CPI Lab

Frequency measure

"Share of weighted CE Diary purchase entries per consumer-unit-week, mapped UCC → ELI → published item stratum." For Diary year y, with consumer-unit-weeks c, final weight FINLWT21, and n_c,u the number of expd purchase entries CU-week c recorded for UCC u: entries_per_cuweek(u) = Σ_c n_c,u W_c / Σ_c W_c.

Index construction

"December-chained Laspeyres over the same leaf universe as the P0 replica" — I_t = I_Dec(y) · Σ_i w_i P_it/P_i,Dec(y) / Σ_i w_i, chained through each December, built identically to the official replica so the two are comparable. Leaf universe = cpi_weight WHERE is_leaf=1 AND matched=1, ~179 items covering ~97% of the CPI; the missing ~3% is unsampled strata with no published index and is renormalised away.

The bill floor (bill_salience)

"max(diary frequency, 1/4.345 entries per CU-week) for the listed items only" — i.e. each of 17 documented recurring-bill items is treated as at least one purchase event per month (0.23015 entries per consumer-unit-week), unless the Diary already sees it more often. The Diary is a two-week purchase log: a household writes down the eggs but not its rent, so owners' equivalent rent scores 25.8% of spending and 0.00% of purchase entries. The floor binds for all 17 items, taking the group from 1.04% to 11.27% of the basket.

This is an assumption, not a measurement. "One bill a month" is a stipulation about what a bill is; the Diary cannot confirm or refute it. It also gives a rent payment and a garbage bill exactly the same weight, which is crude.

The 17 items: rent, owners' equivalent rent, tenants'/household insurance, motor vehicle insurance, health insurance, electricity, utility gas, water & sewerage, garbage & trash, wireless phone, residential phone, internet, cable/satellite/streaming, college tuition, elementary & high-school tuition, day care & preschool, and state motor-vehicle registration & licence fees. Deliberately excluded as not bills: hotels, video and music subscriptions (they mix subscriptions with one-off purchases), parking fees, and episodic medical services.

Multi-ELI split

"Expenditure-weight split of the 29 multi-ELI UCCs (see sensitivity variant split_equal)". ELI → published item is 'SE'+ELI[:4], walking up to [:3] and [:2] when the finer stratum is unpublished.

Aggregation level

"Published item (see sensitivity variant agg_class)".

Coverage gate

A month is priced only if leaves carrying ≥ 90% of the scheme's weight have both a December base and a current index. Only 11.3% of the official-replica weight had an October-2025 observation, so that month is dropped from every scheme, every sensitivity variant and every bootstrap replicate rather than being priced off a handful of items.

Sampling error

σ = √( Σ_r=1..44 (θ_r − θ)² / 44 ) over the 44 CE PUMD balanced-repeated-replication weights, where θ is the gap (scheme YoY − official replica YoY) in percentage points. PUMD replicate weights are already doubled half-sample weights, so no Fay factor is applied; under the Fay k=0.5 convention every SE here would be exactly 2× larger.

How much do the arbitrary choices matter?

Eight specification variants are rebuilt as separate indexes.

Expenditure weight is the December relative importance of the item (the published CPI's own weight). Frequency weight is the item's share of weighted CE Diary purchase entries per consumer-unit-week, normalised to 100 over the same leaf universe. Both are percentages of the basket for the latest month.

Purchase-entry counts partly measure classification depth: an item BLS splits into several UCCs accumulates more entries than an equally common item recorded as one. The agg_class sensitivity variant measures this and finds it small in aggregate (0.0004 pp on the headline) — but that is an aggregate result and can still be large for an individual item.

About 5% of Diary entries have no concordance row and are dropped; 99.4% of the concordance-mapped entry mass lands on a published leaf.

contribution = weight × item YoY / 100, computed separately with the expenditure weight (contrib_exp) and the frequency weight (contrib_sal); the bar is gap_driver = contrib_sal − contrib_exp. Blue bars push the frequency-weighted index above the official replica, red bars pull it below. Bars sum across all ~175 items to approximately the total gap; the 12 largest in each direction are shown.

Universe: cpi_weight WHERE is_leaf=1 AND matched=1 for weight_year = calendar_year − 1 (the December relative importance of year Y applies to the months of Y+1), joined to cpi_item_month — 179 items covering ~97% of the CPI-U. Weights are renormalised each month over the items that actually have a value.

The yoy view shown here is NSA, 12-month, and covered 175 items / 96.5% of the basket in the latest month. It is the view to show a consumer; the SA-based 3-month-annualised and 1-month views describe a smaller basket (~86%) because seasonally adjusted indexes exist for fewer strata.

Median / percentiles

"The value of the first item at which the cumulative weight reaches p" — the step definition, no interpolation, the same convention the Cleveland Fed uses for its median CPI.

Trimmed mean (16%)

Each item keeps the overlap of its weight interval [W_i − w_i, W_i] with [α, 1−α], so boundary items are split and the estimator is continuous in α; then a weighted mean over the retained weight, with α = 0.16 per tail. This is the standard "underlying inflation" measure; the Cleveland Fed publishes a 16% version, against which ours runs −0.13 to +0.34 pp over the last 12 months (expected: they trim ~45 aggregated components, we trim 142–175 leaves).

Ex-shelter universe

Drops SEHA, SEHC01 and every leaf whose cu_item ancestry contains SAH1 — in practice rent, owners' equivalent rent, lodging away from home and tenants' insurance, about 34.4% of the 2025 weight.

Trimmed means are not monotone in α. In July 2026 the 8% trim gave 2.75 but the 16% trim gave 2.95, because 15.2% of the basket was falling and widening the trim removed most of the negative tail. That is correct behaviour, not a bug — the two trims are two different descriptions, not a robustness ladder.

Weights only exist from December 2017 in cpi_weight; months before 2021 reuse the oldest available vintage and are a genuinely fixed-weight historical cross-section, not the weights of the day. Treat pre-2021 history as indicative.

Let x_it be item SA month-on-month change and w_it its renormalised weight, and define the contribution c_it = w_it x_it (percentage points). Then X_t = Σ_i c_it is our replica of headline SA MoM, and exactly Var(X) = Σ_i Cov(c_i, X), so share_i = Cov(c_i, X) / Var(X) and Σ_i share_i = 1. A negative share means the item consistently moved against the total, damping it.

Windows are 24 and 60 calendar months ending at the latest month; months with no CPI are dropped, never shifted, and a window needs ≥90% of its months present. X is not the published series — it omits the strata without SA indexes (~86% of the basket) — so the decomposition describes X, not SA0 exactly. Validated: Σ share_i = 1.000000000 for both windows; replica SA MoM vs published SA0 MoM since 2000-01 has correlation 0.972 and mean absolute difference 0.053 pp over 317 months.

rho

OLS slope of the item's monthly rate on its own one-month lag, with a constant, over a rolling 120-month window ending at the latest month (equivalent to a demeaned AR(1)). An item enters a window only with at least 96 monthly rates inside it.

Half-life

ln(0.5) / ln(rho) months; 0 when rho ≤ 0 — the deviation is gone the following month — and capped at 60 when rho ≥ 1.

The monthly rate, and the two source groups

Each item is assigned one rate definition for its whole history. sa items use the published seasonally adjusted month-over-month change (146 items, 86.0% of CPI-U). Items whose SA series is too short use nsa12: 100 × (ln Pt − ln Pt−12) / 12 (36 items, 10.2%).

The nsa12 rate manufactures persistence. It is a 12-term moving average, so applied to pure noise it yields rho ≈ 0.917 and a half-life of about 8 months analytically — the pipeline re-derives 0.896 / 6.3 months by simulation. That is why the two source groups are shown as separate faceted charts with separate rankings on this page and never sorted into one list: a ranking that mixes them is a chart of the filter, not of the basket.

Raw vs relative universe

"Relative" repeats everything on the item's rate minus the headline rate put through the same filter, so a slowly-fading common shock is differenced out. Spearman correlation between the two half-life rankings at the latest month is —: the ordering is robust, the levels are not. Note the mechanical bias — owners' equivalent rent is ~25% of the headline, so subtracting the headline subtracts a quarter of OER from itself, and the relative measure is biased toward finding low persistence for high-weight items. The honest answer for a large item is somewhere between the two. Both are published; never quote one alone.

Uncertainty

rho from 120 observations carries a standard error of roughly 0.09, so a half-life of 4 months and one of 6 are usually not distinguishable. lp12 (the local-projection slope at a 12-month horizon) and its Newey–West HAC standard error are in the data table; they are the only published uncertainty numbers here. Read the league table as indicative ordering.

Weights

With s_i the within-source-group rank percentile of rho, sticky CPI weights item i by wi · si and flexible CPI by wi · (1 − si). The _binary schemes split hard at the within-source weighted median; the _pooled schemes deliberately do not normalise within source and are shown as the sensitivity run that makes the nsa12 artefact visible. All schemes run on exactly the same item set, so they are comparable with each other and with the official replica.

No look-ahead

The classification used for the months of year Y+1 is estimated on the window ending December of Y, so the index never uses data from the months it is applied to.

Why owners' equivalent rent dominates the sticky index

OER is 25.2% of the basket and the most persistent seasonally adjusted item. Weight × persistence puts it at 33.4% of the sticky index's weight, with rent of primary residence adding another 10.5% — shelter alone is about 44% of our sticky CPI. Read every movement in it as, to a first approximation, a shelter index with a tail of vehicles and restaurants. It is not an independent read on the economy.

Why ours and the Atlanta Fed's differ — vehicles

The Atlanta Fed splits the basket by how often prices change (Bils–Klenow: sticky = repriced less often than every 4.3 months) — a fact about price-setting. Ours splits by how persistent the published index's inflation rate is — a fact about the time series. The classifications genuinely disagree on big items: used cars and new vehicles reprice constantly, so the Atlanta Fed calls them flexible, but their index has serially correlated monthly changes, so P3 calls them sticky. That single disagreement explains most of the divergence — in June 2021 our sticky CPI hit 6.31% against our flexible 3.12% because the used-car spike landed in our sticky basket and in their flexible one, and in April–May 2026 the same thing runs the other way. Treat the overlay as a sanity band, not a validation. Neither is "right"; they measure different things and the labels happen to collide.

Validation

This pipeline's own official_replica reproduces the published headline within ±0.15 pp in every one of the last six months. The Atlanta Fed rows use a ±2 pp sanity band, not a tolerance.

Source: the ALFRED vintage archive at the St. Louis Fed, which keeps every FRED vintage. Thirteen series are collected; headline goes back to 1972, the aggregates to 1996, and the nine item-level CUSR0000* series to 2011, so only the 2011+ window is comparable across all thirteen.

First print is the month-over-month change in the earliest vintage that contains that observation month. A series' oldest archived vintage also carries back-history that had already been revised before ALFRED ever saw it; those months are stored with true_first = 0 and excluded from every statistic, because including them would bias revisions toward zero.

Each vintage is reindexed onto a complete monthly calendar before differencing, so a month-over-month change is a true calendar lag. October and November 2025 are therefore absent from every vintage of the aggregates — there was no October CPI. Two series are exceptions and do have an October value, new vehicles and used cars, because those indexes are built from transaction data rather than field collection; this is detected from the data per series, not hardcoded.

Revision = latest − first_print in percentage points, averaged over every genuine first print for that calendar month. The magnitude is flat across the year — headline mean |revision| is 0.068 pp for January+February against 0.066 pp for the rest of the year, a ratio of 1.03. It is the sign that carries the seasonal signature: early-year prints are systematically revised down, late-year prints systematically up.

Month-of-year statistics pool across each series' whole history, which for headline mixes several seasonal-adjustment methodology regimes (X-11 through X-13ARIMA-SEATS, with intervention analysis introduced at different dates per series). The sign pattern is stable across sub-periods; the magnitudes are not directly comparable across decades.

Reporting rounds a monthly CPI change to one decimal, so the honest question is whether the printed number survives. Both the first print and the latest value are rounded to tenths and cross-tabulated; the full transition matrix for all thirteen series is in p4_transition and in the JSON.

The flag is |first_print − trailing 12-month mean of MoM in that same vintage| < 1.5 × SE. The trailing mean is computed from the first-print vintage, not from today's data, and requires at least 9 of the 12 prior calendar months. Standard errors are BLS's own annual median SE of a one-month change (Table 1V of the variance workbooks, parsed by P5), matched to the observation month's year.

Three series are matched exactly to a published group (headline, core, energy); the rest are components of, or spread across, a broader group, and are marked with an asterisk in the summary table. For those the group SE is a lower bound on the component's sampling error — narrower cells carry fewer price quotes — so their "within 1.5 SE" share is conservative, i.e. too low.

The one number

gap_yoy = CPI-U 12-month % change − C-CPI-U 12-month % change, in percentage points. Both indexes are taken as published; nothing is re-aggregated. The CPI-U leg is NSA, matching the C-CPI-U, which BLS never seasonally adjusts. All 12-month lags are calendar-aligned joins, never row offsets.

Which substitution this is

Upper-level only — shifting spending between categories. Swapping between varieties inside a category is already handled in both indexes by the same geometric means BLS has used since 1999, so that layer is not in these numbers.

The three publication stages

The chained formula needs to know what people spent that month, and that arrives about a year late, so every month is published three times: initial with the month itself, interim each quarter, and final 10–12 months later. The stage is carried in the footnote code on each data row of the flat file (I, U, blank), and is marked on the month axis. The newest number on this chart is the least reliable one.

Universe

There is no fine-grained C-CPI-U. The chained index does not exist below these 29 published aggregates, anywhere public — so "item level" here means these 29, not the ~179 leaf strata used by P1 and P2.

October 2025

No CPI and no C-CPI-U, with exactly one exception: new vehicles has an October value in both surveys, so that is the single 2025-10 cell on this chart. Everything else is absent by design.

The flat file only ever carries the current stage of each month, so it is a snapshot rather than a history. Publication history is therefore rebuilt from BLS's own monthly supplemental tables — Table 1C, xlsx from 2021-11 and pdf back to 2017-07 — each of which prints the index as it stood at that release. 107 files, publications 2017-07 … 2026-07. Nothing before 2017-07 is recoverable: the 2012–2016 archives contain no C-CPI-U table at all, and BLS publishes no consolidated revision history.

rev_yoy_pp is the number to quote. rev_pct compares the level published for that month then and now, and because the whole index path is revised together it is not a clean per-month statistic.

The two variants

raw is the items as published, and it is dominated by energy: fuel oil, gasoline and energy services appear to move first and everything else to follow, because energy is what moves the overall inflation number and the overall number is inside every series. common_removed takes each item's own idiosyncratic movement after regressing out the headline — and is the variant to read. Almost every energy→transport link is strong in raw and gone in common_removed: those items do not have a private relationship with each other, they are both riding the same tide.

Leave-one-out headline

The headline regressor has the item's own contribution removed. This matters more than it sounds: gasoline is 2.9% of the basket but its monthly rate has 19× the headline's standard deviation, so on the plain headline it largely regresses on itself, takes a coefficient of +16, and its residual becomes roughly minus-the-headline — which then shows up as a flat ρ ≈ −0.5 arrow from gasoline into every persistent service item at every lag from 1 to 12, the signature of a level artefact rather than a lead. The fix drops energy sources from 12 to 3 while leaving every genuine edge intact or slightly stronger.

ρ, lag and q

ρ is the cross-correlation at the best lag; ρ = 0.43 means roughly 18% of the follower's month-to-month wiggle lines up with the leader's three months earlier. Lag is the argmax of |ρ| over 1…12 months and is not sharply identified — neighbouring lags usually differ by a few hundredths. q is the FDR-adjusted p-value of a Newey–West HAC Granger-style F test; a low q means the leader says something the follower's own history did not already say. A high q with a high ρ — rent → OER, ρ 0.80, q 0.78 — means "these two move together with a one-month offset, but neither predicts the other beyond its own past". Both are shown because they answer different questions.

Stability

Everything is recomputed on four overlapping ~15-year windows. A pair passes a window if |ρ| > 0.25 or q < 0.05 there; an edge is stable only if it passes in at least 3 of the 4 windows, with the same sign and every best lag within ±2 months of their median. Note that 20 of the 70 stable edges do not clear the screen on the full sample — they pass in three windows but wash out when all 26 years are pooled.

rho_other

Every edge carries the same pair at the same lag measured in the other variant. An edge whose rho_other is near zero or of the opposite sign is a creature of the common-factor step, not a robust feature of the data. Three weak negative gasoline / fuel-oil edges survive into the main view and are residual artefact.

Colour and position

Major group is carried by position first — each group occupies one labelled arc of the ring — and by colour second, which is what allows eight group hues in a form where any two can end up adjacent. Every node is also directly labelled, and the full edge list is in the data table.

The arithmetic, exactly

For a pair with best lag L at month m: base is the mean of the PPI month-over-month rate over the 60 months ending at m. Each of the L months ending at m contributes its deviation devm−k = ppi_momm−k − base, discounted by the pass-through it has already delivered: an innovation k months old keeps the factor (β₁₂ − βk)/β₁₂, where β_k is the same local projection estimated at horizon k. The pair's contribution is Σk=0..L−1 devm−k · max(β₁₂ − βk, 0) · weight, which lands directly in basis points of the headline. This removes the double-count that a plain shock × β₁₂ would carry — β₁₂ is the cumulative 12-month response, so applying it in full to an innovation that has already passed through for months over-counts, and it also up-weighted longer-lag pairs mechanically. The freshest innovation (k=0) keeps almost all of its effect; one older than twelve months keeps none — the method assumes pass-through is essentially complete within a year. A pair whose response is entirely contemporaneous (β₀ ≈ β₁₂) contributes ~0: its moves are already on the shelf. The ±1 SE band uses |Σ dev·(β₁₂−βk)/β₁₂| · se(β₁₂) · weight, a deliberately conservative choice — the true SE of β₁₂ − βk is smaller because the two are estimated together.

β₁₂ — the pass-through coefficient

Jordà (2005) local projections: Σj=0..h cpi_momt+j = α + βh·ppi_momt + γ₁₋₃ cpi_momt−1..t−3 + ε, so β₁₂ is the cumulative pp response of the CPI item over 13 months to a 1 pp producer-price move. Overlapping horizons make ε an MA(h) process by construction, so the standard errors are Newey–West with a Bartlett kernel truncated at h + ⌊4·(n/100)^(2/9)⌋ and a small-sample factor — OLS standard errors would be meaningless here.

Lag 0 contributes exactly zero, by construction

22 of the 61 pairs peak at lag 0 — gasoline (ρ = 0.82), electricity, fuel oil, airline fares, milk, eggs, piped gas. For those the producer move and the consumer move land in the same month, so by the time the PPI print exists there is nothing left in the pipe and the pair contributes zero. That is the correct answer, not a coverage failure. The same applies to the 16 weak pairs (β₁₂ ≤ 0 or |β₁₂/se| < 1.96) and to the import-price overlays, which would double-count the PPI pair on the same CPI item.

One pair per CPI item, finer wins

The concordance deliberately nests (both "Meats" and "Beef and veal"). A pair is dropped from the sum — never from the tables — when another pair's CPI item is a descendant of it in the BLS item hierarchy.

The two bands

The ±1 SE drawn on the chart treats the β errors as independent across pairs. They are not — the pairs share macro shocks — so it is a lower bound, and the comonotone bound Σ|contrib_se| is printed beside it. Neither band allows for the βs being wrong in a regime sense, which is normally the larger uncertainty by a wide margin; that is what the rolling range in the pass-through table is for.

PPI is preliminary for four months

Every PPI index is recalculated with late reports at each release, so the most recent readings — exactly the ones this figure is built from — will move. The pipeline vintages them.

weak

Full-sample β₁₂ is ≤ 0 or not distinguishable from zero at |t| < 1.96. 16 of 61. Several are relationships you would have bet on — dental care, hospital services, car rental, motor-vehicle maintenance. They stay in the tables and are kept out of the sum. A concordance that only reported its winners would be worthless.

CPI leads

The correlation at a negative lag beats the one at lag 0 — the consumer price is moving before the producer price, so the pipeline story does not hold for that pair. 16 of 61, three of them in the sum (poultry, fish and seafood, canned fruits and vegetables). Negative lags are an alignment diagnostic only and are never chosen as the best lag.

NSA-dm

BLS has discontinued the seasonally adjusted twin of that PPI series (or the CPI item has no SA index), so both sides have their calendar-month means removed over the estimation sample instead. A crude X-11 substitute — without it, a January producer bump and a January consumer bump would correlate whether or not one caused the other.

β₁₂ above 1

Seven pairs have the consumer price moving more than one-for-one — cereals and bakery products at 3.5 is the extreme. Not automatically wrong: the PPI series usually covers a slice of what goes into the CPI item, so a 1 pp move in the measured slice can accompany a larger move in the whole thing. It does mean "pass-through" is a loose word for those pairs.

Confidence

A property of the hand-built concordance row, not of the statistics: high means the producer series is the right upstream price for that consumer item, low that it is the closest available proxy. Imported-goods pairs — apparel, footwear, consumer electronics, much of furniture — are labelled med or low because the domestic PPI is measuring the wrong upstream price; the BEA import-price overlays are published next to them for that reason.

Direct import content of a commodity

sd_c = M[c, F01000] / U[c, F01000] — the import matrix over the Use table, both at producers' prices, on the PCE column, clipped to [0,1]. For the 127 of 402 commodities with negligible PCE use the denominator falls back to total use excluding BEA's negative imports-reconciliation column.

Embodied (indirect) imports — the Leontief round

mi = ΣM[·,j]/x_j per dollar of industry output, re-expressed per dollar of commodity output through the Make table, then run through the domestic commodity-by-commodity total-requirements matrix: t = mc·L. Total content is st_c = sd_c + (1 − sd_c)·t_c — the dollar splits into an imported part and a domestically produced part, and only the domestic part carries embodied imports. Industry and commodity output sums must exclude BEA's total row/column; including it silently halves every intensity.

Why the PCE bridge matters

The margin share of a consumer dollar is not a constant. The 2017 detail bridge puts the producer's share of the purchaser price at 12.5% for games and toys — nearly seven eighths of that price is transport plus wholesale plus retail margin — and at 73.8% for nonelectric cookware. A uniform "goods at retail ≈ 30–40%" haircut would have mis-stated import content by a factor of two at both ends of that range. Margins and transport are domestic services and carry only the embodied content of the trade and transport commodities (2017: wholesale 0.053, retail 0.041, transport 0.094).

2017, on purpose

BEA publishes the detail Use / Make / import / total-requirements set only for benchmark years, so exposure is measured before every tariff in the timeline. That is an advantage for the event study — the tier assignment cannot be contaminated by the tariffs being studied or by re-sourcing in response to them — and a limitation for the descriptive series: any supply chain that moved after 2017 is mis-classified, in the direction of overstating exposure for anything that reshored. The 2007 and 2012 bases are computed too; mean share across the 179 leaves is 0.190 / 0.197 / 0.177.

Where the cutoffs sit

0.25 and 0.10 on total share, and they are not percentile splits. The distribution is bimodal — services and shelter pile up below 0.10, durable goods above 0.25 — so the cuts fall in the sparse middle. Moving the high cut to 0.30 would move 12 items; to 0.20, 21 items.

Coverage gate and the October hole

A month is dropped unless leaves carrying 90% of the tier's weight have both a December base and a current index. October 2025 (no CPI release) fails for every scheme; November 2025 additionally fails for the mid tier at 86.0% coverage, so the mid line has a two-month hole where the others have one. That is the gate working on a narrower weight base, not a bug. The all-items replica matches P1's official_replica to 0.0000 pp at every month.

The dependent variable

G_t = Σexposed w·x / Σw − Σdomestic w·x / Σw, where x is an item's NSA month-over-month change minus that item's own mean for that calendar month, computed over all months more than 12 away from the event so the event cannot demean itself away. NSA throughout, for both legs — mixing SA items with NSA-fallback items would put a seasonal wedge straight into the difference.

The regression

±24 months around the event; dummies for relative months −6…+12 with −24…−7 as the omitted baseline; Newey–West HAC standard errors, 6 lags. The cumulative 0…+12 effect and its SE come from the full coefficient covariance — which is why the cumulative path on the chart carries an interval only at its endpoint. Constructing a band for the intermediate horizons would mean inventing the covariance the pipeline does not publish.

Pre-trend test

OLS slope of G_t over relative months −6…−1 with HAC(2) standard errors; flagged as failing when |t| > 2. A failure says the two tiers were already diverging before the action.

Event selection and bundling

87 timeline rows → 28 events → 25 estimated. A row is eligible if its confidence is high or medium and |ad-valorem change| ≥ 5 pp. Each month contributes at most one event, chosen by duty impulse (|rate change| × annual import value covered), not by the largest rate change — a 50 pp rate on one partner moves the consumer basket far less than a 10 pp rate on everything. Every other row sharing the month is listed as a bundled action. Import values rank events and are never published as facts about trade volumes.

Known weaknesses, in the pipeline's own words

Tier assignment is not treatment assignment. Windows overlap. One-per-month bundling is a real loss of resolution, not a technicality. And the exposed tier is durable-goods heavy while the domestic tier is shelter heavy, so anything that moves durables relative to shelter shows up here regardless of trade policy.

Wages lag by 7–8 months, and it is structural

QCEW posts a quarter roughly 5–6 months after it ends, so the newest wage quarter here is 2025Q4 against CPI through 2026-07. The real-wage panel will always trail the price panel by two to three quarters. This is not a stale build.

The 2025Q3+ ownership fallback — read this before comparing quarters

MSA-level private average weekly wage is suppressed for every MSA from 2025Q3 onward (disclosure code -, published as literal zeros, which is why they are stored as nulls rather than as a −100% change). Those quarters fall back to total covered employment, which includes government. Where both bases exist they differ by 0.34 pp on average, so the latest cross-metro ranking sits on a slightly different footing than earlier quarters. The last fully private-basis quarter is 2025Q2, whose ranking is led by Urban Alaska (+3.8 pp), Phoenix (+3.5) and San Francisco (+3.4) with Seattle (−2.1) and Los Angeles (−1.5) at the bottom — a broadly similar picture, which is reassuring. A year-over-year change is never taken across bases.

Matching a quarterly wage to a monthly price

The CPI reading for a quarter is the last month inside that quarter for which the metro actually published. Always inside the quarter, so no information leaks in from the next one; always the freshest reading available; and data-driven rather than schedule-driven, so it self-corrects when BLS reassigned bimonthly slots in the 2018 revision and when a month is missing. The October 2025 hole is visible in it: every even-schedule metro's 2025Q4 row resolves to 2025-12, every odd metro's to 2025-11. real = wage_yoy − cpi_yoy; the exact ratio form differs by under 0.1 pp at these rates.

The composition flag

comp_delta = mean[t] − mean(mean[t−1], mean[t], mean[t+1]), flagged when |comp_delta| > 0.3 pp. Because t−1 and t+1 carry the complementary panel, the three-month centred average is close to composition-neutral and the residual is the composition effect. It is null — and the flag with it — when either neighbour is missing, including the months either side of the October 2025 hole and the latest month. 15 of 212 months are flagged, clustered exactly where you would predict: mid-2022 to mid-2023, when metro rates were furthest apart and the swap mattered most. The sign alternates — 2022-06 +0.73, 2022-07 −0.59, 2022-08 +0.49, 2022-09 −0.35 — which is the signature of a composition artefact rather than of inflation. In that stretch the even panel (Phoenix, Miami, Atlanta) ran roughly a point hotter than the odd panel (New York, Boston, Washington).

Rules

Only metros with a published value that month enter the cross-section, and at least 8 are required. Weights are constant 2020 CBSA population. The series starts 1998-01 because BLS redefined the metro pricing areas in 1998 and again in 2018; before 2018 it covers 20 metros, not 23, on the earlier definitions.

What a metro difference is worth

Metro indexes are built on much smaller samples than the national index and carry materially larger sampling error. A 0.3 pp difference between two metros in one month is not a finding. Metro CPI is also NSA only — there is no seasonally adjusted metro CPI at all — so only 12-month changes are used here and month-to-month moves are not comparable with the seasonally adjusted national headline.

"BLS variance tables: the median (across the year's published changes) standard error of the CPI price change, by item and horizon, U.S. city average. Published annually, one value per calendar year." Standard errors are deliberately not part of the stress composite: an annual number cannot be z-scored against a 12-month 2019 baseline.