Methodology and How to Use

• gdp-jp is the Cabinet Office's 四半期別ＧＤＰ速報: real (chained 2020 prices), nominal and the deflator, all seasonally adjusted at annual rates, quarterly from 1994 Q1, read from the three e-Stat tables the Cabinet Office overwrites at every release. Levels are stored in the ¥ billion published; pages show ¥ trillion (÷1,000, exact). A quarter is dated by its first month: 2026-04-01 is April–June 2026.
• Only levels are stored. The Cabinet Office also publishes growth rates; they are not ingested. Quarter on quarter: qoq % = (level[t] ÷ level[t − 1 quarter] − 1) × 100. Annualized: annualized % = ((level[t] ÷ level[t − 1 quarter]) ^ 4 − 1) × 100. Year on year: yoy % = (level[t] ÷ level[t − 4 quarters] − 1) × 100. The Cabinet Office computes its own rates before rounding the levels, so a rate here can differ from the published one by about a tenth of a point. That is disclosed, not adjusted.
• Contributions to growth: contribution pp = (component[t] − component[t − 1 quarter]) ÷ GDP[t − 1 quarter] × 100 (and the same at four quarters for the year-on-year view). In chained-price accounts the real components do not add to real GDP; the Cabinet Office publishes the gap (開差) as a line of its own. The page bundles that gap with rounding into one residual bar, residual pp = GDP growth % − Σ contributions pp, so the bars always add to the line and nothing is hidden.
• Three gates tie the three tables to one release: the nominal components must sum to nominal GDP to the tenth of a billion; the real components plus the published discrepancy must sum to real GDP; and nominal ÷ real × 100 must reproduce the published deflator to within a quarter of a point. A release where any of these fails is not published.
• Vintages: the real-time record, 2002 to today. The same table ids are rewritten at every release, so each revision is captured as a new vintage from our first ingest onward. Behind that sit 134 backfilled releases — every first and second preliminary estimate the Cabinet Office published between August 2002 and March 2019, loaded from its archived tables with the date it published each one. ?as_of= on the API re-draws any chart as it stood on a past date, and /releases marks every vintage with whether its date is the agency's own (known_at_basis: agency publication) or our fetch.
• Where those publication dates come from. e-Stat carries none — every archived GDP table reports the day the whole archive was uploaded. Dating them by the usual rule of thumb would have put releases on the wrong side of a month end and corrupted every as-of answer in the span. Two of the Cabinet Office's own records are used instead: its machine-readable release calendar, which covers 2009 Q3 onward, and the statistics-list page each release publishes, which reaches back to 2002. On the 24 releases where both were compared they agreed 24 times; that is what justifies using the page for the earlier ones. Where both exist they must still agree or the release is refused. The stamp is 08:50 Japan time, the hour GDP is released.
• Levels are not comparable across vintages, and that is the point. Each release is on the price base and the SNA vintage in force when it was published — real GDP appears in 1995, 2000, 2005, 2011 and 2020 chained prices at different dates, and nominal GDP steps up by about ¥30tn at the December 2016 move to the 2008 SNA. So a difference between two vintages can be a rebasing, a benchmark revision or a genuine data revision, and the three are not the same thing. Each archived release records its own base. Two renamed lines are carried across the break as one series each (民間在庫品増加 → 民間在庫変動 and its public twin); the four "excluding FISIM" lines published only during the changeover are kept as series of their own rather than folded into the main ones.
• One gap, stated rather than papered over. e-Stat's archive of past estimates stops at the 2018 Q4 second preliminary (published March 2019) and our own ingests begin in September 2026. Nothing is recorded between those dates, so an as_of anywhere in that window returns the March 2019 vintage — the last thing we can honestly say was published on or before it.

• unemployment-jp is the Statistics Bureau's 労働力調査 (Labour Force Survey), long-run table 1 a-1: labour force, employed, employees, unemployed, not in the labour force and the unemployment rate, for both sexes, men and women. Seasonally adjusted from January 1953, original series from July 1972. Headcounts are in ten thousand persons (万人) as published — 177 is 1.77 million people. Figures before 1974 exclude Okinawa; March–August 2011 are the Bureau's supplementary estimates after the earthquake.
• The Bureau replaced October 2005 – December 2021 with re-benchmarked figures after each census. Each cell was re-estimated on its own, so in a dozen months of that window men plus women, or employed plus unemployed, miss the published total by up to 50,000. The figures are stored as published, not forced to add up.
• job-openings-jp is the Ministry of Health, Labour and Welfare's 一般職業紹介状況, long-run tables 2 and 3: new and active job openings per applicant at Hello Work, the public employment offices, including part-time (the headline), excluding part-time and part-time only, original and seasonally adjusted, monthly from 1963. A ratio of 1.18 is 118 openings per 100 applicants. Jobs and job seekers that never pass through Hello Work are not counted. The ratio's change is read in points; no percentage change is offered.
• wages-jp is the Ministry's 毎月勤労統計調査 (Monthly Labour Survey), from its long-run cumulative files: pay in yen (total, contractual, scheduled, overtime and special), hours, days, employment, hire and separation rates, and the Ministry's indices (2020 = 100), for all industries and manufacturing, establishments with 5+ employees (from 1990) and 30+ (from 1970), all, full-time and part-time workers. Final figures only; the preliminary release two weeks earlier is not in these files.
• Wage growth is the Ministry's published rate. The series ending _yoy are the Ministry's own year-on-year growth rates, stored as published, and the page leads with them. They cannot always be recomputed from the indices: the index level steps up between December 2023 and January 2024, and in every month of 2024 the change computed from the index overstates the published rate by 2 to 4 pp (June 2024 total cash earnings: +8.0% from the index, +4.5% published). Outside that break the two agree to within 0.3 pp in 95% of months since 2012. The yen averages jump whenever the Ministry rotates its sample, so no percentage change is offered on the yen, hours or days series.
• The real wage index is the Ministry's own: the nominal index deflated by the CPI excluding imputed rent (持家の帰属家賃を除く総合). Nothing is deflated by this platform.
• Wages are not seasonally adjusted. June, July and December carry bonuses; the page's tiles compare each month with the same month a year earlier.

• corporate-finance-jp is the Ministry of Finance's 法人企業統計調査 (quarterly), the survey of companies with capital of ¥10 million or more, read from e-Stat's time-series table for industries other than finance and insurance. A pinned subset is carried: 21 lines (sales, operating and ordinary profit, personnel costs, directors' remuneration, headcount, capital investment with and without software, depreciation, cash, inventories, equity holdings, the balance-sheet totals, borrowings, bonds and retained earnings) across 31 industry aggregates and four capital classes, quarterly from 1954. Amounts are in ¥ million as published; counts are counts.
• Not seasonally adjusted. The right reading is year on year: (value[t] / value[t−12 months] − 1) × 100. A quarter-on-quarter change on these levels is mostly seasonality and no surface shows one.
• Margin: margin % = ordinary profit ÷ sales × 100. Trailing four quarters: ttm = sum of the four most recent published quarters, flows only. The Ministry's own published ratios are not ingested.
• Three identities are gated on every ingest, for every industry, size and quarter: total assets equal liabilities plus net assets; manufacturing plus non-manufacturing equal all industries; and the three capital classes sum to all sizes. Yen lines reconcile to the ¥ million; the company count is an estimate rounded within each class, so it is allowed five parts in ten thousand.
• The "up to fiscal 2008" legacy industry lines created by the 2008 reclassification are not carried; finance and insurance are a separate table with a different line structure and are not included.

• Eight datasets, one page. fiscal-jp is the central government's general account from 1875; fiscal-jp-revenue its revenue heads from 1982; fiscal-jp-expenditure its spending by the Ministry's major-expense classification from 1967; fiscal-jp-ministry the same account by the body that spends it, from 1877; tax-receipts-jp every national tax against budget from 2022; govt-debt-jp the quarterly debt stock; govt-accounts-jp the whole of government on the IMF GFS basis from 1994; and local-finance-jp the 47 prefectures from 2002.
• Budget and settlement are different measures and never share a series. 予算額 is what the Diet voted — the initial budget plus every supplementary passed during the year — and 決算額 is what was finally collected and spent. Japan under-spends its budget in most years. The settlement runs about two fiscal years behind the budget because a year's accounts close roughly twenty months after it starts, so a settlement series ending earlier than a budget series is the calendar, not a gap in the data.
• Bond issuance is revenue. 公債金 is a revenue head of Japan's general account, so the published revenue total includes borrowing. Tax and borrowing are separate series and are never merged.
• A fiscal year is dated to the 1 April it begins on. Before 明治19年度 (1886) the Japanese fiscal year began in October and then in July, so that dating is a convention for those years rather than a fact. The eight pre-1875 accounting periods (明治第1期 to 第8期) run from four to fifteen months and are not fiscal years; they are in the Ministry's file and deliberately not in the dataset.
• Units. The Ministry's fiscal tables are stored in ¥ million; Table 1 is published in 円 to 1946 and 千円 from 1947 and is normalised on ingestion so the series does not change unit mid-run, with the multiplier read from each sheet's own 単位 cell. The debt statement is stored in ¥100mn, the national accounts in ¥ billion and the prefecture accounts in ¥ thousand, each as published. Conversions to ¥tn or ¥bn on a page are exact divisions and are stated wherever they appear.
• Two revenue heads exist only in a footnote, and are ingested from it. The Ministry's 合計 includes bonds with no column of their own — the 1991 Gulf War 臨時特別公債, the 減税特例公債, the 2011 復興債 and the 年金特例公債 — and Table 4 also carries receipts from the 決算調整資金 in five years. They are parsed out of the note text and published as their own series, which is why the printed heads reconcile to the printed total in every year.
• Where the source does not add up, the figures are published as printed. Three places in about 460 year-and-measure pairs do not reconcile in the Ministry's own files: the fiscal 1973 予算現額 for 国債費 (an ¥11bn typo its own 差引額 column contradicts), the fiscal 1977 budget and fiscal 1978 settlement by ministry (¥8mn and ¥91mn on ¥29tn and ¥34tn), and the fiscal 2026 revenue budget (¥732.5bn, unexplained). Nothing is silently corrected; each is pinned so every other year stays checked to the yen.
• Ministries are published under the name they had. The 2001 reorganisation split and merged ministries rather than renaming them, so 大蔵省 and 財務省 are two series, as are 厚生省, 労働省 and 厚生労働省. A body that appears in one year and not the next has ceased to exist, not moved.
• The general account is not "the government". It is one legal account of the central government. govt-accounts-jp consolidates central government, local government and the social security funds and nets out the transfers between them; the consolidation (部門間調整) is published as its own series so that central + local + social security + consolidation = general government is visible and checked on every line of every year. Net lending and the operating balances are the Cabinet Office's own balancing items, taken from the source rather than derived by subtracting one series from another.
• Rolling windows are stitched, not truncated. The debt statement publishes only the last five years in one file, one quarter arriving and the oldest leaving. Every edition is archived and read together, newest edition winning where two cover the same quarter, so the history grows rather than shrinking.
• Not ingested. Table 19(1) (fiscal 1949–1984) uses the older 重要経費 vocabulary and needs its own mapping; the line items beneath each major-expense heading are re-cut by the Ministry every few years and only the headings are comparable across the run; the GFS revaluation, other-volume-change and closing balance-sheet blocks are stocks rather than operations; municipalities have their own MIC survey and are not in the prefecture dataset; and tax receipts before fiscal 2022 sit in the National Diet Library's WARP archive, which refuses our requests.

• Two datasets sit behind this surface and they are not interchangeable. population-jp is the Basic Resident Register (住民基本台帳) as published by the Ministry of Internal Affairs and Communications: an administrative count of everyone on a municipality's register at 1 January, not a survey estimate, split three ways — all residents, Japanese residents and foreign residents. population-jp-history is the Statistics Bureau's System of Social and Demographic Statistics (社会・人口統計体系), which republishes census, population-estimate, vital-statistics and register figures on one prefecture grid back to 1975.
• Reference dates differ by series and are never mixed on one axis. Register counts are as of 1 January. Register flows — births, deaths, moves — cover the calendar year before that date, so the count at 1 January 2026 less the count at 1 January 2025 equals the flows dated 2025. Census and population-estimate figures are as of 1 October and are a different measure: the total population line and the registered residents line are close but neither is read off the other.
• Change rate: net change ÷ (population − net change) × 100. The register publishes the closing count and the change over the year, but not the population it started from, so the denominator is the published count less the published change.
• Natural and social change: (births − deaths) ÷ opening population × 100 and (net change − natural change) ÷ opening population × 100. Social change is not migration. The ministry defines it as net change less natural change, so it also carries naturalisations, losses of nationality and register corrections. In-migrants less out-migrants is a different figure and does not equal it for any prefecture.
• Foreign share: foreign residents ÷ all residents × 100, always over all residents whichever segment the table is showing — a share of one nationality within itself has no meaning.
• Aged 65 and over: (sum of the published five-year bands from 65 upward) ÷ all residents × 100. A small number of residents have no usable birth date on the register, so the bands fall slightly short of the published total — 43 people in 123.8 million nationally at 1 January 2026, but 1,258 of Hokkaido's 75,392 foreign residents. The shortfall stays in the published total rather than being spread across the bands, so a share computed from the bands is a share of everyone, not of everyone with a recorded age.
• Crude birth and death rates, on the register: births (or deaths) over the calendar year ÷ mid-year population × 1,000, where mid-year population = (population at 1 January + population a year earlier) ÷ 2. A crude rate is per thousand of the population that lived through the year, which is neither endpoint; the register gives both, so the mean of the two is used rather than the closing count.
• Crude birth and death rates, on the long run: births (or deaths) in year Y ÷ Japanese population at 1 October Y × 1,000. Different inputs, so a slightly different number from the register figure above, and the two are never drawn on one axis. The vital statistics count Japanese nationals, so the denominator is the Japanese population and not the total; the 1 October estimate is the mid-year population the vital statistics themselves use. The flow and the population are joined on the year, not on the reference date. This reproduces the Ministry of Health, Labour and Welfare's published rates — 6.0 births and 13.0 deaths per 1,000 for 2023.
• Population pyramid: the bars are published counts. share % = residents in the band and sex ÷ residents with a recorded age and sex × 100 — the denominator is the sum of the bands, not the published headcount, so the bars add to 100%. The gap between the two is stated under every pyramid. aged 65+ per 100 aged 15–64 = residents 65 and over ÷ residents 15 to 64 × 100; males per 100 females = male residents ÷ female residents × 100. Japan's profile is drawn behind a prefecture's as a dashed outline on shares only: two areas cannot share a count axis.
• The total fertility rate is the one published rate on this platform that is stored rather than calculated. It is the sum of birth rates at each single year of a mother's age from 15 to 49; its denominator, the female population by single year of age, is in no table held here, so it cannot be rebuilt from the counts on the page. It is therefore served exactly as the Statistics Bureau publishes it and carries the Official Statistic badge on that basis. Being a rate, it never sums across prefectures, and the ingest's reconciliation gate skips it. Prefecture figures are published for 1980, 1985 and then every year from 1986; the missing years are drawn as gaps. Vital statistics run about two years behind the register, so this series ends earlier than everything else on the page. The 2.07 reference line is roughly the rate at which a generation replaces itself — a reference level, not a published series.
• Two published figures that will not reconcile, disclosed rather than adjusted. The national age bands hold about 6,200 more foreign residents than the 47 prefectures band, because a prefecture leaves those people unbanded and the national row does not; summing the prefectures therefore gives a slightly smaller age total. And the Statistics Bureau's own ageing rate excludes people of unrecorded age from its denominator, so in census years it differs from 65+ ÷ total population by real amounts — 28.7% against 28.0% in 2020. Every rate the platform can rebuild from published counts it calculates itself and shows the formula for; the total fertility rate is the single exception, above.
• Index view: value[t] ÷ that series' value in the first year of the window shown × 100. The base is the first year in view, not a fixed year, so changing the range rebases the chart — the base is printed on the axis and in the export header. Foreign residents enter the register only in 2013, so on a window that starts earlier that line is based on 2013 and the others are not; the lines are then comparable in shape, not in level.
• Log axis: an option on the level views, not a default. Registered residents and foreign residents are two orders of magnitude apart, so on a linear axis the smaller line sits on the baseline with no readable shape. On a logarithmic axis equal vertical distances are equal proportional changes, not equal numbers of people; the axis is labelled "log scale" wherever it is on.
• Municipality level mixes four kinds of area in one column, and three of them overlap. Under each prefecture the workbook lists the prefecture, then designated cities and their wards, districts and the towns in them, and — in Tokyo — one 島しょ row covering the islands as well as each island town. Adding every municipality row together gives 160 million people in a country of 124 million. The 328 rows that are sums of other rows are identified by two independent signals (a grouping row's name is a strict prefix of the rows it groups; and the workbook is an outline in which a grouping row equals the run that follows it) and the result is proved by reconciliation — the 1,898 municipalities add to their prefecture exactly, in all three resident segments. Grouping rows are published here too, marked level: group, and are never included in a total.
• Small cells are suppressed. At municipality level the ministry withholds counts small enough to identify someone — 17,829 cells in the 2026 foreign-resident age workbook. A withheld cell is missing, not zero: a village with three foreign residents has three, and its age bands are simply not published. One Hokkaido village has every foreign age band withheld, so 100% of its foreign residents fall outside the bands, which is correct rather than an error. Across all municipalities the unbanded share is 0.155% of foreign residents and 0.005% of all residents.
• Six municipalities are carried on the register with a population of exactly zero — the villages of the disputed Northern Territories. Any per-capita figure has to guard against them; none is computed here.
• The ministry keeps only the current year's workbooks online — last year's file is deleted and its URL reused — so the archive of past releases here is the history.

• Why the House of Councillors and not the House of Representatives. The upper house's 45 prefectural constituencies (選挙区) are exactly prefectures — every prefecture on its own, except two merged pairs — so a prefecture headcount divides into a seat count with nothing left over and nothing counted twice. The lower house's 289 single-member districts cut through cities below the municipality level, and no boundary file on this platform can resolve them; that surface is not offered rather than approximated.
• The seat complement is law, not a statistic. 公職選挙法 別表第三 (Public Offices Election Act, Appended Table 3) as amended in 2018 and in force since the 2019 ordinary election: 148 district seats across 45 constituencies, plus 100 elected from a single nationwide proportional list, which no prefecture's population bears on and which is excluded here. The table carries no vintage and no freshness — it changes when the Diet changes it. It is held as a checked constant, cited to the Ministry of Internal Affairs and Communications on the surface itself, and self-checked on every request: 45 constituencies, 148 seats, every prefecture in exactly one constituency, and every complement even. Half of each constituency's seats are contested every three years (改選数 = 定数 ÷ 2, whole for every one), so the per-seat ratios are the same whichever complement is divided.
• Tottori with Shimane, and Tokushima with Kochi have been single two-seat constituencies since 2016 (合区). Their two prefectures are summed and they are never shown as four.
• Four denominators, because who counts as the electorate is the argument. adults18 is Japanese residents aged 18 and over, the franchise as it has stood since 2016; adults20 is Japanese residents aged 20 and over; japanese is Japanese registered residents of all ages; residents is everyone on the register, foreign nationals and children included. The last is not an electorate: it is the base most often quoted in the apportionment argument, and the gap between it and the first is itself a finding, since a prefecture with many foreign residents is allocated representation for people who cannot vote in national elections.
• One quantity is estimated, and only one. The register publishes five-year age bands, so the two single years 18 and 19 are taken as two-fifths of the 15–19 band: adults18 = Σ bands from 20–24 upwards + 0.4 × the 15–19 band. Every other year is a published count added as published. The surface serves the headline on all four denominators side by side so the influence of that one estimate is visible rather than asserted; on the January 2026 register it moves the widest disparity by 0.014×, against the 20-and-over basis that splits nothing.
• The register is not the electoral roll. Residents on the Basic Resident Register are close to, but not identical with, the people on the 選挙人名簿 — the roll applies a three-month residency requirement, carries overseas voters separately, and excludes people disqualified from voting. This is an electorate measured from the register, not a count of registered voters, and nothing here should be quoted as the latter.
• The formulas. per_seat = electorate ÷ seats; vote_weight = min(per_seat over the 45) ÷ per_seat, so the best-served constituency is 1.00 and every other is the fraction of one of its votes that a vote cast there is worth; disparity = 1 ÷ vote_weight, the figure the courts and the press quote; the headline 最大格差 is max(per_seat) ÷ min(per_seat). seat_share = seats ÷ 148 × 100, electorate_share = electorate ÷ total × 100, and over_representation = seat_share − electorate_share in percentage points, which sums to zero across the chamber by construction.
• The drift series is a counterfactual and is labelled as one. It applies the seat table in force today to every year of the fifty-year prefecture panel, using registered Japanese residents of all ages (indicator A2101, the only register count published on one basis for fifty years). It measures how far the population has moved under a map that does not move with it — not what the disparity was at the time, which was governed by whatever table was then in force. A year missing any prefecture is left out rather than compared against a different set of areas.
• A constituency with no count is null, never zero. If any prefecture in a constituency has no published figure the constituency's electorate is missing, and it is excluded from the totals, the extremes and the disparity rather than imputed. A constituency whose electorate is a genuine zero keeps that zero and takes no ratio: it can neither anchor the best-served end nor produce a division by zero.

• Monthly counts of foreign visitors, in persons, by market, shown exactly as published. JNTO computes them from Ministry of Justice immigration statistics: foreign nationals entering Japan, less those whose main residence is Japan, plus landings at a port of call, landings in transit and cruise passengers. Crew are excluded; expatriates, their families and students are included.
• Three releases a month, all official. An estimate (推計値) about three weeks after the month, a provisional figure (暫定値) with full market detail two months later, and a definitive one (確定値). The estimate covers only the largest markets and is rounded to the nearest 100, so on estimate months the named markets do not add to the headline and the regional totals are absent; those months are marked on every surface. Measured across three releases of the source workbook, definitive figures are never subsequently revised and the estimate-to-final change is rounding, about 0.006% at the headline.
• Growth: (arrivals[t] / arrivals[t−12 months] − 1) × 100. Month-over-month rates are not shown: the series is strongly seasonal and an unadjusted monthly change is not interpretable. Recovery index: (arrivals[t] / arrivals[same month of 2019]) × 100 — 2019 is the last full year before the border closed, and comparing like months removes the seasonality.
• Share: arrivals[market] / arrivals[Total] × 100 for the same month. The market table therefore shows the latest month with a complete breakdown, not each market's own latest month — a share is only comparable when every row is the same month.
• Contributions: (arrivals[market, t] − arrivals[market, t−12]) / arrivals[Total, t−12] × 100, in percentage points. The residual is growth[Total, t] − Σ contribution[named markets, t], taken from the published headline rather than by summing the markets left out, so the segments sum to the headline exactly even on estimate months. It is always shown, never dropped.
• Year to date sums the published months of the current year against the same months of the prior year. Where a month is still an estimate, JNTO rounds both it and its own cumulative to the nearest 100, so this sum can differ from the press release by a few tens of people. Neither figure is adjusted to match the other.
• Market mix: Σ arrivals[market, t−11 … t] / Σ arrivals[Total, t−11 … t] × 100. Shares are taken over a trailing twelve months so a market's summer peak reads as seasonality rather than a change in the mix. A share is a percent of arrivals, not of spending — markets differ several-fold in what each visitor spends.
• Concentration: HHI[t] = Σ share[part, t]² over the parts that actually partition the total — every leaf market, every residual JNTO publishes, and whatever is left of the headline not itemised in that era, since the residual rows only begin in 2016. Including that remainder is what lets the index mean the same thing in 2003 as today. A higher value means arrivals come from fewer places. The series ends before the headline does: an estimate month does not publish every market, and an incomplete partition is not an index, so those months are gaps.
• The seasonality view groups the same published counts by calendar month, one line per year. 2020–2022 are left out of it: with the border shut those lines sit on the axis and flatten every other year into an unreadable band. They remain in the table, the other charts and every export.
• Where arrivals are shown against the consumer price index for hotel charges, both are indexed to the same month of 2019 so they share one comparable scale; neither is read off the other's units, and the two are published by different agencies on different schedules, so they can end in different months. That chart shows two series moving together. It is not evidence that one causes the other.

• Japan Tourism Agency Accommodation Survey (宿泊旅行統計調査), shown exactly as published. Two measures live here and are never mixed: guest nights are person-nights — one visitor staying three nights is three person-nights and one person, which is why they are never compared with or added to visitor arrivals — and room occupancy is a published percentage. Pages show guest nights in millions (person-nights ÷ 1,000,000); no other transformation is applied.
• These are sample estimates, and the survey says so. Figures are grossed up from a sample of properties, and the Agency publishes its own standard error rate for each prefecture and breakdown — about 1% at the national headline, near 30% for the smallest properties in a thin prefecture. It also marks individual weak cells. A marked value is kept and counted rather than dropped, and the count of marked values is recorded on every release. A room occupancy rate above 100% is rare but real for the same reason — Tottori's city hotels print 104.3% for August 2011 — and is shown as published, never capped.
• The January 2026 break is a change of definition, not of demand. From that survey the Agency stratifies properties by room count instead of employee count, and states that year-on-year comparisons spanning the change may carry its effect. Every surface marks the break with a rule on the chart and a notice whenever a comparison straddles it. Three consequences are structural: property-size series begin at the break, because the pre-2026 employee-count bands measure something else and are deliberately not spliced on; the nationality list widens from 21 categories to 24, so the Nordic region, the Middle East and Mexico begin there; and the residual "Other" is carried as two separate series, because before 2026 it also contained those three. Nothing is adjusted to bridge the break.
• Three releases a month, all official — a first preliminary covering national totals only, a fuller second preliminary about a month later, and annual definitive figures the following year. Unlike visitor arrivals, this survey genuinely revises, so its point-in-time vintages carry real information. A release is assembled from the Agency's trend workbook, its annual definitive workbooks and the monthly second-preliminary files; where two carry the same figure the more settled one is used, and the agreement rate between them is recorded as a check that no column has been misread.
• Growth: (guest nights[t] / guest nights[t−12 months] − 1) × 100. Occupancy is a ratio, so it is differenced, never divided: occupancy[t] − occupancy[t−12 months], in percentage points. An occupancy rate is never averaged across areas — the national figure is the Agency's own, not a mean of the prefectures. Recovery index: guest nights[t] / guest nights[same month of 2019] × 100.
• Shares: foreign guest nights[t] / guest nights[t] × 100, and guest nights[area] / guest nights[All Japan] × 100. The 47 prefectures partition the country; the ten transport-bureau regions republish the same nights and are never added to the prefectures or given a share.
• Contributions: (guest nights[prefecture, t] − guest nights[prefecture, t−12]) / guest nights[All Japan, t−12] × 100, in percentage points. The residual is taken from the published headline rather than by summing the prefectures left out, so the segments sum to the national rate exactly. It is always shown, never dropped.
• Nationality shares use the nationality table's own published total as the denominator. That table covers larger properties only — 20 rooms or more from 2026, 10 or more employees before — so its categories do not sum to the all-properties foreign guest nights and must not be divided by it. Guests whose nationality the survey does not identify are shown as a disclosed residual, normally a few percent of a prefecture's foreign nights and far more in a thin cell, and are never spread across the named categories.
• The purpose split classifies the property, not the guest: "leisure-majority" means at least half of that property's guests stayed for leisure. It is a proxy for the business/leisure mix, not a measurement of it. Neither the purpose split nor the residence split sums to total guest nights — the survey also publishes nights whose purpose or residence is unknown — so they are drawn as separate lines, never stacked into a false whole. The five property-size bands do sum to the total exactly.
• Municipality figures are counted, not grossed up: the Agency publishes them as responses received rather than as an estimate for the whole municipality, so they are not comparable with the prefecture totals. A municipality appears only in months where at least ten of its properties responded, so the list changes month to month; a town absent from a month is missing, never zero. City names are carried in Japanese as published, with the prefecture romanised — the Agency publishes no romanisation and one is not invented here.

• Ministry of Finance customs figures for Japan's trade by partner country, from January 2001 — from the principal-commodity by country tables (概況品別国別表) of the Trade Statistics of Japan, shown exactly as published. Six slices of the one table, each its own dataset and page: semiconductors, components and chipmaking equipment (trade-semis); motor vehicles, parts and motorcycles (trade-autos); mineral fuels — crude, LNG, coal, refined products and LPG by supplier, refined products by destination (trade-energy); general machinery (trade-machinery); medical products (trade-pharma); and food (trade-food). Every rule in this section applies to all six. Value is published in thousands of yen; pages show billions (value ÷ 1,000,000), and no other transformation is applied. Quantity is published in the commodity's own unit — pieces for a chip, kilograms for a machine — which is why unit values are never compared between commodities.
• Export and import commodity codes are separate vocabularies and do not correspond. Code 70311000 is semiconductors and electronic components on the import side and audio equipment on the export side; 70503000 is motor vehicles on the export side and vehicle parts on the import side; 30301000 is crude oil on the import side and refined petroleum products on the export side. Every surface validates a commodity against the direction being read, and the two directions are never joined on a code.
• A published group is never the sum of the items shown beneath it. The Ministry publishes components inside a group that this dataset does not carry, so Semiconductors & electronic components always exceeds integrated circuits plus discrete semiconductors plus thermionic tubes. Groups and items are drawn together for comparison and must not be added.
• Months the table has not reached are published as zero, and are dropped. The current-year table returns 0 for every month after the last one compiled. Coverage is derived per year as the last month carrying any non-zero value anywhere in the slice, and later months are not stored. Zeros inside the covered window are real — no customs entry was recorded with that partner that month — and are kept.
• World totals: Σ value[partner, commodity, t] over every partner the Ministry publishes for that commodity, including the non-country entries (goods shipped to order, unknown destinations, bonded areas). This table publishes no world row, so the total is summed here and carries no official badge.
• Twelve-month total: Σ value[t−11 … t]. Trade is lumpy and a single large shipment moves a partner's monthly figure several-fold, so shares, unit values and the destination mix are read off this sum rather than one month. The sum is left blank until twelve months of history exist.
• Growth: (value[t] / value[t−12 months] − 1) × 100. Month-over-month rates are not shown: the series is seasonal and lumpy. Share: 12-month total[partner, t] / 12-month total[world, t] × 100, with the destination-mix residual 100 − Σ share[named partners, t], so the bands sum to 100% in every month and no partner is dropped.
• Unit value: (12-month total of value[t] × 1,000) / 12-month total of quantity[t], in yen per unit shipped. This is an average realised price across a year of shipments — not a price index, not quality-adjusted, and not comparable between commodities or against the consumer price index. The quantity unit differs by line — pieces for a chip, kilograms for a machine or for vehicle parts, number of vehicles, kilolitres for crude and refined products, tonnes for coal, LNG and LPG — and the mineral-fuel total is published with no quantity at all, so it has no unit value.
• A figure too small to show at the displayed scale reads as <0.1 rather than rounding to 0.0, so a rounded-down value can never be mistaken for a genuine zero or a month with no entry.
• Four release stages, all official — preliminary (速報), confirmed (確報), revised (確々報) and final (確定). The current year moves through all of them, so a recent month will change; a new release is stored whenever the Ministry republishes any year of the table, and the as-of view returns the figures as they stood on any past date.
• Silicon wafers come from the HS-detail tables (品別国別表), dataset trade-inputs, as HS 3818.00 — doped chemical elements in the form of wafers. e-Stat publishes no names for HS lines, so the names are set from the Customs tariff schedule; the export and import schedules split the heading differently (-100/-900 and -010/-020), and the import silicon line was confirmed by its value-per-kilogram signature. The residual "other doped elements" line is kept separate and never added to silicon. On the supply-chain chart the three steps — wafers, integrated circuits, equipment — are indexed to a common base month so lines of very different size can be read together; they are three different products and are never summed.

• Rice contract prices are the volume-weighted average of contracts actually struck between shippers handling 5,000 tonnes a year or more and wholesalers — not a quote and not an auction clear. Yen per 60kg bag of brown rice, first-grade table rice, including freight, packaging and consumption tax, published by MAFF for each 産地品種銘柄 origin-and-variety brand. A brand-month with no contract, or with under 100 tonnes contracted, is not published; the Ministry does not distinguish the two and neither does this site. Both stay gaps.
• The all-brand average is the Ministry's, not ours. It is weighted by the previous crop year's inspected quantity by brand, stored exactly as released, and never recomputed from the brand series shown alongside it — which is why it need not sit at the arithmetic mean of the lines on the chart.
• A regional Koshihikari is one series across a layout change. Until the 2017 crop year MAFF carried the growing area in a separate 地域区分 column; from 2019 it is inside the brand name as コシヒカリ（魚沼）. Both are stored as the same series, so the history does not break in half at 2018.
• The crop-year average price is calculated: mean of the published monthly contract prices for September of year y through August of year y+1 — the marketing year of the crop harvested in year y, which is the year the cost survey prices. A crop year missing any of its twelve months is left out rather than averaged short.
• Three published cost levels, and the difference between them is the point. 費用合計 is what was spent; 生産費（副産物価額差引） nets off straw and other by-products; 全算入生産費 adds imputed interest on the farmer's own capital and imputed rent on the farmer's own land. All three are stored as published. A farm-gate price can sit above the first and below the third, and which one is quoted is usually where an argument about farm policy is being had.
• Rice inventories: the crop-age split does not add to the total, and is not meant to. MAFF itemises the current crop and one-year-old rice; rice two years old and older is inside the stage total with no row of its own. The residual is the difference and is never shown as zero. Shippers and wholesalers do sum to the combined stage, bar the Ministry's independent rounding of each row to whole 10,000 tonnes. Stocks are private trade stocks only — rice on the farm and the government reserve are outside the table.
• The farmer's terms of trade is calculated: (farm output price index[t] ÷ farm input price index[t]) × 100, from published index values on the same base year. Above 100 means output prices have risen further from the base than input prices have. It is not an official statistic and carries no badge.
• Co-operative amounts are thousands of yen as published; pages show billions (value ÷ 1,000,000), an exact conversion. 営農指導事業分配賦額 — the farm-guidance loss charged back to each earning business — is a published line of the Ministry's own statement, not an allocation made here. 共通管理費等 is overhead belonging to no single business: it is part of the group total, so adding the five businesses without it does not reconcile.
• Fiscal and crop years are dated to a fixed month. A co-operative fiscal year runs April to March and is dated to the 1 April it begins (令和5事業年度 → 2023-04-01); a rice crop year in the cost survey is dated to 1 January of its harvest year (令和5年産 → 2023-01-01). Both are conventions of this site, chosen so a year-on-year measure compares like with like.

• Holdings (end of month) and flows (during month) are official values exactly as published in the Bank's “Monetary Base and the BOJ's Transactions” table (MD09), in units of ¥100 million. Pages display ¥ trillion, an exact division by 10,000 — no rounding happens before display.
• Change from peak: value[latest] − max(value[m]) over all published months; the percent version divides by that maximum. The peak month is named wherever the figure appears. 12-month pace: mean(value[t−11] … value[t]) over the trailing twelve published flows. A negative net flow is balance-sheet runoff.
• Flows can be genuinely negative (redemptions carry a minus sign), so no percentage rates are computed on a flow series — a percent change across a sign change is arithmetic noise, and the API refuses it rather than serving it.
• Gross outright purchases, redemptions and “other transactions” do not sum exactly to the net JGB flow in every month: some components, such as repo sales to the government, have no live flow series. No surface presents those three as a complete decomposition.

• boj-loan-rates: the average contracted interest rate on loans and discounts (貸出約定平均金利, database IR04) by type of lender — domestically licensed banks, city banks, regional banks, second-tier regional banks and shinkin banks — on new loans made in the month (total, short-term, long-term; from October 1993) and on the outstanding loan book at month-end (total, short, long, overdrafts; loans on deeds and discounted bills; from 1976). Sixty-three series, every code checked against the BOJ's own catalogue.
• boj-deposit-rates: the average of the rates financial institutions post on ordinary deposits and on time deposits by term (1 month to 10 years) and by size of deposit (under ¥3 million, ¥3–10 million, ¥10 million and over), monthly from April 2022 (database IR02). The BOJ's earlier weekly regimes of the same survey carry different codes and are not spliced on.
• A rate is a level in percent per year. The change in a rate is in percentage points and is shown as value[t] − value[t−1 month] or value[t] − value[t−12 months]. The API refuses the ratio measures (year over year, month over month, annualised) on a rate: a loan rate going from 1% to 2% is not "+100%".
• The lender groups were re-based at March 2002: the trust-bank, long-term-credit-bank and sub-total lines end there and are kept as their own series. Shinkin banks' outstanding-rate series have 32 unpublished months in 1976–1978; the gaps are the BOJ's.
• Both datasets carry the BOJ's mandatory credit line wherever they are shown or exported.

• Yields are the Ministry of Finance's daily constant-maturity par yields for 15 tenors (1–10, 15, 20, 25, 30, 40 years), in percent to three decimals, computed by the Ministry from Japan Securities Dealers Association over-the-counter reference prices. They are shown exactly as published — nothing is fitted, interpolated or smoothed. A tenor with no published value on a date is a gap, and the curve line breaks rather than drawing across it.
• Coverage starts 24 September 1974. Longer tenors begin later — 10-year in July 1986, 20-year in December 1986, 15-year in August 1991, 30-year in September 1999, 25-year in March 2004, 40-year in November 2007 — and the 1-, 2- and 3-year series have real mid-series gaps where short tenors were not quoted.
• Spreads: 2s10s[t] = 10Y[t] − 2Y[t] and 10s30s[t] = 30Y[t] − 10Y[t], in percentage points. Changes: value[t] − value[b], where b is the latest business day on or before the same calendar date one month or one year earlier.
• Yields can be genuinely negative — the 1-year traded below zero through the negative-rate era — and no surface treats a negative yield as missing.
• The policy markers on the history chart are Bank of Japan announcement dates (ZIRP, QE, QQE, NIRP, yield curve control and its end, and subsequent policy-rate changes). They are dates only: annotations, not data.

• Every figure is the FSA's own, from table 1 of its half-yearly 金融再生法開示債権 release: claims and total credit in ¥100mn (the FSA rounds each line to the nearest ¥1bn), the bad-loan ratio in percent, the disposal loss and real net business profit in ¥tn. Units differ by line and nothing is rescaled.
• No single release covers the whole history — each carries about twenty half-years — so the dataset stitches table 1 of every release since 2006 and, where two overlap, the newer release wins. That is how the FSA's own restatements are carried; the vintage store keeps what each ingest changed.
• The disposal loss and real net business profit are the full fiscal year at March and the first half at September (the FSA's note 9). They are stored as two series each, .fy and .h1, and are never drawn as one line.
• A half-year is dated to the first day of its closing month (3月期 → 1 March, 9月期 → 1 September), so the twelve-month comparison lines March up with March.

• The FSA's two long-run workbooks behind its results summary: major banks from March 2008 and regional banks from March 2005. Income-statement lines in ¥100mn, loans and bad loans in ¥tn, as published. Ratios the FSA stores as fractions are multiplied by 100, exactly.
• The 3月期 sheet is the full fiscal year and the 9月期 sheet the first half, so every income-statement line is two series (.fy, .h1). Stocks and ratios are one series.
• The major-bank income statement is consolidated (the FSA's （連結） lines) and the regional-bank one is not; the majors' 業務純益 and the regionals' 実質業務純益 are different measures and carry different codes. Capital ratios split at March 2013 into the international and domestic standards; the pre-split ratio is its own series.

• Balance sheet and income statement of every JBA member bank, non-consolidated and consolidated, plus the association's aggregates by bank type, in millions of yen exactly as published in 全国銀行財務諸表分析. The share and change columns the association prints are not stored — they are calculations, and where a page shows one it shows the formula.
• Series are {s|c}.{bank}.{line}: the bank is its 金融機関コード (0001 Mizuho, 0005 MUFG Bank, 0009 SMBC) or a group (all, city, regional, regional-2, trust, other); the line is the association's reference code (a015 deposits, d395 loans, e800 ordinary profit, f850 net income). A code survives renamings and mergers, so a merged bank's series runs on under the surviving code from the merger date.
• Each edition is one point in time, so every edition since fiscal 2001 is read and stitched (earlier editions are Excel 5/95 files the platform cannot open). Editions before fiscal 2014 numbered the lines differently and their consolidated sheets carry no codes; those are matched by line name within the section of the statement, and a line with no modern counterpart is dropped and counted in the release's validation record (about a fifth of lines before fiscal 2006, 2–9% to fiscal 2013, under 1% today).
• Income-statement lines are .fy on March points and .h1 on September points; balance-sheet lines are one series. Trust-account (信託財産) tables are not ingested.

• The FSA's annual 中小・地域金融機関情報一覧: every regional bank, second-tier regional bank, shinkin bank and credit co-operative with deposits, loans, the capital ratio, the bad-loan ratio, branches, loans to SMEs and the number of SME borrowers for two year-ends, compiled by the FSA from each institution's own disclosure. Series are {金融機関コード}.{line} (0134 Chiba Bank, 1001 Hokkaido Shinkin).
• The bank file is in 億円 and the co-operative file in 百万円; the bank figures are multiplied by 100 — an exact change of unit — so every amount is in 百万円 and one loan-to-deposit ranking can run across all five hundred. Ratios are in percent as published.
• The FSA overwrites the same file each year, so the history is only the editions ingested here. Loans to SMEs and SME borrowers carry two year-ends per edition; every other figure is dated to the later one. An institution listed with no figures (one being wound up) is a gap, never zero.

• Maturity ladder (金銭債権及び満期のある有価証券の連結決算日後の償還予定額): securities and loans by remaining term in six buckets — within 1 year, 1–3, 3–5, 5–7, 7–10 and over 10 years — with the securities split into held-to-maturity and available-for-sale and each into JGBs, municipal, corporate, foreign and other; and deposits, CDs, borrowings and bonds by the same terms. Read from the HTML tables of the financial-instruments note in each bank's 有価証券報告書 on EDINET; nothing in it is tagged as a number. Amounts are principal, as the note states, not book value. A few filers print five buckets with an open-ended 7年超; that column is stored as such (bucket_scheme = 5_open7) and never read as 7–10.
• Unrealised gains and losses (（有価証券関係） その他有価証券): book value, cost and the difference by type — equities, bonds (JGBs, municipal, corporate), other (foreign bonds, investment trusts) — in the note's own two blocks (above cost, below cost) and the total; held-to-maturity bonds' book value, fair value and difference from the neighbouring table.
• Gates. The current-year ladder's 合計 must equal the sum of its top-level rows in every bucket (a 有価証券 line the filer leaves blank takes the sum of its printed parts, which is what the 合計 adds); the available-for-sale table's 合計 must equal its two blocks and its difference must equal book value less cost. A filing that fails is partial: shown as printed on its own page, excluded from rankings and group sums.
• Rate risk (IRRBB1). ΔEVE and ΔNII under the six Basel rate shocks, the largest of them and Tier 1 capital, exactly as each bank printed them in its own Pillar 3 disclosure (自己資本の充実の状況 → 金利リスク). That table is not on EDINET: it is found by starting from the disclosure page the FSA records for the institution, walking that site at most two pages deep, converting candidate PDFs to text and accepting a document only when its text carries the IRRBB1 template. The source URL, the PDF's SHA-256 and the page are stored and the table's text is archived. Gates: 最大値 must equal the largest of the six ΔEVE values; Tier 1 must be positive and larger than the largest ΔEVE. An as-of date read from the document rather than the table is recorded as such.
• Derived, with formula: unrealised bond loss as a share of net assets = available-for-sale bond difference ÷ consolidated net assets (純資産, from the same report) × 100; share beyond ten years = over-10-year securities ÷ the sum of all six buckets × 100; ΔEVE ÷ Tier 1 = 最大値 ΔEVE ÷ Tier 1 × 100, both from the same table. Group sums add only filings that passed every gate.
• Keys. The report data is by EDINET filer (securities code); the rate-risk table by the FSA's 金融機関コード, because the bank publishes it, not its holding company. The two are linked by name plus a fixed table for banks whose report is filed by a listed holding company (Yokohama Bank → Yokohama FG); fi:0138 addresses a bank by its FSA code.

margin-jp is Japan Exchange Group's 信用取引現在高: how much stock is held long on borrowed money and how much is held short on borrowed stock, on the Tokyo and Nagoya markets, every week since 2 August 2002. Levels are official statistics exactly as published; the margin ratio and every change are calculated here.

• Two workbooks, one balance, and they check each other. The exchange publishes the same total cut by whose account it sits in (委託 customers, 自己 the member firm's own book) and by which kind of margin it is (一般 negotiable, 制度 standardised). Each cut sums to the same published total, and both identities are checked at every ingest, in both units and on both sides. Those checks are also what confirm the column mapping: these workbooks label their columns with merged header cells that no parser can read unambiguously.
• The two workbooks restate the same total, and disagree in two weeks out of 1,224. At 2003-06-13 and 2004-03-12 the account file and the kind file print slightly different totals — at most 0.43%. Both are JPX's own figures; the account cut is the one stored, and the discrepancy is disclosed rather than reconciled away. An ingest fails if any week's two totals differ by more than half a percent.
• Shares and value are different measures. Every balance is published twice, as a share count in 千株 and a value in 百万円, and each is stored in its published unit. The value of an account balance is struck at contract prices, so it is not a mark-to-market of the share count, and the two are never summed or ranked against one another. Charts show one unit at a time.
• A definition break in July 2013, marked on the chart. Figures up to the 2013-07-12 application date cover the Tokyo, Osaka and Nagoya markets; from 2013-07-19 they cover Tokyo and Nagoya, the Osaka cash equity market having merged into Tokyo. The level steps at that week. It is stored exactly as published either side of it and the break is drawn on the chart, never smoothed.
• The margin ratio is calculated, and it is 信用倍率, not 取組比率. ratio[t] = total shares bought on margin[t] ÷ total shares sold short[t], from published share counts — the number a Japanese equity desk quotes. It is the reciprocal of the exchange's own 取組比率, which is published per issue and divides the other way round.
• A week is dated by its application date (申込日), as the exchange dates the table — a Friday in all but about forty of the weeks, and earlier when the market closed for a holiday. A week the exchange did not publish is absent, never zero. Because the periods are weekly, the year-on-year and month-on-month measures the monthly datasets serve are not offered here: a twelve-month lag does not land on a week.
• Per-issue balances are not in this dataset. The weekly table by company and the daily one for the watchlist issues are separate publications; this is the market aggregate only.

investor-flows-jp is Japan Exchange Group's 投資部門別売買状況: what each kind of investor bought and sold in Japanese equities each week, from the returns of trading participants with ¥3bn or more of capital. Sales and purchases are official statistics exactly as published; the net is calculated here.

• The weekly history cannot be downloaded, so it is kept as it appears. JPX leaves only the current month of weekly files on its site; the monthly page holds the current year and the annual page ten yearly totals. Every weekly workbook is therefore archived here under its own SHA-256 as it is published, and the full history is rebuilt from that archive at each ingest. The record begins when capture began and grows a week at a time; it cannot be extended backwards by us or by anyone else.
• Every week is published twice, and both copies must agree. Each workbook carries the week it reports and the week before it, so consecutive files overlap. That overlap is not redundancy: two independently typed copies of one figure must match exactly, and an ingest stops if they do not — which is how a revision announces itself. A revision to a week that has already fallen off the site cannot reach us, and nothing here pretends otherwise.
• The table adds up four levels deep, and every level is checked. Proprietary plus brokerage make the total; institutions, individuals, foreigners and securities companies make brokerage; investment trusts, business corporations, other corporations and financial institutions make institutions; insurers, banks, trust banks and other financials make financial institutions. All four identities are checked at every ingest, in both units and on both sides, and a file that breaks one is not published.
• Net purchases are calculated. net[investor, t] = purchases[investor, t] − sales[investor, t], from the published gross figures; it equals the 差引き balance column JPX prints. It is the only figure in this dataset that can be negative — a gross sale or purchase never is.
• The published ratio columns are not stored. Their denominator changes down the page: the proprietary and brokerage ratios are shares of the grand total, the investor-category ratios shares of brokerage alone. One column meaning two things is worse than no column, so a share is computed here against whichever total is meant.
• Value and shares are different measures. Every figure is published twice, in ¥ thousand and in thousands of shares, each stored in its published unit and never ranked against the other.
• What the survey covers. Trading participants with capital of ¥3bn or more, in domestic common stocks only (no preferred shares), including ToSTNeT off-auction trades. It is most of the market's turnover, not all of it, and the exchange says so on the file. Weeks are dated by the Monday they begin.

A holder whose short position reaches 0.5% of a listed company's shares must report it, and Japan Exchange Group publishes every such report the next business day. The ratio (空売り残高割合), the share count (空売り残高数量), the trading-unit count, the calculation date (計算年月日) and the publication date are official statistics exactly as published. Whether a position is still open, a company's total across its holders, and the move since a holder's previous report are calculated here and carry their formulas.

• The archive cannot be back-filled, and that is the headline limitation. JPX keeps roughly twelve business days of these files on its site and then deletes them. There is no archive, no API and no second source, so this record begins the day capture began and can never be extended backwards by us or anyone else. Every surface states how many publication days it holds, because that number is the completeness of everything computed from it.
• Each file is a tape of reports, not the outstanding book. A daily workbook carries the reports received that day, dated by their own calculation date — usually two business days earlier. A holder files again only when the position moves, so consecutive files overlap only partly. The open book is therefore derived: the newest report per (issuer, holder, discretionary manager, fund) by calculation date, counted as open unless that report is a below-threshold closing report. A position opened before the archive began and not moved since has never been published into it and is not in the book.
• A closing report's zero is a real zero. A holder falling under 0.5% files one final report saying so, with a ratio at or near zero. That is the position going away, not a missing value, and it is what removes the position from the book. Missing is still shown as an em dash and is never exported as 0.
• A total is the disclosed short, never short interest. Only positions of 0.5% or more are disclosed at all, so a company's total across its holders understates the true short by whatever sits below the threshold. It is labelled disclosed short everywhere and is never presented as short interest.
• Ratios are added only within one company. Holders of the same issuer share a denominator, so their published ratios sum. A short of 3% of one company and 3% of another are different measures: no total across companies is shown, and a holder is ranked by its count of positions and its largest one instead. Share counts follow the same rule.
• The ratio is JPX's, not the statutory one. 空売り残高割合 is measured against shares outstanding. It is not the 株券等保有割合 used by the 5% long filings, whose denominator includes the holder's own potential shares, so the two are never netted against each other or ranked together. The workbook writes the ratio as a decimal fraction; it is stored and shown as the percentage JPX's own PDF prints, which is the same number.
• Holders are grouped by a folded name. Case, punctuation, spacing and the legal-form tail (plc, Ltd, LLC and the like) are folded so one holder is one row however a file spelled it. Nothing that distinguishes one arm of a group from another is removed: Morgan Stanley & Co. International plc never merges with Morgan Stanley MUFG Securities. A holder reporting through a discretionary mandate names the manager and fund, and those are part of the position's identity rather than decoration.
• Every daily file is a vintage. Each workbook is archived with the SHA-256 of its bytes before anything is parsed, and a stored file is never rewritten: reading the same file twice stores nothing, and a revised file for the same day lands beside the first. The Daily Files Held table on the page lists what is held, so a day that is missing is visible rather than inferred.
• A file that fails a gate is not stored at all. The parser refuses a workbook whose Japanese header row has changed, whose internal publication date disagrees with its filename, that carries fewer than fifty reports, that omits an issuer code, holder or share count, or that states a ratio outside 0–50% or a calculation date after publication or more than thirty days before it. The days already held stay live.

• Nothing is recomputed. Every figure on the Financials page and under /api/v1/equity/financials/ is the value tagged in the company's own annual securities report (有価証券報告書) on EDINET: the filer's five-year summary (主要な経営指標等の推移) and its balance sheet, income statement, statement of comprehensive income, cash flow statement and statement of changes in equity, consolidated and parent-only. ROE, equity ratio, PER and payout ratio are the filer's figures — filers differ in how they compute them (average versus year-end equity, for instance) and this platform does not harmonise them.
• Percent fields. The taxonomy stores ratios as fractions (0.101 for 10.1%); fields named *_pct are that fraction × 100 and nothing else.
• Statement order and membership come from the filing's own presentation linkbase, so a statement prints line for line as the filer laid it out, including the filer's own extension line items. A line the filer marks with a negated label is one it prints with the sign reversed (a cost shown positive); the API flags it and the value is left as tagged.
• Standardised field names (revenue, profit, roe_pct…) are a mapping from element names, one entry covering the Japan GAAP element and its IFRS twin; every row names the element behind each field. Where a filer tags no standard revenue element — Toyota's summary uses its own — revenue is the first yen line of the filer's summary table and the row says so (revenue_source = first_line). Bank and insurer "revenue" is 経常収益 (ordinary income), as those filers report it.
• Restatements are kept. Each filing restates five fiscal years. The company panel takes each fiscal year from the latest accepted filing that covers it and names that filing on the row; ?as_filed_in=YYYY serves one filing's five years exactly as first published, which is the point-in-time view. Fiscal-year labels name the month and year the company's year ends — FY Aug-2025 for Fast Retailing's year to 31 August 2025 — as the company itself does (2025年8月期), on every company page, segment table and customs lens. They are the filing's period end shifted by whole years, so a filer that changed its year-end inside the window is labelled by that arithmetic, not by its old year-end. (The government's April–March year keeps its own convention on the fiscal pages: FY2024 is April 2024 to March 2025.)
• Split-adjusted per-share figures are calculated, and sit beside the filed figures, never in place of them. Filers restate EPS and book value per share for a stock split only as far back as they choose, and often print dividends as paid: Fast Retailing's report for the year to August 2025 shows FY Aug-2021 EPS of ¥1,663 on the pre-split share count beside FY Aug-2022 EPS of ¥892 after its 3-for-1 split. The adjusted figure puts every year on the latest year's share count: split-adjusted = filed ÷ the split ratios between that year and the latest. Which share count a filed figure used is read from the figures themselves — profit ÷ EPS (else net assets ÷ BPS) from one year to the next — and a jump counts as a split only when a split of the same ratio also shows in the issued-share count, or in the filer restating its own earlier EPS by that ratio between two reports. A move no split explains (a buyback, a share issue) is left as a real change. Dividends per share get their own test, because a filer may restate EPS and not dividends: dividend ÷ EPS is compared with the filer's payout ratio, and a clean multiple of it shows the dividend is on an older count. Where a year's share count cannot be read this way — no payout ratio, a loss year, a dividend paid half before and half after a split — the adjusted figure is — rather than a guess.
• The only derived numbers on a statement are the change columns: change = current − prior on the tagged values and change % = change ÷ |prior| × 100, shown as — where the prior value is missing or zero.
• Labels. Japanese labels are the filer's own. English labels are the filer's own where the filing carries one; otherwise a readable form of the XBRL element name, marked derived — never a translation of the Japanese by this platform.
• Gate. A filing is accepted as clean when its tagged balance sheet satisfies assets = liabilities + net assets on every basis it reports, and its summary's net assets do not exceed total assets; otherwise it is partial with the reason stated, and a cross-section says which. A filer whose statements are not tagged (US GAAP filers tag them as text) still has its summary figures.

The Screener and the "Calculated Ratios" block on each company's Financials page are the one place this platform computes a number from the filed statements. Same rule as every other derived figure: no badge, the formula on the page, every input named, and the filer's own figure beside ours wherever the filer prints the same ratio.

• ROE = profit attributable to owners of parent ÷ average of opening and closing equity attributable to owners × 100. ROA = the same profit ÷ average total assets × 100. Averages are the convention Japanese filers use for their own ROE; a filing whose opening balance is not tagged gets no ratio rather than a closing-balance one. Across the current universe our ROE is within 0.5 pp of the filer's own for over 99% of companies; the difference is shown per row, and a gap over 1 pp is marked. Every remaining difference is a filer convention, not an error: all six companies that differ by more than 2 pp divided by closing equity rather than the average, and our figure recomputed on their basis reproduces each of them exactly. Filers disagree with one another here — Chiyoda uses the average and matches this platform at 123.1%, Kodama Chemical uses the closing balance and prints 82.8% where the average gives 144.9% — which is why one convention is applied to every company and the filer's own figure is always shown beside it.
• A return is not calculated on a negative base. Where average equity over the year is zero or negative the ROE is withheld and the company is flagged negative equity instead. Dividing a loss by negative equity inverts the sign: Japan Display closed FY2026 with equity of −¥7.5bn against +¥6.7bn a year earlier, and its ¥19.8bn loss over that negative average would otherwise read as an ROE of +4,570% and rank first in the market. The test is applied to the denominator actually used, so a company that recovered from negative equity during the year still gets a ratio; across the universe this withholds exactly the three companies whose own filings also decline to print an ROE. The equity ratio is still shown for these companies — it is negative, which is the point, and it matches the filer's own figure.
• Where the base is not available, no ratio is invented. US GAAP filers publish their statements as text, so their equity comes from the five-year summary, and only from a line that names the owners of the parent — Komatsu, ORIX, Fujifilm and Omron all reconcile to their own published ROE within 0.05 pp on that basis. Nomura tags only equity including minority interests, which is the wrong base for a return to shareholders, so this platform shows no ROE for it and prints the company's own figure instead.
• Equity attributable to owners (自己資本): under IFRS the tagged element; under Japan GAAP net assets less non-controlling interests less subscription rights to shares. A balance sheet that prints no such line has none, so an absent subtrahend is nil — the only case in which an absent tag is treated as zero, and only for a subtractive component under a total that is present.
• Revenue and profit are the filer's own summary lines, resolved as on the key-indicators panel; banks and insurers report 経常収益 as revenue. Operating income is the summary's where printed, otherwise the income statement's; banks, insurers and most IFRS filers print none, so their operating margin is —.
• Margins: operating income ÷ revenue, profit ÷ revenue. Asset turnover: revenue ÷ average total assets. Growth: current ÷ prior − 1 from the same filing; profit growth is not computed off a prior-year loss. Cash conversion: operating cash flow ÷ profit, not computed for a loss. Free cash flow is operating cash flow less capital expenditure: the cash flow statement's purchases of property, plant and equipment and of intangible assets (or the filer's one combined line, never both). A filer's own element for the same payment is read by its name; retirements, equipment leased out to customers and "other" fixed assets are not counted. Until September 2026 it was operating plus all investing cash flow, which counted securities bought with spare cash as spending and a cross-shareholding sale as free cash — Sony read −¥25bn against ¥1.49tn on capex, Denso +¥494bn against ¥77bn. No capex line tagged, no figure; banks, insurers and brokers get none, because their operating cash flow is the balance sheet moving.
• Cash position, read from the lines printed on the face of the balance sheet, three figures side by side. Liquid cash: cash and deposits plus short-term securities (有価証券 in current assets) under Japan GAAP; under IFRS, cash and cash equivalents only, because IFRS balance sheets put time deposits and short-term investments in "other financial assets" together with derivatives and loans. Net cash: liquid cash less interest-bearing debt — short- and long-term borrowings, bonds, convertible bonds and commercial paper, current and non-current, non-recourse loans included. Lease liabilities are left out: under IFRS nearly every lease is on the balance sheet and under Japan GAAP most are not, so counting them would make an IFRS filer look more indebted for an accounting reason. A debt line the balance sheet does not print is a debt the company does not have. Where one line mixes borrowings with leases or other liabilities — including an IFRS "interest-bearing liabilities" line on a balance sheet with no lease line of its own, which is how Toyota files — net cash is not computed. The figures are consolidated, so a captive finance arm's borrowing counts as debt: a carmaker's auto-loan funding makes it a large net borrower. Investment securities (投資有価証券 — shareholdings in other companies, cross-shareholdings among them, and bonds held for over a year) are shown on their own and never counted as cash; IFRS has no equivalent line. Each is also shown ÷ total assets. Banks, insurers and brokers get none of these: cash and securities are their operating assets. Until September 2026 "cash" was the summary's cash and cash equivalents where printed and the balance sheet's cash and deposits otherwise — two definitions in one column, and one that left out Keyence's ¥897bn of short-term securities.
• Implied valuation. This platform holds no prices. Where the filer prints its own year-end PER, an implied price = PER × EPS gives an implied PBR (÷ book value per share) and an implied dividend yield (DPS ÷ implied price). Both are labelled implied and are only as current as the filer's year-end. Per-share dividends are printed in the reporting company's own summary table, so the yield is computed entirely inside that table and never mixes the group's per-share figures with the parent's. A share split between the fiscal year-end and the filing can still leave the paid dividend and the restated EPS on different share bases; the filer's own payout ratio exposes this, and where it disagrees with DPS ÷ EPS by more than 25 pp no yield is derived and the row says why.
• Basis and ranking. Consolidated where the filing carries consolidated statements, parent-only otherwise; every row says which. Rankings use one filing per company, its latest accepted annual report; companies without the ranking metric are left out and counted on the page. Filters are floors and caps in each metric's own unit.

• Revenue by region, the customers a company names, and its reportable segments come from the segment-information note (セグメント情報等) of the annual securities report on EDINET, for the current and prior fiscal year. The note is a text block, not tagged numbers: every figure is parsed from the filer's own HTML tables, and each filing carries a status. A filing whose region revenues do not sum to its consolidated revenue within 0.5% is marked partial with the numbers in the reason; nothing is adjusted to make it reconcile. Across the June 2026 filing season, 95.9% of the filings that carry a region table reconcile, and 91% of all filings are clean.
• Regions are on the filer's own basis — customer location for most, shipping destination for some — and the basis sentence is stored and shown with the figures. The filer's label is kept verbatim; a small lexicon maps it to a key (JP, CN, TW, KR, US, NA, EU…) so two filers' "China" can be read together, and a label the lexicon does not know keeps no key rather than a guess. A sub-note such as 「北米のうち、米国は…」 is its own row, marked as a sub-note of its parent, and is never added to the region it sits inside.
• A filer that omits the region table says why — no overseas revenue, or domestic revenue above 90% — and that reason is stored. The absence is then a fact, not a gap, and no region row is invented for it.
• Named customers are every customer the filer discloses at ten percent or more of revenue, with the revenue booked from them, as written. Some filers give them in a table, some in prose; the source is marked. Fiscal years are named the Japanese way, for the calendar year in which they begin: the year ending March 2026 is FY2025.
• Fiscal roll-ups of customs data: Σ over the fiscal year's months of the published customs value for every partner the region key maps to. Only monthly flows are summed — customs values and quantities, visitor counts — and the API refuses to sum an index, a stock or a rate. Only complete periods are shown: a fiscal quarter or year with an unpublished month is left out, never summed short, and the months left out are listed.
• The commodity-to-company mapping is a platform classification, not an official source label. Customs records carry no company identity. A company is mapped to a customs line only where its own annual report describes that product as a reportable segment or principal product; the file is versioned in the repository with an owner and a review date, and every surface that uses it says so.
• Implied share = revenue the company filed for a region / customs exports to that region's partners over the same fiscal year × 100. An indicator of size and direction, never an accounting identity. A company ships from overseas plants, sells domestically, and books revenue on acceptance rather than shipment; the customs line includes its competitors and trading houses; and the company's region and the customs partner list are different definitions that share a name. An implied share above 100% is that caveat made visible. Only regions that name a place are joined — China, Taiwan, Korea, the United States, North America, Europe — and "Asia" never is, because one filer's Asia includes Japan and another's excludes China.
• Named customers and dependence. A listed company must disclose any customer worth ten percent or more of consolidated revenue and state the amount. dependence[filer, customer] = customer revenue / the filer's consolidated revenue × 100, both as filed for the same fiscal year; combined dependence is the sum over every customer that filer names. It cannot exceed 100%: a filing where it did was a parsing fault, and two gates now catch that class of error — a single customer larger than the filer's whole revenue, and a combined total above it. A company absent from the list has no customer at or above the threshold, which is not the same as having a diversified book. Where consolidated revenue has not been extracted the share is blank, never assumed.
• English customer names are looked up, never translated. The English name beside a filed Japanese one comes from the EDINET company registry where the buyer files its own annual report — which also yields a securities code, so the buyer's own page can be opened from the row — or from a small curated list, versioned in the repository with an owner and a review date, for government bodies, unlisted subsidiaries and foreign affiliates the registry cannot cover. Where no published English name exists, none is invented. Of 1,010 distinct customer names, 363 resolve to an English name and 132 are already in Latin script; the remaining 515 are shown as filed. A subsidiary keeps its own identity — Amazon Japan G.K. is not Amazon.com, and TEPCO Energy Partner is not TEPCO — because rolling a subsidiary up to its parent would state a relationship no filer disclosed.
• Customer names are never merged. They are shown exactly as each filer wrote them: Toyota appears as both 「トヨタ自動車(株)」 and 「トヨタ自動車株式会社」, and Samsung under three different strings, because that is what the filings say. Combining them would state a total no company disclosed, so search matches text and the page tells the reader to try more than one spelling.

A listed company publishes its 決算短信 on TDnet, the Tokyo Stock Exchange's disclosure wire, weeks before the statutory report reaches EDINET. Since the statutory first- and third-quarter reports were abolished in April 2024 it is the only source for those quarters. Amounts, per-share figures, the company's forecast with any upper and lower bound, and the year-on-year percentages are official statistics exactly as each company tagged them in the release's Summary. Progress against the forecast, which release is current for a period, and the fiscal period are calculated here and carry their formulas.

• Progress (%) = cumulative result ÷ full-year company forecast × 100, for the same line and basis, both from the same release. It is not calculated when the forecast is absent, a range, or zero or negative. Companies are ranked within one reporting period only: three months of progress and nine months of progress do not compare.
• Results are cumulative. A second-quarter release covers six months and a third-quarter release nine. A stand-alone quarter is never derived, because the difference of two rounded cumulative figures is not a published number.
• Results and forecasts never share a column. They are the same element in the same unit, so every table and export keeps them apart. A forecast the company did not give is missing, never zero.
• Year-on-year changes are the company's own. The release tags the percentage it prints; it is shown as published and not recalculated from the two amounts.
• One period can have several releases. A company re-issues its release when the auditor's interim review completes, or to correct a number. Every release is kept as disclosed; the newest by disclosure time is marked current and the earlier one names it.
• Accounting standards are mapped, not equated. The headline lines use the nearest published element for each standard and always show which one: ordinary income under Japanese GAAP and profit before tax under IFRS share a row because they sit in the same place in the Summary, and they are not compared across standards.
• The fiscal period is read from the headline. 2027年3月期 becomes 2027-03, and an era year is converted (令和9年 is 2027). A headline that names no month has no fiscal period.
• The record cannot be back-filled. TDnet deletes a disclosure after about a month. Capture began on 10 July 2026; nothing before it can be recovered. Each package is archived with the SHA-256 of its bytes before anything is parsed.

Every annual securities report carries a section, 事業等のリスク, where management lists the risks it believes could materially affect the business. The text is an official disclosure exactly as filed: spacing is tidied and any table inside the section is shown one row per line, but no word is changed, summarised, scored or translated. Two things are calculated here and say so.

• Items. The section is cut at the company's own numbered headings — (1), ①, 1., (ア), (a). A line starts a new item only if it carries the next number in its sequence, so a numbered list inside a risk never splits it. A second numbering style inside the first becomes sub-items. A section with no numbers is cut at short bracketed headings, such as （為替変動について）, when at least two occur and none repeats. A section with neither is shown whole. The text before the first item is kept as the preamble. The preamble and the items together reproduce the full section line for line; a filing where they do not is flagged.
• New and dropped. An item is marked new when its heading, with the numbering removed, is not among the headings of the same company's previous annual report on file. The comparison is of exact text, not meaning: a reworded heading shows as one new item and one dropped heading.
• Text search is exact. It matches the filed Japanese character for character. 為替 finds 為替レート, but サイバー does not find 情報セキュリティ.
• Companies, not survivors. Every annual report carrying a securities code is read, including companies since delisted. Each filing is archived with its SHA-256 before parsing.

A rank across the whole listed market answers almost nothing: a 3% female board is unremarkable among 3,700 issues and conspicuous inside TOPIX Core30. The Peer Groups page and the cohort parameter on every equity screen restrict a cross-section to a peer group before anything is ranked. Every figure compared is the same one its own page publishes; the statistics about a cohort are calculated here and carry their formulas.

• The classification is not ours. Market segment (Prime, Standard, Growth), the 33- and 17-industry classifications and the TOPIX scale band (Core30, Large70, Mid400, Small 1, Small 2) are taken exactly as Japan Exchange Group publishes them in the listed issue list (東証上場銘柄一覧), and are never inferred, recomputed or back-filled. The issues carrying a scale band are the TOPIX constituents — 1,636 of them at 2026-08-31. This platform does not compute index membership and does not maintain an index.
• Membership is a vintage. Each accepted read of the JPX file is stored under its own effective date and the SHA-256 of its bytes, and a stored vintage is never rewritten: a revised file lands beside the first, so what a cohort contained on a given date stays recoverable. Index reviews, segment migrations, listings and delistings accumulate as history from the first read forward. Neither publisher offers back-history, so a request for a date before collection began answers with nothing rather than quietly returning today's membership.
• Index membership is not published here. The Nikkei 225 constituent list is Nikkei's copyrighted work and may not be redistributed. Where it is held it is held for internal comparison only, behind an explicit switch, and it is absent from the public catalogue, the API and the assistant tools unless a deployment is entitled to it.
• Baskets are not stored. A basket of securities codes is carried entirely in the address bar (cohort=codes:8001,8002,…), which is what makes any comparison built on one a permanent, citable link. Names given to baskets are kept in the reader's own browser and are never uploaded.
• Median and quartiles are taken over the cohort members that report the measure, by linear interpolation between the closest ranks of the sorted values (the conventional method, R type 7): for n values the p-quantile sits at position (n − 1) × p. Percentile = members strictly below the value, plus half of those equal to it, ÷ members with a value × 100 — the mid-rank convention, so a company in the middle of its cohort reads 50 whichever way ties fall. vs median is the company's value minus the cohort median, in the measure's own unit; for a measure in per cent that difference is in percentage points.
• A company that does not report a measure is not a company reporting nil. Members without a value are excluded from the distribution and from every rank, listed separately with the reason, and counted on the page. They are never imputed and never treated as zero.
• Reference periods differ across a cohort. Each company contributes its most recent annual report and Japanese fiscal year-ends are staggered, so a cross-section carries no single as-of; every row shows its own period end. The board, register, financial-statement and cross-shareholding figures for one company can also sit on different filings, and each is labelled with the filing it came from.
• The distribution chart may cut its axis. Where a cohort's tails are extreme — operating margin across electrical-machinery makers runs from a four-figure negative to a four-figure positive — the plotted range is cut to the 1st–99th percentile so the shape is visible. Values beyond it are counted in the end bars rather than dropped, the caption says how many, and the table beneath carries every value in full.

• cpi-us, cpi-us-sa and cpi-us-areas are three cuts of the BLS CPI for All Urban Consumers (CPI-U) flat files: U.S. city average not seasonally adjusted (about 400 series, all items from 1913), seasonally adjusted (about 325, from 1947), and regions, size classes and 23 metro areas plus urban Alaska and Hawaii (about 3,400). Codes are the BLS series ids (CUUR0000SA0), so every figure can be checked on bls.gov. Rates use the same formulas as the Japanese CPI (see Calculations) and can differ from the BLS's own published percent change by about a tenth of a point.
• Reference periods differ by series. Most indices are 1982-84 = 100; an item introduced later carries its own base (Dec 1997 = 100, Dec 2017 = 100, …) and names it in its title. Rates of change do not depend on the base.
• Not carried: semiannual series, alternate-base (“old base”) duplicates, the purchasing power of the consumer dollar, and the BLS annual averages (M13).
• Gaps. October 2025 is blank throughout, flagged by the BLS “data unavailable due to the 2025 lapse in appropriations”, so October 2026's year-on-year rate will be missing too. Most metro areas are priced every other month; the months between are absent. The BLS prints 0.0 for some months that were not priced (quarterly area pricing in 1935–40, a few metro fuels in 1996); an index or a price cannot be zero, so those cells are treated as missing.
• No contribution-to-headline split for the US. The BLS publishes relative importance for December only, and it moves every month with relative prices, so a fixed-weight split would not add up to headline. The December weights are served as cpi-us-weights (Table 1, CPI-U, percent of all items, December 2020 onward), matched to CPI item codes by name; about 27 rows with no published index keep an RI- code. The eight major groups must sum to 100 within 0.01 or the release is not published.
• us-avg-prices are the BLS average prices in dollars per the unit in each item's name (per dozen, per lb., per gallon, per kWh). They are not quality-adjusted and are not a substitute for the item index; two items' dollar levels are never charted on one axis.
• us-wages are average hourly and weekly earnings from the BLS payroll survey (CES), total private and its two halves, current and 1982-84 dollars. The real series are the BLS's own; nothing is deflated here. Each month is revised twice and the whole series is re-benchmarked every February; each revision is a vintage.
• us-wage-tracker is the Atlanta Fed Wage Growth Tracker: the median of 100 × ln(wage[t] ÷ wage[t − 12]) for the same people a year apart. Headline cuts are 3-month moving averages, demographic cuts 12-month; the window is in every series name. Values are rates as published; a percentage change of one is refused.

• Yields are the Treasury's own daily par yield curve rates, interpolated by the Treasury from end-of-day bid-side quotations, from January 1990; and its par real yields on inflation-protected securities (TIPS) at 5, 7, 10, 20 and 30 years, from January 2003. Both are shown exactly as published.
• A maturity not yet introduced on a date is a gap, never zero. The 1-month starts in 2001, the 2-month in 2018, the 4-month in 2022 and the 1.5-month in 2025. The 30-year was discontinued in February 2002 and reintroduced in February 2006; that gap is in the published data.
• Real yields are often negative; that is the published value, not an error.
• These are daily series, so the monthly rate measures (year over year, month over month, 3-month annualized) do not apply and are refused.
• Breakeven inflation is not served. It would be nominal minus real at the same maturity: a calculated figure, which would carry its formula rather than an official badge.

• Market value is the latest close times the shares outstanding stated on the cover of the company's newest 10-K or 10-Q (the figure the company itself gives a few weeks before filing). A company with several share classes is valued class by class, each at its own close; an unlisted class at the conversion ratio its own 10-K states, and where no ratio is on file the market value is left blank with the reason. Where no cover count is on file the year-end balance-sheet count is used, and the response says so; so it is where the cover count is ten, a hundred or a thousand times the balance-sheet count (or a tenth of it) and no split in the price history explains it — a scale slip on the cover, as Alibaba's 20-F carried in 2026. A count with no second count to check it against is checked against the holders: a value above $3bn of which the 13F filers together report holding under 0.1% is held back with a note (a 20-F cover tagging 150 billion shares implied $417bn for a company no fund holds).
• Depositary shares. A foreign company's US listing is usually an American Depositary Share, and one can stand for any number of the shares the filing counts: three for HDFC Bank, eight for Alibaba, thirteen for BeOne, 4,000 for Scinai. That number is read from the company's own annual report — the Section 12(b) table on its cover, or where the cover only says "American Depositary Shares", the sentence in its text that states it — and the count is divided by it before it meets the price. P/E multiplies the filed earnings per share by the same number. A filing that names a depositary share without stating the number, or a 20-F filer not yet read, gets no market value and a note, never a guess. The source sentence is kept with each ratio. (Until October 2026 the count was used as filed, which overstated HDFC Bank three times and Alibaba eight.)
• Currency. Market value is in US dollars whatever currency the filing reports in, because the close is. P/E, dividend yield and buyback yield are computed only for filings in US dollars; rankings by any filed figure take only US-dollar filers, so a yen revenue is never ranked against a dollar one.
• Revenue is the first of the standard revenue elements a filer tags, with two exceptions where that element is not the company's revenue. A bank tags its fee income as revenue from contracts with customers and its interest income separately, so for banks (SIC 6020–6036 and 6099–6199) revenue is total interest and dividend income plus non-interest income. A landlord tags rental income as operating lease income beside a small contract line, so where lease income exceeds the revenue element found, revenue is the two together. The element recorded with the figure names both parts; a filer's own "Revenues" line is never replaced. (Until October 2026 Cullen/Frost's $122mn of fees stood as its revenue against $649mn of net income.)
• Dividends paid are dividends to common shareholders where the filer tags them separately; otherwise the all-in dividends line, which can include preferred dividends.
• What is not a stock. Funds, commodity trusts and blank-check companies file annual reports too; they are left out of rankings and counts of companies.

• Insider trades. "Insiders bought" means transaction code P, an open-market or private purchase. Awards (A), option exercises (M), shares withheld for tax (F) and gifts (G) are counted separately and never as buying or selling. A trade's value is shares times the price the filer reported; an award or exercise with no price has no value, never zero. Form 4 is due within two business days of the trade and is read the day after it is filed. Amendments keep their own row; the original is not changed.
• 13F holdings. Long positions of managers over $100m, reported up to 45 days after the quarter ends. A share of reported 13F value is not a share of the company. The filings identify a security by CUSIP, not by company, and each manager spells the issuer's name its own way. Values are in dollars from 2023 and thousands of dollars before. A line whose value per share is more than three times off the median for that security is treated as a units slip: its value counts, its share count does not.
• 5% stakes (13D and 13G). A 13D is filed by a holder who may seek to influence the company and must state its purpose; a 13G by a passive holder. A filing's stake is the largest percentage on its cover pages: the reporting persons are usually one holder's affiliates reporting the same shares, so percentages are never added up. The percentage is the filer's own. Only the structured filings since 18 December 2024 are read; earlier ones were free text.

An answer about one US company sets its figure beside one or two references, so a reader can tell whether it is high or low. Every reference is calculated here from the same figures the answer uses, and its formula is under "How This Was Worked Out" on the answer.

• Sector peers. The middle value (median) of the same figure across companies worth over $2bn in the same broad sector, each on its latest annual report, the company itself left out. The sector is read from the SEC industry code and is approximate. With fewer than eight such peers no comparison is shown. A P/E is compared only among companies with positive earnings.
• The S&P 500's middle company. The median across S&P 500 members, taken from the holdings of the SPY fund.
• The S&P 500 over the same days. The SPY fund's price change between the same two dates as the company's, adjusted for splits, before dividends.
• Insider trading. Value traded is set against the company's market value today and, when the filings reach back that far, against the same dates a year earlier. A window that starts before the first insider filing on file is cut back to it, and the answer says so.
• The largest 13F holders. How many of the fifteen largest added or cut shares in the quarter. Their changes are not summed: a manager that begins filing for its affiliates shows shares it already held as an addition.
• Words, not causes. "Above", "below" and "about the same as" describe the gap (a gap within 10% for a multiple such as P/E, 1 pp for a margin or growth rate, or 2 pp for a price change reads "about the same"); no answer says why a figure differs.

A ranking that keeps or leaves out a sector ("tech stocks", "non-tech stocks") reads each company's sector from its SEC industry code, grouped into eleven broad sectors. Technology means software, chips, computers, communications equipment and IT services. The grouping is not the S&P's own (GICS): Alphabet and Meta count as Technology here, and a company filed under an old code can land outside the sector most readers expect.

• Read as. Every ranking states the reading it ran on: the stocks ranked, the size floor, the figure and the period.
• Switches. Up to three switches re-run the same question: S&P 500 members only; stocks under $2bn included; and Amazon, Tesla, Netflix, Uber, Visa, Mastercard, PayPal, Booking, Airbnb, DoorDash and Spotify, filed outside Technology, counted as tech. The switches are part of the answer's link (opt=sp500,all,tech), so a switched view can be cited.
• When a switch matters. If counting those companies the other way would change the list shown, the answer says who would drop out and who would come in.

• Values are exactly as the C&SD publishes them. Its own percentage changes are not stored, except the official CPI inflation rates, which are kept as published in cpi-hk-rates (see Calculations for why they can differ from a rate calculated here). Series codes are the C&SD's own statistic codes, so every figure can be found in the C&SD's web tables under the same code.
• CPI. CPI(A), (B) and (C) cover households in the relatively low, medium and relatively high spending ranges, some 50%, 30% and 10% of households; the Composite CPI combines them. The underlying indices remove the effect of one-off government relief. The indices are not seasonally adjusted; the C&SD's seasonally adjusted figures exist only as a rate.
• GDP. A quarter is dated by its first month: 2026-04 is Q2 2026. Not seasonally adjusted: compare a quarter with the same quarter a year earlier. Imports are positive amounts that GDP subtracts. Changes in inventories cross zero, so a percentage change of them is refused.
• Labour. Every figure covers a rolling three months, dated by the last: 2026-08 is June to August 2026. Neighbouring months share two months of data.
• Retail. The volume index is the value with price changes removed: how much was bought, not what it cost. Supermarkets (outlet type 49) is a memo line that overlaps two other types and is not part of any sum.
• Trade. Total exports are domestic exports plus re-exports; imports are valued c.i.f., exports f.o.b. The trade balance crosses zero, so a percentage change of it is refused. A year-on-year rate calculated here matches the C&SD's own to 0.1 pp from 1968; in the 1950s and 1960s it can differ by up to 0.4 pp.
• The C&SD's provisional and revised flags are not stored; a revision arrives as a new vintage, and the release history keeps every earlier value. Annual figures are not stored: they are the year's months or quarters, summed or averaged.