₲ True Inflation

How the index is built

What True Inflation measures, how each number is produced, and where the method is weak. Figures marked “live” below are read from the index tables when this page is generated (last complete night: 2026-11-03).

1.What is measured

A chained, night-over-night change in the shelf prices of food products, per store and then per country, in each country's own currency. It is not a price-level comparison between countries and not a cost-of-living index: there is no fixed basket and no household weights. Exchange rates play no part in the index.

2.Collection

Each store's website or app is scraped once a night. Since 3 September 2026 every store's night is recorded with its status and row count, and a night recorded as partial or failed is never used, either as a measurement or as the baseline the next night is compared with. Nights before that date have no such record; they are checked only by the slice rule below and the thin-night rule (section 9).

A scrape can also end cleanly with only a slice of the catalogue (a few categories, or one promotion page). A run that offers fewer than half the store's usual number of products (the median of its last 14 runs, once it has at least 5) is therefore recorded as partial even when the scraper reported success. Earlier slices found in the history by the same rule are excluded from the index in the same way. Their rows are kept.

3.Price sanity at insert

Since 3 September 2026, a product whose price more than doubles, or falls by more than half, against the same store's previous scrape is treated as a probable scraper error and is not stored. The comparison is only made when that previous scrape is at most 7 days old; across longer gaps ordinary price movement can exceed the band, so nothing is rejected. When a large share of a store's catalogue trips the rule at once (a unit or currency change on the site, or the scraper reading the wrong field), the night is recorded as partial and left out of the index. Prices scraped before that date were stored as scraped; the same band is still applied to them when the index reads them (section 4), but a whole night of wrong prices was not caught at the time.

4.Matching products

A price is compared only with the identical product in the same store on the previous usable night: exactly the same name, weight and unit. There is no fuzzy matching and no grouping of similar products. Prices must be positive, and the pair must lie strictly between 0.5× and 2× (the same band again, at index time). Units stay as the store publishes them; US stores keep ounces and pounds.

5.A store's change

For each store and night, the change is the geometric mean (Jevons) of the matched price ratios: exp(mean(ln(pnow/pbefore))) − 1. It is symmetric, so a promotion that cuts a price and later restores it nets to zero rather than drifting upward. A night with fewer than 10 matched products gives no reading. The 7-day and 30-day views compare against the newest usable night at least 7 or 30 days earlier, and give no reading when that baseline is more than 2 days further back. The same floor applies to them: a 7-day or 30-day change resting on fewer than 10 matched products is not shown for that store.

6.Food-only filter

Products are kept or dropped by name, using a multilingual list of 720 patterns that removes alcohol, household goods, personal care, pet food and similar items. It is pattern-based and therefore imperfect: some non-food items still get through, notably in stores whose catalogues mix groceries with stationery, toys, merchandise or medicine. In samples of such stores a noticeable share of the remaining “other” items is still non-food. A few scrapers also skip a store's non-food departments at the source.

7.A country's figure

Each night, the country's change is the average of its stores' changes, weighted by how many products each store matched — not by market share or by what households buy. A large online catalogue therefore counts for more than a small one. The nightly changes are compounded from the country's first measured night (or its restart, section 13), which is set to 0%. Countries therefore start on different dates (see the coverage table), and their cumulative figures cover different periods.

8.Global average

Where a global average is shown, it covers one common window. Its start is the date by which three quarters of the countries had begun (or restarted) their lines. For each country that had begun by then, the change from its level on that date to its latest value is taken, and the global figure is the plain, unweighted mean of those changes: every country counts the same, whatever its size. Countries that began later are left out, and the app says how many. It is a summary, not an index.

9.Thin nights

A night on which fewer than 0.6× the usual number of stores reported (the median of the 14 nights before it) is excluded, both as a measurement and as a baseline, so the night after it is compared with the last good night. The rows are kept; they are simply not used.

10.Gaps and later revisions

When a store misses nights, the change it measures across the gap is spread evenly (geometrically) over the published nights inside the gap instead of landing on a single night. A store that comes back after weeks therefore does not produce an overnight jump. Nothing is spread across a structural break, or back to a country's first night.

A change measured across missed nights moves the country line as much as that many ordinary nights would, so a long gap needs a larger sample than one night does. When the gap spans 2 or more missed index dates, the change is used only when at least 50 products matched across it. Below that minimum the gap contributes nothing: the store simply picks up again from the night it returns. A store that misses fewer index dates than that (typically a single night) needs only the ordinary minimum of 10 products (section 5).

11.Structural breaks

Some nights mark a change in how a store was measured — the scraper switched from promotional to regular prices, or the site served a different branch — rather than a change in prices. On these listed nights the store's change is set to 0%.

• hktvmall — 2026-04-15
• hktvmall — 2026-04-16
• auchan — 2026-09-18
• carrefour-ar — 2026-10-07
• carrefour-ar — 2026-10-08

12.Mirror stores

12 store tables turned out to copy another store's prices (the site ignored the branch or country requested). They are excluded from the index, from every count on this site, and from the global comparison, for their whole history.

13.Dropped countries and restarts

A country is dropped from the current ranking when none of its stores has a measured change within 7 days of the newest index night. Its history is kept, and it returns when a store does.

When a country has no measured change from any store for more than 14 days, and no store's later change is measured from before that hole, its line restarts at 0% on the first night after the hole instead of being spliced across it. What follows such a hole is usually a different basket, often a different retailer; chaining across it would count the empty weeks as 0% and stitch two shops into one series. A hole that a returning store bridges with a measured change is not a restart (section 10). The rows before a restart stay in the database; they are not part of the published line.

A country whose stores matched a median of fewer than 200 products a day over the last 30 index nights is flagged as thin in the app.

14.Category breakdown

The per-category view uses the same matching, food filter, gap spreading and geometric mean, split by a category assigned to each product when it is scraped (by a multilingual name classifier), with at least 3 matched products per category and night. The minimum sample for changes measured across 2 or more missed index dates (section 10, at least 50 products, counted over the store's categorised products) and the restart rule (section 13) apply to this view as well. Products without a category are left out of this view, and categories are not re-weighted to the country line, so they need not add up to it.

15.Global product comparison

The cross-country comparison of staples (eggs, ground beef, butter, milk, bananas) and global brands works differently from the index. It shows one current shelf price per product and country, converted to US dollars, and is chosen like this:

• Fresh, complete scrapes only. Candidates come from a store's latest complete night, only when that night is no more than 7 days old and the store has at least 3 complete nights with products in the last 7 days, so a store that has only just come back cannot set a country's price. Mirror stores never set a price.
• The same pack. A candidate must be a single pack (no multipacks or cases) at the product's exact stated pack size. Eggs, ground beef and Doritos have no single standard pack, so they use a stated retail range instead: 6–30 eggs, 250–1,000 g of beef, 85–450 g of Doritos. A size written in the product's name must agree with the stored size.
• Filters. The food filter applies, and so does the price-sanity rule: a pack whose price has more than doubled or halved since that same product was last seen on a complete night is rejected. Prices confirmed by another complete night in the last 7 days are preferred.
• A plausibility band. A candidate below 0.25× or above 4× the product's cross-country median (in US dollars) is rejected as a probable mismatch or price error.
• The median, not the cheapest. The published price is the median remaining candidate in the country. Where a country has no plain pack, a flavour, sugar-free or organic version stands in; it is marked as a stand-in on the page.

Each country's figure is one local product, so it is not a basket, and coverage differs by product. A country with no candidate left is shown without that product rather than with a wrong one.

Exchange rates. Current prices are converted at the latest rates from open.er-api.com; if that service is unreachable, a fixed table of approximate rates is used instead. Price histories use the daily rates stored for each date, from the same service and, for older dates, the open currency-api. Both are market reference aggregators that track the official rates; they are not any central bank's published fixing. Where a parallel-market rate differs from the official one, the dollar figure overstates what the price is worth locally. These requests are made by the server and carry no visitor data.

16.The real cost: what the government takes

Next to shelf prices, the app shows what food really costs you in each country: the 100 you pay at the till, plus what the government takes from you on top. For example: “Canada: $100 of groceries really costs you $179. You pay $100 at the till. The government takes another $79 from you in taxes and borrowing.”

How 179 is worked out.

1. Take everything the government raises in a year — taxes, social security contributions and its other income — as a share of everything the country produces (GDP).
2. Add what the government borrows that year, also as a share of GDP. Borrowing counts because it is paid back later out of taxes. A government that saves instead lowers the figure.
3. Divide that total by what people spend (household spending), again as a share of GDP. Because both are shares of the same GDP, GDP drops out: this says, for every 100 that people spend, how much the government takes on top.
4. Round it to a whole number and add it to 100. That is the real cost.

In round numbers, for Canada (2024 government figures, 2025 spending): the government raises 42 and borrows 2 for every 100 the country produces, and people spend 56 of that 100. (42 + 2) ÷ 56 × 100 = 79. So 100 of groceries really costs 100 + 79 = 179. The app uses the exact published figures, not these rounded ones.

Where the figures come from.

• What the government raises and borrows: the IMF (Fiscal Monitor / World Economic Outlook), for all of government — national, regional and local government and social security together, 2024. This is the first choice for every country.
• The World Bank's figure for central government, used only where the IMF's figure covers less of government than that one does. The breakdown by type of tax also comes from the World Bank (years vary by country, back to 1980) and is not used for the number itself.
• What people spend: the World Bank; where it publishes nothing or only an old year, the UN Statistics Division, or Taiwan's statistics office (DGBAS).

Figures more than 3 years apart are never put together. For 14 countries the IMF's figure covers only part of government (central government only, or the public sector including state-owned companies, by the IMF's own notes). There the bigger of the IMF and World Bank figures is used, because it leaves less out.

How sure we are. Each country is marked H (high), M (medium) or L (low). H: the figures cover all of government. M: an approximate figure — this country's government figures are incomplete (those 14 countries). L: people spend under 35% of GDP, a sign that GDP itself is inflated — usually by foreign companies' profits booked there; the real cost is not affected, but the country's figures are less certain. Where the published figures cannot be put together, the app shows an older estimate instead and marks it as one we can no longer check. On the ranking, low-confidence countries carry a ~.

What this number is not.

• It is an average. It is the whole country's figure per 100 spent, not your own tax bill. Some people pay far more, some far less.
• It counts everything the government raises and borrows — from companies, oil and gas, state-owned businesses and debt, not only from your pay and your shopping.
• It doesn't count what you get back, like roads, schools, hospitals and pensions. Nothing is subtracted for them.
• Household spending stands in for what people take home after tax. The two are close but not the same: what people take home also includes what they save.
• Shelf prices would not simply drop by that amount without the government. The real cost shows how much you pay in all, on average, for every 100 you spend; it is not a forecast of what groceries would cost without tax.

17.What is never done

Stored prices are never deleted or overwritten to make the index look better. Since 3 September 2026, a price is not stored in the first place when it fails the price-sanity rule at insert (section 3); earlier prices were stored as scraped and are screened by the same band at index time (section 4). A stored value found to be wrong is corrected to the right value rather than removed. Exclusions (partial and slice nights, thin nights, breaks, mirrors, restarts) are applied when reading, not by removing rows, so every such judgement stays reversible.

18.Known limitations

• Uneven coverage. Some countries rest on a single store; one store's promotion cycle then moves the whole country.
• Promotion swings. Stores with weekly promotions produce a weekly saw-tooth of several percent in either direction. The geometric mean cancels a round trip, but a single night can look dramatic.
• Partial scrapes. Slices of less than half the usual catalogue are caught (section 2), but a scrape that silently returned, say, two thirds of a catalogue is still recorded complete and enters the index on that sample.
• Small samples across gaps. A store that misses a single night can still carry a change measured on as few as 10 products over that short gap; longer gaps need at least 50 (section 10).
• Older nights were checked less. The scrape status record (section 2) and the price-sanity rule at insert (section 3) exist only since 3 September 2026. Earlier nights were checked only by the slice rule (section 2) and the thin-night rule (section 9), plus the 0.5×–2× band at index time (section 4), so a scrape that ended early or read wrong prices on a whole night could enter the index then.
• Revisions. Past values change when gaps are filled (section 10).
• Residual non-food. The name filter misses some items (section 6).
• No shrinkflation. Pack-size changes are not counted (section 4).
• Weights. Stores are weighted by matched products, not by spending (section 7).

Faults that are found are corrected; the index is then recomputed for the affected period.

19.Coverage by country

(live) 72 countries measured in the 7 nights to 2026-11-03. Stores = stores with a measured night in that week; products = store-product price comparisons on 2026-11-03, summed over the country's stores (a chain scraped in several cities counts its catalogue once per city table). “Measured since” is the country's first measured night; where a restart applies (section 13), the published line begins later.