait for.

So the summary is not “this system beats the Sahm Rule.” The summary is a trade, with both sides priced. On one side: months of earlier warning. On the other: more false alarms, and, at the default settings, one missed recession. The 54-cell map shows how the dials move you along that trade. The paper’s contribution is not a winner; it is the price list.

The clock between the alarms

Requiring two alarms creates a bonus the paper had no need to build separately: information in the timing between them.

When the national tripwire and the propagation index fire almost together, the damage arrived everywhere at once — the signature of a sudden shock, like the 2020 shutdowns. When the tripwire fires first and the propagation index takes a long while to confirm, the damage started somewhere and traveled — the signature of a slow, spreading downturn, like the one that gathered through 2007 and 2008. The gap between the two keys turning is a free diagnostic: it tells you not just that a storm has arrived, but what kind of storm it is. The paper reads this gap across all eight historical recessions and shows it sorting the sudden ones from the spreading ones, at no extra cost — the machine’s own internal clock doing double duty.

Finding the room where the fire started

The third layer answers the question people ask first when trouble is confirmed: where did this begin?

The method uses a different public dataset — earnings by industry, in each state — and asks, for each recession: in the year before the recession’s peak, which industry was already declining, and across how many of the declining states? The industry whose pre-recession decline covered the largest share of the deteriorating map gets a score the paper calls its footprint: the fraction of declining states in which that industry was among the decliners. A footprint near 1.0 means the industry’s trouble was nearly everywhere the trouble was.

For the 2007 recession, the answer is emphatic: Real estate, with a footprint of 0.87 across 46 declining states. The paper then checks the finding against an entirely separate government dataset — an independent count of employment by industry — and prints the check’s results state by state, including the one state where the two datasets disagree. The runner-up industries and their scores are printed too, so a reader can see how close the race was rather than being told to trust the winner. What the machinery has done, working only from data available before the peak, is rediscover what history knows: 2007 was a housing-led recession, and the map said so at the time.

Just as telling is where the method declines to answer. For the 2001 recession, the highest footprint any industry achieves is 0.196 — a noise-level score. The paper says plainly that no strong epicenter is identifiable: the dot-com damage was too concentrated and too fast to spread a broad earnings footprint before the peak. For 2020, the top industry reaches only a middling 0.60, and the paper leaves it below the bar rather than rounding it up to a story. A locator that names a culprit only when the signal is overwhelming, and says “cannot tell” otherwise, is a different kind of instrument from one with an answer for everything. And the difference is enforced by a scoring bar set in advance, not by mood.

Three steps of alarm, and a promise in writing

What does all of this become going forward? Not a prediction in a press release — a standing procedure, with its own failure conditions written down.

The paper registers a three-step alert ladder. Step one, WATCH: the standard national setting (10% for 4 weeks) confirmed by the propagation index (25% of the economy, 8-week test). Step two, ELEVATED: a stricter national trigger (18% for 3 weeks) confirmed the same way. Step three, CONFIRMED: the Sahm Rule itself crossing its public half-point trigger — the slow thermometer as the final word.

The registered claim: ahead of the next recession the referee eventually declares, the ladder escalates in that order. And the failure conditions are part of the registration. A WATCH that ELEVATED never joins is, by the written rule, a self-cancelling false alarm that counts against the system. A recession that arrives in the fast, concentrated, 2001-like shape is named in advance as the design’s known weak spot. Even the starting position is frozen and printed. At the paper’s data vintage, the national signal sits at −8.47% — filings well below last year. The propagation index sits at 2.11% — almost no economic weight in deteriorating states. The Sahm value sits at 0.067, far from its 0.50 trigger. Every light green. A quiet board, timestamped, with the rules for judging it already public.

Anyone can keep score from here. The data is public, the rules are published, the pipeline is downloadable, and the referee’s eventual call will settle it. The future does the grading.

What the checking caught in this paper

The general machinery every paper in this series runs through — the sealed data, the locked ledger of numbers, the verification program, the adversarial review, the from-scratch rebuild — is described once in the series’ shared verification note, which follows every companion on LaggingTruth.com. What belongs here is what that machinery caught in this paper specifically.

The adversarial review returned twenty-five findings; twenty-four led to fixes, and one was rebutted with the reasoning committed alongside it. One of the twenty-five was a count in the manuscript that had been typed by hand and disagreed with the data it described — the fix now forces such numbers to come from the ledger like all the others. The from-scratch rebuild failed on its first run, exposing result files that had never actually entered the repository and a copying quirk that silently altered sealed files; both produced permanent repairs, and the re-run rebuild reproduced every ledger value identically. And the document builder was once caught reporting success while leaving a stale file on disk; it now checks its own output’s timestamp. Each break added a safeguard — which is why the checking is layered in the first place.

What this cannot do

The boundaries, stated as plainly as the results. Silence from this system is a statement about one thing only — spreading deterioration in unemployment filings — and never a promise of safety; recessions can arrive by roads this instrument does not watch. The known weak spot is on record: a fast, concentrated onset, the 2001 shape, is exactly what a breadth-based, confirmation-gated detector handles worst. All results live on one sealed vintage of revisable government data, and the 2001 knife edge is the standing reminder of what a revision can do to a borderline case. The evaluation scores the system against the referee’s final dating, which is itself a product of revised data. How the system would fare scored against only first-draft data is left as an open question, stated in the paper’s limits. And none of this is investment advice or a forecasting service; it is a measured, pre-registered instrument for one narrow question, with its prices printed on the label.

The takeaway

A machine built from the economy’s earliest paper trail — weekly unemployment filings from every state, each state weighted by its share of the economy, with two independent alarms required to agree before it speaks — was constructed to verify a bold claim and refuted it instead: on freshly sealed data it catches seven of eight modern recessions, not all eight, and its false-alarm record is worse than the claim said, not better. What survived is sharper than what failed. Two-thirds of its pre-committed settings do catch all eight, and the only recession any setting ever misses is the same fast, concentrated 2001. Its near-perfect cleanliness in the confirmed configurations is mechanically traceable to the two-key rule. The timing between its own alarms sorts sudden storms from spreading ones. Its industry layer named housing as 2007’s origin from pre-peak data, and declined to invent an answer for 2001. And the whole instrument now stands behind a dated, public, self-cancelling prediction about the next recession, checkable by anyone with the data and the patience. The test said no. The paper kept the no, and kept what the evidence actually earned.

This companion is licensed CC BY-NC 4.0. The research paper it accompanies is licensed CC BY-NC-ND 4.0, and the analysis and verification code is MIT-licensed. Education, not advice: nothing in this document is financial advice, an investment recommendation, or a forecast.