How the data is sourced, verified and refreshed

1 · Primary-source policy

We do not re-sell aggregated feeds. Each indicator pulls from the primary issuer — the national statistics office, the central bank, the finance ministry, or a multilateral of record (IMF, OECD, UN Comtrade, BIS). Where a central bank and a statistics office both publish a series, we take the one that is the official release for that concept (e.g. CPI from the statistics office, the policy rate from the central bank).

Every series carries its provenance on the card: the issuing institution, the specific table or dataset code, the base year, and the source URL. When a value is derived (for example a real series deflated by CPI, or a trade balance built from exports minus imports), the card says so and names its inputs.

2 · Sources by country

The live region is Latin America & the Caribbean plus the United States as global benchmark. Primary sources per market:

Economy	Primary source(s)	Domain
Argentina	INDEC · BCRA · Secretaría de Hacienda / MECON	datos.gob.ar · bcra.gob.ar
Bolivia	INE · Banco Central de Bolivia · MEFP	ine.gob.bo
Brazil	Banco Central do Brasil (SGS, Focus) · IBGE (SIDRA) · MDIC (Comex) · OECD	bcb.gov.br · ibge.gov.br
Chile	Banco Central de Chile (Base de Datos Estadísticos)	bcentral.cl
Colombia	DANE · Banco de la República (totoro / SUAMECA) · DIAN · MinHacienda · Fedesarrollo	dane.gov.co · banrep.gov.co
Costa Rica	Banco Central de Costa Rica	bccr.fi.cr
Dominican Rep.	Banco Central de la República Dominicana · ONE	bancentral.gov.do
Ecuador	Banco Central del Ecuador (IEM, IMAEc) · INEC	bce.fin.ec
Guatemala	Banco de Guatemala (Banguat) · MINFIN · SECMCA	banguat.gob.gt
Honduras	Banco Central de Honduras · SEFIN	bch.hn
Mexico	Banco de México (SIE) · INEGI · SHCP	banxico.org.mx · inegi.org.mx
Panama	INEC Panamá · Superintendencia de Bancos	inec.gob.pa
Peru	Banco Central de Reserva del Perú (BCRPData)	bcrp.gob.pe
Paraguay	Banco Central del Paraguay · MEF · INE	bcp.gov.py
El Salvador	Banco Central de Reserva de El Salvador	bcr.gob.sv
Uruguay	INE · Banco Central del Uruguay	bcu.gub.uy · ine.gub.uy
Venezuela	Banco Central de Venezuela · OPEC	bcv.org.ve
United States	BEA (NIPA/ITA) · FRED / St. Louis Fed · BLS · US Treasury · OECD	bea.gov · stlouisfed.org
Barbados	Central Bank of Barbados · Barbados Statistical Service IMF NSDP	imf.org
Belize	Statistical Institute of Belize · Central Bank of Belize IMF NSDP	imf.org
Jamaica	STATIN · Bank of Jamaica IMF e-GDDS	imf.org
Suriname	Centrale Bank van Suriname · ABS · Ministry of Finance IMF e-GDDS	imf.org
Trinidad & Tobago	Central Statistical Office · Central Bank of T&T IMF NSDP	imf.org

Cross-country projections and debt ratios also draw on the IMF World Economic Outlook; commodity prices on the relevant exchange / agency feeds. Caribbean markets currently carry quarterly national-accounts and external coverage via the IMF standardized dissemination standards, with the monthly dashboard in progress.

3 · Refresh cycle

Pipelines run on scheduled GitHub Actions — one workflow per country — that re-fetch from source, re-verify, and commit the refreshed panel daily, aligned to each release calendar. There is no manual data entry in the loop.

• Self-healing: a source outage does not blank the series — the last good value is carried forward and the run flags the gap, so a single failed fetch never corrupts the panel.
• Deterministic aggregates: cross-country indices (heatmaps, nowcast inputs, the MCP catalog) are re-projected from the committed country files on every run, so they never drift from the underlying data.
• Detected dates: each print records both its reference period and the date it was ingested, surfaced in the terminal's "Latest Updates".

4 · Seasonal adjustment

Many activity, trade and monetary series are seasonal. We are explicit about which adjustment a chart shows, tagged per series:

• Official SA — where the issuing institution publishes a seasonally adjusted series (e.g. INEGI's IGAE SA, DANE's ISE Cuadro 2, the BCCh Imacec SA), that official series is used and labelled as such. The non-adjusted original is kept alongside it.
• Own SA (X-13ARIMA-SEATS) — where the source publishes only the original (NSA) series, we compute the adjustment in-house with the US Census Bureau's X-13ARIMA-SEATS program (v1.1 build 62), the same engine national statistical offices use. It is tagged own_sa and recomputed from the NSA on every run (never chain-linked). Each series gets automatic ARIMA model selection, automatic outlier detection (additive, level-shift and transitory), and trading-day and moving-holiday (Easter) regressors — the last of these matters across Latin America, where Semana Santa migrates between March and April.
• Fallback — where X-13 cannot run (a series with interior gaps, or shorter than three years), a simpler in-house ratio-to-moving-average adjustment (“X-11-lite”) is used instead. Every in-house SA series records which engine produced it.

Adopting X-13 mattered most for the pandemic. A ratio-to-moving-average filter has no outlier detection, so the 2020 collapse was absorbed into the March and April seasonal factors permanently and kept distorting later readings. Re-estimating the in-house series with X-13 identified 337 outliers across 84 of them — 66 series carried a 2020–21 shock — and a significant Easter effect in 32.

Two caveats we would rather state than bury. First, an X-13-adjusted series does not preserve the annual mean of the original the way a pure seasonal filter does, because trading-day and holiday effects are removed as well; that is intended, not drift. Second, X-13 reports its own quality verdict, and for a handful of volatile series (mining, fishing, some construction) it finds no identifiable stable seasonality at all. We publish those adjustments but flag them, rather than presenting a seasonal adjustment the diagnostics do not support.

Both the seasonally adjusted and original series are retained wherever the source allows, so you can see the adjustment rather than take it on faith.

5 · Accounting identities

National-accounts and external blocks are verified to close, not just displayed:

• Balance of payments — BPM6. Current, capital and financial accounts follow the IMF BPM6 framework; the current-account and financial-account contribution charts are built from the component series, and the identity is checked against the published aggregate.
• National accounts. Quarterly GDP is carried on both the expenditure side (C, G, gross fixed capital formation, change in inventories, exports, imports) and the production side (value added by sector), with the expenditure identity verified to close on the nominal series. Real chain-linked series can be mildly non-additive; where a source documents this, we do not force the real identity.
• Fiscal. Where sub-aggregates are published (revenue by tax, spending by function), the components are checked to sum to the headline before wiring.

6 · Unit & convention notes

• Local currency is labelled with each country's ISO-4217 code (ARS, CLP, BRL, …); dollarized economies (Ecuador, El Salvador, Panama) are in USD.
• Percent series are stored as the percentage number (a 4.1% print is 4.1, not 0.041), so cross-country comparisons line up.
• Index series carry their base year on the card (e.g. base 2018 = 100); re-based vintages are not spliced across a break without a continuity check on the overlap.
• Real series deflated in-house name their deflator and base period.
• Fiscal sign convention is stated per source (some publish deficit as positive financing need, others as a negative balance); the card follows the issuer.

7 · Policy models

The Signals › Policy models view replicates the macro model each of four central banks runs for a policy round — an IS curve, a Phillips curve and a forward-looking reaction function, quarterly — fits it on this dashboard's own data and shows what the bank's own framework implies today. Each bank gets the replication of its model, not one engine stretched across four economies: a small semi-structural block for Brazil (BCB WP 1 / SAMBA lineage), the 4GM for Colombia (BanRep, Borradores 1106), the MPT for Peru (BCRP, DT-011-2022) and a semi-structural block for Mexico (Banxico, Informe Trimestral). Every panel re-estimates in the browser.

• Two objects, never mixed. The rule prescribes a level. A second, deliberately small model — an ordered logit fitted on that committee's own decision record — says how much of the distance to it the board travels at one sitting, as a distribution over the bank's own 25bp grid. Both are shown, labelled, and scored: a rolling-origin backtest re-estimates everything on data ending before each past meeting, and reports the hit rate against the naive benchmark that record makes relevant, because "74% correct" means nothing on a board that held 66% of the time.
• Estimated where the data supports it. Equations are fitted by OLS with Newey–West standard errors, excluding the COVID quarters. Each coefficient is gated separately on sign, plausible size and significance; one that fails ships the published calibration instead, and the panel labels every coefficient estimated or calibrated with the reason.
• The published parameters are the default. Target, band, neutral real rate, β_π, β_y, ρ, the meeting count and the decision grid are shown with their source, and each panel opens on the bank's own convention. Switching to the coefficients estimated here is one click, and the panel says which is in force.
• Smoothing is quarterly, meetings are not. The published inertia parameters are quarterly. The per-meeting value is ρ^{4/meetings per year} — √ρ for the Copom, BanRep and Banxico (8 a year), ρ^⅓ for the BCRP (monthly). Using the quarterly value per meeting would move about three times too fast.
• The decomposition always adds up. Brazil's block adds expected inflation and carries the response above one in β_π; the other three add the target and carry the total response. Each panel prints the term its own rule adds, so the column sums to i*.
• Every panel cites its sources. The documents the structure comes from, the feeds it is fitted on with their series codes, and one citation per published parameter — printed, not hidden behind a hover. Where a bank's own calibration could not be transcribed, the panel says so: the structure is the published one, the coefficients are estimated here.
• Failure is reported. A forward-looking rule on a strong interest-rate channel need not have a stable solution; where it does not, the view says so and falls back to survey expectations rather than drawing an explosive path.

8 · AI-readable by design

The same verified catalog is exposed to AI assistants over an open Model Context Protocol (MCP) connector, so the numbers an assistant reasons over are exactly the ones in the terminal — same sources, same units, same freshness. See the AI assistant section for how it works.