# Central Bank of Brazil series: SGS, Focus and PTAX without the traps
Sidney Bissoli
2026-08-31

- [The SGS: everything is a series
  number](#the-sgs-everything-is-a-series-number)
- [Three API limits, measured](#three-api-limits-measured)
- [The trap that raises no error: variation on a series that already is
  variation](#the-trap-that-raises-no-error-variation-on-a-series-that-already-is-variation)
- [Focus: the policy rate’s axis is not a
  horizon](#focus-the-policy-rates-axis-is-not-a-horizon)
- [PTAX: not every quote is compiled by the Central
  Bank](#ptax-not-every-quote-is-compiled-by-the-central-bank)
- [The licence asks for things](#the-licence-asks-for-things)
- [Where an AI assistant fits](#where-an-ai-assistant-fits)

*Published 31 August 2026. Every behaviour and value described here was
measured live against the Central Bank APIs on the publication date;
nothing is inferred from documentation.*

The Central Bank of Brazil publishes three different things, by three
different routes, and beginners tend to treat them as one:

- the **SGS** (time series management system) — thousands of numbered
  series: the policy rate, inflation, exchange rates, credit, monetary
  aggregates;
- **Focus** — the weekly market-expectations survey, which is not a time
  series but a matrix of indicator × horizon × statistic;
- **PTAX** — the official closing exchange rates, served by a separate
  API.

## The SGS: everything is a series number

You do not query “inflation”. You query **series 433**. Not “the
dollar”; **series 1**. The policy rate target is **432**. Finding the
code is half the work.

Real response on 2026-08-31 — monthly IPCA inflation:

``` json
[{"data":"01/05/2026","valor":"0.58"},
 {"data":"01/06/2026","valor":"0.16"},
 {"data":"01/07/2026","valor":"0.07"}]
```

## Three API limits, measured

**1. A window longer than 10 years on a daily series is refused with
HTTP 406.**

    .../bcdata.sgs.1/dados?formato=json&dataInicial=01/01/2011&dataFinal=31/12/2025
    → HTTP 406

A 3-year window on the same series returns 755 observations.

**2. An open window is refused too**, for the same reason: with no end
date the API assumes today, and series 1 begins in 1984.

**3. `ultimos/N` is capped at 20**, in every periodicity — and above it
the request *fails* with HTTP 400 rather than returning 20.

Put the three together and a 15-year daily series has to be sliced,
fetched and merged without duplicating the seams. In practice slice at
**3 years**, not the permitted 10: a decade-long daily window costs
10–20 seconds upstream and can be cut off around 30.

## The trap that raises no error: variation on a series that already is variation

Monthly IPCA **already is a percentage change**. Summing twelve monthly
values to get the annual figure is wrong, because inflation compounds:

    compounded = [ Π (1 + vᵢ/100) − 1 ] × 100

With the three real months above, the sum gives 0.8100 and compounding
gives 0.8114 — the sum understates. Small over three months; not small
over twelve, or in any period with high inflation.

The rule: **before computing a variation, ask whether the series is a
level or a rate.** Levels (the exchange rate, credit stock, GDP) admit
endpoint-to-endpoint change. Series that already are period-on-period
rates require compounding. The same applies to resampling: turning
monthly inflation into annual is geometric composition, not an average.

## Focus: the policy rate’s axis is not a horizon

For most indicators the expectation has a **horizon** — monthly,
quarterly, annual, or inflation accumulated over 12 and 24 months. Each
response carries mean, median, standard deviation, minimum, maximum and
the number of respondents, plus a **Top 5** variant for the best recent
forecasters.

The policy rate is different: its axis is the **monetary policy
committee meeting** (in the form `R1/2026`), not a time horizon.
Treating it as one more horizon is the most common conceptual error
here.

Another asymmetry: **the set of indicators changes with the scope** — 9
in the monthly scope, 26 in the annual. There is no single “list of
Focus indicators”; there is the list *of that scope*.

## PTAX: not every quote is compiled by the Central Bank

The Central Bank compiles the US dollar PTAX. **Cross-currency parities
are not**: they come from an information agency (Refinitiv) and are
redistributed by the Central Bank. The difference is legal and
methodological, not cosmetic — and the Bank’s own liability disclaimer
must be passed through verbatim.

## The licence asks for things

Central Bank data is released under the **Open Data Commons Open
Database License (ODbL) v1.0**. That is **not** CC0, not CC BY and not
public domain. The ODbL carries three obligations that survive
redistribution: **attribution**, **share-alike on derived databases**,
and an **anti-DRM** clause.

Checked against the source on 2026-08-13: 4,259 of the 4,260 datasets in
the Bank’s open-data portal declare `license_id: "odc-odbl"`.

## Where an AI assistant fits

Asking a model *“what is the policy rate today?”* returns a number from
training: possibly right, possibly three committee meetings out of date,
with no source. Plugging in the SGS buys **provenance** — the series,
the period, the canonical URL, the real extraction instant and the
licence — and moves the traps above to the server side.

That is what [`bcb-br-mcp`](https://github.com/SidneyBissoli/bcb-br-mcp)
does: 15 tools over SGS, Focus and PTAX, with a curated catalogue of 135
series verified against the source, for clients that speak
[MCP](https://modelcontextprotocol.io).

With or without a tool, the two questions that avoid most errors are the
same: **is this series a level or a rate?** and **who compiled this
number?**

------------------------------------------------------------------------

*Server: [`bcb-br-mcp`](https://github.com/SidneyBissoli/bcb-br-mcp) ·
hosted endpoint at `https://bcb.sidneybissoli.com/mcp` · code under MIT.
Central Bank data under ODbL v1.0 —
https://opendatacommons.org/licenses/odbl/1-0/.*