How the Cost Index works

What this index is, and how it differs from an assessed benchmark

For each common restaurant ingredient, the index publishes a wholesale dollar level and a direction, with a confidence and a dated public citation. It is a decision tool for one question an operator actually asks: is this a real market move, or just my vendor?

The reference bar for proteins is Urner Barry, and we take its rigor seriously. But Urner Barry is an assessed, judgment benchmark: market reporters speak with participants and publish a normalized private assessment. Daily timeliness is its strength; the closed method is the trade-off. We do not try to out-quote it on daily granularity. We out-explain it — every number traces to a public report, every estimate is labeled as one, and every gap is named with its structural reason.

That posture is deliberate. An opaque, self-reported commodity index can fail badly: the Georgia Dock broiler index was discontinued in 2016 after questions about thinly-sourced, unverified contributions, with USDA and Urner Barry cited as the better-verified alternatives. Our position is the inverse — verification is the moat, and opacity is the risk.

Source: Georgia Dock discontinuation & the assessed-benchmark comparison

USDA Agricultural Marketing Service & Urner Barry — public reporting on poultry price benchmarks.

The Georgia Department of Agriculture's "Georgia Dock" broiler price was discontinued in 2016 after scrutiny of its thin, self-reported contributor base, and was widely replaced in contracts by USDA and Urner Barry quotes. Urner Barry is described here as an assessed, judgment-based market benchmark — a fair characterization of how a price-reporting agency operates, used to contrast its private method with this index's public, reproducible one. We credit both organizations by name rather than deep-linking a specific page; see USDA AMS for the public market reports this index builds on.

Specification & quotation basis (what, exactly, is priced)

The hallmark of a defensible methodology — and where a silent number falls short — is stating precisely what each quote represents. Every ingredient carries an explicit specification:

• Commodity & grade. The exact source term and grade — meat cuts by their IMPS/NAMP number (for example, a striploin is IMPS 180 boneless strip), produce by its AMS commodity term (for example, "Lettuce, Green Leaf").
• Unit. The quoted unit (pound, carton, hundredweight); where a unit conversion is involved, the conversion is declared, not silent.
• Pricing basis & point. Whether the figure is a wholesale dollar or an index, and the market level it represents (a USDA terminal-market shipping point, an FOB cutout, and so on). A figure on an index basis is never shown as a dollar.
• Geographic basis. A national composite or a named multi-market range; whether the range spans market families or a single source is recorded, because it drives the shippable bar below. Regional splits are a future item, not yet quoted.

A quote is published only once its specification is verified — staged sources never reach a reader.

Sourcing principles, in priority order

1. Public and public-domain. We build only on sources we can cite by report identifier and date. USDA data is republishable, including derived values.
2. Mandatory and transaction-verified first. USDA Livestock Mandatory Reporting (cattle, hogs, beef, pork — transaction-reported under federal law), the mandatory weekly National Dairy Product Sales Report, and AMS terminal-market reports outrank survey or index series.
3. Verification before publication. A source is staged as unverified and only flips to verified after a live probe confirms its report identifier, commodity term, unit, and price field all resolve. Until then it never reaches a reader.
4. Name the ceiling, don't fake it. Where public data is structurally coarser than an assessed benchmark — poultry and shell-egg granularity, daily seafood — we label the limit rather than paper over it.

The assessment pipeline

The path from raw report to published reading is automated and logged: fetch, normalize, bound, reduce, composite, assess, gate. No human keys in a number; the audit trail is the version history of the data files.

• Normalize. Each source family has an adapter that maps a raw payload to a value, a basis, and an as-of date, on a declared basis (wholesale versus index).
• Bound. Every ingredient carries per-source sanity bounds; an out-of-bounds reading is rejected before it can move a level — the guard against a mis-parsed unit or a misplaced decimal. Staying in-bounds is a hard gate, not a warning.
• Reduce. A single report that spans a range of markets or grades collapses to one number via a declared reducer — trim-the-tails-take-the-middle for terminal produce, a volume-weighted average for mandatory cuts. The reducer is part of the spec, so the same report always reduces the same way.
• Composite. Independent sources blend into a level (the dollar) and a trend (the direction) as separate sub-objects.
• Assess. The point is stamped with its level and trend confidence and its provenance — which source families and types backed it.

Collection runs on each source's native cadence — AMS terminal reports daily, mandatory reporting daily or weekly, the dairy report weekly, BLS and FRED monthly — and records each report's own publication date as the value's vintage.

The measured / derived / absent honesty spine

Every ingredient is tagged as exactly one of three states. Making absent a first-class, explained state is more trustworthy than a silent number — and it is the honest answer for the handful of ingredients where no public series exists.

The three-tier spine. An untagged ingredient fails the build.

State	Meaning	How it is shown
measured	A published USDA, BLS, EIA, or dairy-report series, shown as-is.	Solid line, source link, report date.
derived	Our own estimate from public inputs (see below).	Dashed line, confidence band, an "estimate" badge, and a link to the method.
absent	No credible public series exists.	A greyed "no public data — here's why" card naming the structural reason (for example, USDA's small-cell confidentiality suppression).

Confidence — level and trend, scored separately

The headline confidence is the lower of two independently-scored stories. Scoring them apart stops a noisy week-to-week direction from stamping a well-measured price as low-confidence.

• Level confidence rates the dollar. High requires at least two independent dollar types that agree within a tight dispersion band; one type is medium; an index-basis or absent level scores no dollar at all. Two types that disagree are capped at medium — independence without agreement is not certainty.
• Trend confidence rates the direction. High requires at least two independent trend types agreeing on direction most of the time; a jagged, noisy path is capped at low no matter how the endpoints line up — a noisy series is noise dressed as a trend.
• The calibration ceiling. The published confidence may never exceed what the data supports, and the health matrix proves on every build that nothing is overstated.

The shippable bar — ship complete, or not at all

An ingredient earns a public reading only if it clears the shippable bar: a credible wholesale dollar level — either a measured cross-market range spanning several market families, or a single authoritative source whose direction is corroborated by at least two independent types and isn't pathologically noisy. No dollar level, or a lone uncorroborated source, and it does not ship.

There are no apologetic "no published figure" readings on the dashboard. Below-bar ingredients live only as honest expanding-coverage pages — present, named, and explained, but without a fabricated price.

Freshness & series rot

Every value stores its vintage. A freshness heartbeat reports the oldest reading, and the weekly refresh turns a persistent stall red. Public series get revised or retired on their own schedules — when one vanishes, the ingredient becomes absent with the reason, never a silent zero and never a frozen last value pretending to be current.

Separately, the percentage we show must describe the line we draw: every displayed trend is reconciled to its own sparkline before publish, so the headline change and the curve never tell different stories.

The derived methods — a labeled estimate, never a measurement

Derived values let a coarse or indirect source still earn a place. Each one carries a confidence band, a link to the method, and degrades to absent when its inputs don't clear the method's own bar. They are tagged derived, never measured.

The historical-ratio bridge (the general principle)

The methods below are instances of one idea: any reputable series can be put to use through a validated historical ratio to the wholesale number we are building. If an outside series overlaps our measured wholesale series on the ingredients where we hold both, we can learn the historical relationship there and apply it where we have the outside series but not a direct wholesale one — even when the outside series isn't itself a clean wholesale number.

That is also exactly where a fact-disciplined index could quietly start lying, so the bridge is published only under firm guards. A relationship that fails any of them yields absent, not a number:

1. Overlap, not assertion. The ratio is measured on a real overlap window where both series exist — never assumed. No overlap, no bridge.
2. Stability gate. The relationship must be stable over that window. A drifting or noisy ratio yields absent; we don't publish off a relationship we can't pin down.
3. The band is the relationship's error. The published band reflects the residual of the historical fit, so a loose bridge reads as a wide, honest band — never a false-precise point.
4. Mechanism, not coincidence. A bridge needs a plausible causal link (feed to meat, import flow to domestic price, farm to retail spread) — not a bare correlation. "It correlates" is not enough; "it correlates and there is a reason" is.
5. Always derived, always re-estimated. The ratio is re-fit as overlap grows, and a bridge that destabilizes reverts to absent. A bridge never overrides a direct measured level.

The instances of the bridge:

• Yield-adjusted true plate cost. The edible-portion cost is the wholesale price divided by a representative trim yield — converting a wholesale tick into the number an operator actually pays. Yields are ranges, labeled illustrative; the operator's own yield governs, so this is a guide to plate cost, not a guaranteed one.
• Seasonal normalization. A month earns a "typical" median and band only once it has been observed across two or more distinct years; until then the card reads a transparent building-baseline state. No "typical June" off a single June.
• Indirect estimation from the retail-to-wholesale spread. Where no direct wholesale series exists but a reputable retail series and a published spread both do, we estimate wholesale from retail and the spread, validated against our own observed wholesale-to-retail ratio. The honesty constraint is built in: retail is sticky and lags, so a retail-derived wholesale understates volatility and lags turns — it is a labeled cross-check or fallback, never a primary level when a direct source is available.
• Import unit-value estimation. For import-dominated ingredients with no domestic wholesale series, a level can be estimated from customs value divided by import volume, paired with an import price index as a direction corroborator. Same rules: derived only, banded, monthly and lagged, and explicitly not wholesale — customs value excludes duty, freight, and the importer's margin.

The leading-indicator pressure overlay

Forward signals — feed prices leading protein, drought leading beef, diesel leading freight — are published as inferred direction only, never a price, and only once a rule's live track record clears a hold-until-proven bar (a minimum number of calls, a minimum hit rate, and enough non-steady calls to be meaningful). The recomputed direction must match what's rendered. This is the one part of the index that looks forward, and it is bounded hard: a direction, with a track record attached, and no dollar figure.

Revision & correction policy

A price-reporting method lives or dies on how it handles a number that changes. Ours:

• Source revisions. When a source republishes a revised value, the next refresh re-vendors it; the new vintage and the version-control diff are the record. History is not silently overwritten — the prior value is recoverable.
• Corrections. A discovered error — a mis-parsed unit, a wrong commodity term — is fixed at the source specification, re-verified, and re-vendored. A correction that changes a published reading is noted in the change log.
• Withdrawals. A discontinued or suppressed series transitions the ingredient to absent with the reason, rather than being frozen at a stale last value.

Governance & fact discipline

The methodology is versioned with the code. A change to how a number is assessed — a threshold, a reducer, a new derived method — is a reviewable commit with a dated change-log entry, so a reader can see not just the current method but how it evolved.

The fact discipline behind it is absolute. No number, date, or claim is invented: every figure is a published source value, a labeled derived estimate, or labeled illustrative in the prose. The narration of this index is read aloud verbatim in six languages, so a fabrication would be spoken aloud — which is precisely why the gate has no soft edge. Every published number traces to a dated public report and a named reducer, and the entire surface is reproducible by re-running the open checks against the committed data. That reproducibility — not any single quote — is the citable asset.

Stated limitations & biases

Naming the limits is part of the method, not an afterthought:

• The public-data ceiling. Poultry and shell-egg granularity, and daily seafood, are structurally coarser in public data than an assessed benchmark's quotes. We label these, and several ingredients are honestly absent rather than faked.
• Derived-method lag. The retail-to-wholesale back-out understates volatility and lags turning points; it is a labeled fallback, never a primary level.
• Yield ranges. The edible-portion cost uses representative trim yields; the operator's own yield governs, so it is illustrative, not a guaranteed plate cost.
• Seasonal thinness. Seasonal normals need at least two years per month; until the corpus fills, the seasonal reading is an honest building baseline, not a normal.
• No forecast. We publish measured levels, measured directions, and inferred directional pressure — never a price forecast.

Questions about the method

Where do the prices come from?

Only from public U.S. market data: USDA mandatory and terminal-market reports, the National Dairy Product Sales Report, BLS, FRED, EIA, and NOAA Fisheries. No private survey, no phoned-in assessment, no vendor feed. Every published number traces to a dated public report a reader can look up.

How is this different from Urner Barry?

Urner Barry is an assessed, judgment benchmark whose value is daily timeliness; its method is necessarily private. This index does not try to out-quote it on granularity. It out-explains it: every number traces to a citable public report, every estimate is labeled an estimate, and every gap is named with its structural reason.

What do measured, derived, and absent mean?

Measured means a published government series, shown as-is. Derived means our own estimate from public inputs, drawn with a dashed line, a confidence band, and an estimate badge. Absent means no credible public series exists, shown as a greyed card that names the structural reason instead of inventing a number.

Why does an ingredient sometimes show no price at all?

Because it does not clear the shippable bar. An ingredient earns a public reading only when there is a credible wholesale dollar level: either a measured cross-market range or a single authoritative source whose direction is corroborated and not pathologically noisy. A lone, uncorroborated source is held back rather than dressed up as certainty.

Is this a forecast of where prices are going?

No. The index publishes measured levels, measured directions, and an inferred directional pressure overlay that carries a direction but never a price. It never publishes a price forecast.

What happens when a source revises or discontinues a series?

A revision is re-vendored on the next refresh with a new vintage date; the prior value stays recoverable in version control rather than being silently overwritten. A discontinued or suppressed series turns the ingredient absent with the reason, never frozen at a stale last value pretending to be current.

Sourced: public U.S. market data — USDA AMS and Livestock Mandatory Reporting, the National Dairy Product Sales Report, BLS, FRED, EIA, and NOAA Fisheries — surfaced live through Cost Pulse. Method last reviewed 2026-06-14.