What this page counts

A composite score tells you how the economy reads once its parts are weighted together. It cannot tell you how many parts agree. Two months can score identically — one because nearly everything improved a little, the other because two heavily-weighted inputs surged while most of the rest slipped. Economic Breadth answers the second question and only the second question: how widespread is the movement?

Each member indicator casts one vote a month — improving, deteriorating, or abstaining — judged against its own history, not against the others. The page reports the tally, plus net breadth: improving minus deteriorating, as a share of the whole electorate. Counts, never probabilities. Historical ranges, never forecasts.

Who votes: the membership line

Economic Breadth counts what the economy does — not what markets price, and not how people feel.

• In — measurements of real activity: production, employment, income, sales, housing activity, trade, inventories, and credit quantities and behaviour (loan volumes, delinquencies, reported changes in lending standards).
• In — reported-activity surveys. The test is the question the survey asks: “what happened at your firm this period” is a measurement collected by questionnaire. Lending standards (SLOOS) and the ISM surveys qualify.
• Out — market prices and their derivatives: equities and equity trend constructions, interest rates, yield curves, spreads, commodities, valuation ratios. A census should count measurements, not priced opinions about measurements. Those indicators keep their own pages elsewhere on the site.
• Out — opinion and expectation surveys: sentiment, confidence, inflation expectations. Mood can decouple from activity for years; letting mood vote would pollute the activity count.

Where the same economic fact is stored in several transforms (level, month-over-month, year-over-year), exactly one representative votes — one voter per fact. Series with no honest “good direction” do not vote at all: the roster census of record classes them ambiguous-polarity, and among them are the personal saving rate, housing inventory, and the CPI and PCE inflation statistics. Membership moves only by amendment to the decision record, never by data availability.

The electorate

Twenty-four voters carry the census today.

Three more are admitted but abstaining until they have enough history of their own to be judged against — they enter automatically, under the same rule as everyone else:

How a vote is cast

A member's vote is the direction of its change over a three-month window, in its own polarity, measured at its own native cadence — a quarterly series votes on its quarter-over-quarter change, and that vote stands until its next release.

The dead band. A change only counts if it is large enough to matter for that series. The threshold is the median of the member's own history of absolute changes, computed only from periods strictly before the month being judged — so the band is calibrated per member, never across members, and never with data from the future. Movement inside the band is an abstention. Chatter becomes silence; that is the anti-noise device.

Abstention is first-class. Abstainers stay in the denominator, so a quiet economy honestly shrinks net breadth toward zero rather than exaggerating whichever few members happened to move. Abstaining counts are displayed, never folded away.

Burn-in. A member needs 60 months (or 20 quarters) of its own history before it can vote — no member is judged against a threshold built from a handful of observations. Months in which fewer than 12 members had burned in are not a census at all and are not stored: the record begins in December 1963, when the twelfth member qualified, and the electorate grows to today's 24 as later series reach their own burn-in.

No interpolation, ever. A member that misses its expected release beyond twice its normal cadence abstains rather than carrying a stale vote forward.

Why a three-month window and a median band

The window and the band were not tuned. Before any breadth number was computed, six candidate combinations were declared — a one-month or three-month window, crossed with a dead band at the 50th, 60th or 70th percentile of each member's own change history — and the selection rule was frozen: eliminate any combination whose abstention share falls outside 20–60% (a census where almost everyone always abstains, or never does, is not measuring), then take the most persistent net-breadth series among the survivors.

All six survived the first filter, with abstention shares between 40.4% and 57.4%. The three-month window with the median band was the most persistent (month-to-month correlation 0.83) and won its tie-break against the 60th-percentile band on the lower abstention share. Crucially, the selection statistics are blind to the adoption tests below: the cell was chosen before the exam was marked.

What the record shows

Adoption was gated on two tests whose numeric bars were frozen before measurement. Both are stated here as historical counts.

Test 1 — is it more than a flat hum? Net breadth had to be persistent (month-to-month correlation of at least 0.70; measured 0.83), had to visit both extremes (at or below −25 in at least 4% of months, measured 4.5%; at or above +25 in at least 10%, measured 64.6%), and no single vote state could swallow more than 90% of all member-months (the largest, improving, was 44.1%). It passed on every bar.

Test 2 — do the most negative months differ from the rest?Months in the lowest decile of net breadth against their own prior history — 83 of the 632 months eligible for the test — were compared against NBER recession dating. Of those 83, 69 were followed by a recession month within six months (83.1%, against a 16.3% rate across all eligible months) and 71 within twelve (85.5%, against 22.0%). Both counts clear a placement placebo of 1,000 seeded random draws (p95 bars of 19 and 24) and a structure control that preserves the series' persistence while destroying its alignment (p95 bars of 8 and 10).

Twelve of those 83 months were not followed by a recession inside twelve months; three of them — February 2003, October 2015, February 2025 — sit clear of any recession in either direction. They are on the chart, ticked in red like all the rest.

What Test 2 does and does not mean

T2’s pass is the instrument agreeing with the definition of recessions — widespread activity deterioration is close to what a recession IS — not a predictive discovery. The 2019-11/12 extremes “preceding” COVID are the standing cautionary exhibit: no census predicts a virus. EB ships as an honest counting instrument; the no-lead-language firewall is permanent, and every EB surface speaks descriptively: counts, never probabilities, historical ranges, never forecasts.

Reading the agreement count

The page counts the months in which the breadth reading and the State of the Economy regime label sit on the same side of the activity line — net breadth at or above zero under an expansion-family label, or below zero under Contraction. Months labelled Unclassified make no claim and are not compared.

That count has a floor built into it: both measures sit on the expansion side in roughly nine months out of ten, so a high agreement rate is partly arithmetic rather than corroboration. The count is published because the disagreements are the informative months, and they are listed by date whether they flatter the instruments or not. Agreement here means consistency, never correctness: neither measure validates the other, and neither is evidence for the other. They are different questions asked of the same economy — how it reads once weighted, and how many of its parts are moving.

Economic Breadth takes no State of the Economy quantity as input. The regime label is joined only for this comparison, at the display layer; no breadth number is computed from it.

Forward Stress and Economic Breadth ask different questions

The site's Forward Stress composite reads inputs that carry expectations — credit spreads, the yield curve, financial conditions — and asks whether stress is building. Economic Breadth refuses those inputs by charter and counts realized activity only, asking how widespread the movement actually is. The two are not rival answers to one question; they are answers to two.

That difference shows up in when each one moves. Prices embed what participants anticipate, so Forward Stress is built to register strain early — that is its stated job. A census of realized activity cannot do that and does not try: it can only report what has already been measured, which is why it either corroborates that strain once it reaches activity, or does not. Neither reading validates the other, and neither is evidence for the other.

When the two sit on opposite sides, the divergence itself is the information — priced expectation and realized participation are disagreeing about the same economy. This page shows such disagreements rather than resolving them, and it draws no conclusion about which one will turn out to have been right.

Basis, and what it forecloses

The history is computed on the revised basis: each month uses the values as they stand today, not the values as they were first published. A census month refines the same way — its count moves as member data arrive and revise — until the month record settles it. That choice is declared on the page with a basis chip and it is permanent — which is precisely why no claim on this page runs ahead of the data. A revised-basis history cannot support a statement about what breadth would have told you in real time, because the numbers it counts were not all knowable then.

The live census — the page hero's basis (EB-B4) — counts every member at its most recent monthly vote: the charter's own persistence, a vote standing until the member's next release, so the full 24-voter electorate always counts and the headline refreshes as releases land — the live census always counts each member's newest available vote, even where the month record's staleness rule would abstain it; the record basis remains the staleness-guarded truth. The vote rules are the monthly engine's own — the same 3-month change against the same dead band; nothing new is computed. When every member's newest observation sits inside one calendar month, the live census equals that month's row exactly. The monthly table is the record; the live census is a count, never a signal.

So Economic Breadth makes no such statement, and never will. It reports what the count was, month by month, alongside what followed — in counts, with the misses shown. It does not forecast, it does not signal, and it does not instruct.

The decision records: the Economic Breadth charter (EB-DR-1), the membership ruling and frozen survey specification (EB-DR-2), and the verdict (EB-DR-3). The prior-art census, the roster census, the survey report, and the build gates are committed alongside them.