090°Données ouvertes
Jeux de données ouverts, entièrement documentés — interrogeables ici, et lisibles par n’importe quel LLM.
Les titres et les descriptions proviennent des sources de données, en anglais.
10 jeux de données
US Business-Formation Signals (derived)
Weekly US business-formation signals from Census Bureau Business Formation Statistics (FRED BUSAPPWNSAUS + HBUSAPPWNSAUS, 2006 ->): total and high-propensity business applications with 13-week momentum, year-on-year change, 30-week change volatility, 3-sigma anomaly flags, naive-drift forecasts, and the high-propensity share of applications (the quality mix of the startup pipeline). The entry-margin lens on the US business cycle — applications lead formations, and high-propensity applications lead employer births. Companion to us-labor-market-signals (established firms) and us-state-coincident-activity-signals (output). All rows normalized to country_code USA. Raw series: U.S. Census Bureau Business Formation Statistics via FRED.
U.S. Census Bureau
Annual US county business-vitality intelligence from the keyless U.S. Census County Business Patterns bulk files (2021-2023): one row per year per county (~3,190 counties) per NAICS 2-digit sector (20 sectors) plus an all-sectors county row. Each row carries establishments, March-12 paid employment, Q1 and annual payroll (normalized to whole dollars), average pay per employee, the small-establishment (<5 employees) share, and the Census noise-infusion flag. Derived: exact year-over-year changes in establishments/employment/payroll (null when the prior year is absent, never interpolated), 2021->2023 two-year CAGRs, each sector's share of county employment, the county employment concentration HHI (0-10000, lower = more diversified), and a documented 0-100 county vitality score (40% winsorized employment momentum + 30% winsorized establishment momentum + 30% sector-mix diversity, min-max over scored county-years) with per-state ranks and quartile tiers, plus per-(state, sector) employment-growth ranks. Method: parse the 21-row sector panel from each CBP county file, resolve FIPS to canonical county names via the Census gazetteer, null suppressed size-class cells (never zero-fill), keep payroll in dollars. Caveats: cells carry Census noise infusion (flagged); payroll is in dollars converted from the published $1,000s; scores are relative history gauges, not forecasts; reference year lags ~2 years. U.S. federal public domain (commercial reuse allowed with attribution to the U.S. Census Bureau).
US Retail Inventory-Cycle Intelligence (derived)
Monthly US retail inventory-cycle intelligence from the U.S. Census Bureau's Monthly Retail Trade Survey (official 'Inventories and Inventories/Sales Ratios' workbook: 9 published industries — retail trade total, ex-motor-vehicle total, motor vehicle and parts dealers (441), furniture/home-furnishings/electronics (4423X), building materials (444), food and beverage (445), clothing (448), general merchandise (452), department stores (4522); 1992-01 ->). Each industry-month row carries published inventory levels and inventories/sales ratios (seasonally adjusted and not adjusted, millions of USD / months of supply), inventory momentum (MoM, 3-month-annualized, YoY), the ratio's YoY change and 36-month z-score, a documented 0-100 inventory pressure score (50% winsorized ratio z-score +-2.5 + 30% winsorized inventory YoY +-25pp + 20% winsorized ratio YoY change +-1.0pp, min-maxed within each month) with deterministic rank and p1-p4 tiers across the 7 peer industries, overstock/understock flags (|z| > 1.0), 36-month ratio record-high and inventory-surge flags, the exact unpublished-kinds inventory residual on the ex-441 row, plus month-level context (total and ex-441 ratios, overstock/understock breadth, the month's top pressure industry, median pressure score). An online shop joins monthly inventory pressure to its own stock planning on (date, industry_code) for overstock/understock and markdown timing; a sales team joins category pressure to pipeline value on date. Method: one polite GET of the official Census workbook; a trailing-36-month verbatim transcription is embedded for offline safety; the transform is pure pandas in map_schema with no imputation (warmup cells are honest nulls). Caveats: the latest month is preliminary and revised next release; the published kind detail is not exhaustive (the unpublished residual covers 446/447/451/453/454); the 2020-04 ratios for clothing (18.63) and department stores (48.93 SA / 51.88 NSA) reflect the COVID sales collapse, not a data error; department stores (4522) are a subset of general merchandise (452) and are excluded from the additive identity.
US Retail Kind-of-Business Sales Intelligence (derived)
Monthly US retail-sales intelligence at kind-of-business grain from the U.S. Census Bureau's Monthly Retail Trade Survey (12 FRED MRTSSM*USS series: NAICS 441, 442, 443, 444, 445, 446, 447, 448, 451, 452, 453, 454; millions of dollars, seasonally adjusted, 1992-01 ->). Each category-month row carries the sales level, its share of combined kind-of-business sales, month-on-month / year-on-year / 3-month-annualized momentum, YoY acceleration, a documented 0-100 category momentum score (50% winsorized YoY + 30% 3-month annualized + 20% MoM, min-maxed within each month) with deterministic rank and c1-c4 tiers, trailing-12-month record-high and contraction flags, plus month-level context (combined sales, total YoY, breadth of categories growing YoY, the month's top category, and the nonstore/e-commerce-adjacent share). An online shop joins monthly category sales to its own revenue on (date, category_code) for demand benchmarking; a sales team joins category momentum to pipeline value on date. Method: keyless FRED fredgraph.csv fetches with 0.3s polite pacing; a trailing-36-month verbatim transcription is embedded for offline safety; the transform is pure pandas in map_schema with no imputation (early-window momentum cells are honest nulls). Caveats: shares are shares of the 12 published kinds (which sum to combined retail-trade sales by construction), not of a separately published total; official e-commerce is quarterly (see companion dataset ecommerce_penetration) — the monthly nonstore share is the closest monthly proxy.
US-China Trade Signals (derived)
US-China bilateral goods-trade signals (Census via FRED, monthly 1985 ->): import/export momentum, 30-period change volatility, 3-sigma shock flags, drift forecasts, the bilateral deficit tracker, its 5-year z-score and the export reciprocity gauge. The bilateral lens: where the decoupling story shows up in the numbers. US government data via FRED (free, keyless-by-reuse of the existing FRED key).
Government of Canada Open Data
Population counts for census metropolitan areas, census agglomerations, population centres and rural areas. Statistics Canada table 98-10-0006-01.
US E-commerce Penetration Index (derived)
Quarterly US e-commerce penetration index, 1999Q4 through 2026Q2, from the U.S. Census Bureau's Quarterly E-Commerce Report (FRED series ECOMPCTSA: e-commerce retail sales as a percent of total retail sales, seasonally adjusted). Ships ML-ready derived features: share_qoq_pp and share_yoy_pp (quarterly and year-on-year changes in percentage points), share_accel_yoy_pp (YoY acceleration), and share_vs_5y_median_pp (deviation from the trailing 20-quarter median). Base values read verbatim from FRED's keyless endpoint 2026-09-28; derived columns are this connector's transformation.
US Housing-Vacancy Signals (derived)
Quarterly US housing-vacancy signals from the Census Bureau (FRED RRVRUSQ156N rental vacancy and RHVRUSQ156N homeowner vacancy, 1956-Q1 ->): 1-year momentum, year-on-year change, 4-quarter volatility, 3-sigma anomaly flags, naive-drift forecasts, 5-year z-scores, a tight-rental-market (<7%) flag, and an elevated-homeowner-vacancy (>=2.5%) flag. The tightness lens on US housing — the direct gauge of supply/demand balance — complementing housing-signals (construction, prices, mortgage rates). All rows normalized to country_code USA. Raw series: U.S. Census Bureau (Housing Vacancy Survey) via FRED.
US Retail Sales Signals (derived)
Monthly US retail-sales signals from Census advance retail data (FRED RSXFS nominal retail trade + RRSFS real retail & food services, 1992-01 ->, seasonally adjusted, keyless except the already-provisioned FRED key): year-on-year change, 3-month momentum, 12-month change volatility, 3-sigma anomaly flags, naive-drift forecasts, 5-year spending z-scores, and contraction/strong-expansion flags. The measured monthly household-spending lens between us-consumer-sentiment-signals (stated sentiment) and quarterly real PCE — the nominal/real split shows when spending growth is prices vs volume. All rows normalized to country_code USA. Raw data: Census Bureau via FRED.
US retail seasonal demand index (derived)
Who joins this: an online shop joins monthly revenue on (category_code, calendar_month) and divides out seasonal_index to separate real performance from calendar seasonality; a sales team joins pipeline value on calendar_month for the same deseasonalizing. Method: the 12 Census Monthly Retail Trade Survey kind-of-business series (NAICS 441, 442, 443, 444, 445, 446, 447, 448, 451, 452, 453, 454) in NOT-seasonally-adjusted form via FRED's keyless fredgraph.csv (MRTSSM*USN, millions of dollars, 1992-01 ->). Classical multiplicative decomposition: centered 12-month moving average -> trend; seasonal-irregular ratio = value / trend; seasonal_index for each calendar month = median of that month's ratios across all complete years (median so one anomalous year cannot drag the profile), normalized so the 12 indices average exactly 1.0. si_stability_iqr is the interquartile range of the yearly ratios behind each point — the honest uncertainty. Category context: peak/trough month, seasonal amplitude, December lift, and median Q4 share of annual sales. 144 rows (12 categories x 12 months), refreshed annually once a new complete year lands. Caveats: the profile is a long-run national average per kind of business — structural shifts move it slowly; pandemic-2020 months are in the medians but outvoted by 30+ normal years; the trailing partial year is never used.
Une fois l’allocation gratuite du jour épuisée, les fonctions d’IA peuvent passer par votre propre compte fournisseur.
Conservée uniquement dans cet onglet (effacée à sa fermeture) et envoyée avec chaque requête d’IA. Nos serveurs l’utilisent pour cette requête et ne la stockent ni ne la journalisent jamais.