Methodology, as questions you'd actually ask

Inflation, GDP, unemployment, policy rates, and bond yields come from World Bank, IMF World Economic Outlook, FRED, and national statistical offices (BLS, ONS, Statistics Canada, Eurostat HICP via FRED, Statistics Bureau of Japan and INEGI via the IMF CPI monthly index). Sovereign 10-year yields are shown only for Australia, Canada, Switzerland, Germany, the United Kingdom, India, Japan, and the United States, and only from a current comparable FRED series. Germany’s figure is the Bund. Policy rates are shown for Germany, the United Kingdom, and the United States. Every other country says there is no comparable 10-year. Import tariffs are the World Bank weighted mean, labeled with the year of that print. A current-year figure stored as a year only is marked as an estimate. A model never overwrites these prints. Hover and popup cards show the office, the definition, and the as-of date, with a sparkline of the last official observations. Where we have a live monthly CPI YoY (BLS, ONS, Eurostat HICP, StatCan, IBGE IPCA, or IMF CPI monthly index from the national office) that print sits next to the IMF annual average, not instead of it. Monthly unemployment for OECD/BLS countries is the same: a live rate when the office series is current, otherwise the annual World Bank/IMF figure and a label that says so. Quarterly real GDP growth is BEA, Eurostat, ONS, StatCan, IBGE and the other national accounts via FRED, shown next to World Bank/IMF annual GDP in current dollars — never mixed into one number. The US page also shows BLS core CPI and BEA core PCE from the same FRED pipe. OECD MEI CPI reprints are not used; they lag the office print. Argentina has no current INDEC monthly feed we can cite, so it stays on the annual print. Each country also has a public series page (for example /country/USA) with history and the next dated office release. Watching SPY, EWZ, or a country puts the same office CPI YoY, unemployment, and next print into the digest when that as-of is new (or when a verified event hits the watchlist), and into a real-time alert when the as-of date changes — office, as-of, and URL, never a GDELT Goldstein hurt/benefit. A quiet day is not an email. Official prints refresh on their own job, separate from Twelve Data quotes.

Country hover cards name the CIA World Factbook archive for a leader, and UN Comtrade (HS2 public preview) for main export, main import, and the largest export partner, only when that source is on file. A leader with no source URL is marked unsourced. A country with no trade line says so. There is no model-written investor note. US crop yields are USDA NASS with the crop year on the figure; other countries show the crop name without an inferred yield. Mines are USGS MRDS producer sites. Rivers are Natural Earth 50m centerlines. Submarine cables are TeleGeography, named, unclipped. Cloud points are Microsoft Azure first-party regions. Shipping chokepoints are labeled as unofficial geometry until IMO, AIS, or Natural Earth publishes a named set. Petroleum field coordinates are not in MRDS; they stay empty until EIA publishes a citable field set.

Company-to-place operators on mines and plants must cite a 10-K/20-F Item 2 (or another named annual report). Product and plant nodes on a ticker schematic use those same filings; a vendor industry string is not a product map, and the caption says so when none is on file. Listed commodity ETFs are labeled proxies, not local producers. A country's listed market fund (SPY for the United States, EWT for Taiwan, MCHI for China) is a labeled vehicle, not a 10-K. GDELT is a news code: it does not mark that fund as hurt or benefit. The country hover card is official prints and the country profile only. A profile line is named with Factbook or Comtrade only when that source is on file. News pins stay on the globe, not mixed into CPI.

Globe pins start from GDELT (machine-coded news), GDACS/NOAA disasters, and USGS earthquakes. GDELT's ActionGeo is often the dateline or publisher city, not the event. We prefer a named place in the article path when it disagrees with ActionGeo. If the URL, the geocode, and the actors conflict and nothing in the path resolves them, we do not pin and we do not email. GDELT also codes sports and tourism metaphors ("battle it out") as fighting. Event rows pass three layers: a high-precision check that the article path is the same kind of event as the CAMEO code, a write-time model that compares the URL to the coded type and parties, and the same path check again at serve so a stale cache cannot wait for the next refresh. The globe request does not fetch article HTML. Watchlist alerts also require the event's own date (not merely ingest time) to fall in the last three days. Conflict density on the choropleth is UCDP, human-reviewed, not GDELT counts.

After you click a country, the column may show what prices did the last times events were tagged to that country: the country fund versus the S&P 500 over the next five trading days, median across those event days, using the last two years of daily bars we store. That is a base rate, not a cluster of "similar" stories and not a forecast. A sample under five events is named as too thin and the percent is withheld. Overnight tests of "elevated readings versus next month's returns" stay in the research database; if that long-run test did not hold up, the dock says so in English. Same-day co-movement with the S&P 500 on a ticker page is correlation, not this event path. A model never overwrites BLS, BEA, or OFAC.

Prices are refreshed on a scheduled job, not a live tape. The popup shows the quote as-of timestamp when we have one. That can be up to a couple of hours old.