Nothing here is a current reading of the economy. The page describes what the section would measure, where the data would come from, how each figure would say what it is and how the prototype was evaluated.

The section we are preparing

A quick, explained reading of where the Portuguese economy stands: activity, the labour market, inflation and recession risk. It would not be a GDP forecast and would not replace Banco de Portugal or INE, which are the official sources.

The principle: every figure would say what it is and what it is not. A modest result would be presented as modest; a provisional one, marked provisional; and when a model lost to a simple benchmark, the benchmark is what would be published.

Where the data came from

The prototype used about 210 public series, all from official or open sources:

  • BPstat (Banco de Portugal): coincident activity, credit, deposits, interest rates, financial stress and wages;
  • Eurostat: national accounts (GDP), industrial production, retail trade, prices, the labour market and the European Commission's confidence surveys (consumers, industry, services, construction and retail, published at the end of each month);
  • FRED / ALFRED: including the historical versions of Portuguese GDP, which let an estimate be scored against the value as it was first published;
  • Google Trends: search series;
  • ECB, REN, ENTSO-E, IMF and others: policy rates, electricity consumption and economic policy uncertainty.

The transformations (year-on-year and quarter-on-quarter changes, adjustments) were automatic and reproducible; no data was edited by hand.

Each figure would say what it is

Each figure would carry one of four labels, with a fixed meaning:

  1. Official data: a value published by an official source (INE, Eurostat, Banco de Portugal, IEFP), reproduced as is, with the date it refers to. It is not ours.
  2. Reading: a summary of official data already published (the activity pulse, the labour market, current recession risk). It describes the present; it is not a forecast.
  3. Indicative estimate: a model estimate of a value not yet released (GDP growth in the current quarter). It is never authoritative: the official number is INE's.
  4. Forecast: a simulated distribution of future outcomes (how the year could end). It would always be published as a distribution: read the intervals, not just the central point.

How the prototype was evaluated

  • As if in real time. We replayed history from the first quarter of 2005 to the fourth of 2025: in each quarter, the model only sees what would have been published by then. The four quarters of 2020 (the pandemic) were left out of the evaluation; those of 2021 are in. One caveat: most series have no archive of past versions, so we used today's revised values, cut off by their publication date. That tends to make the model look better than it would have been in real time. The confidence surveys, which carry much of the signal, are hardly ever revised.
  • Against the first release. Where past versions exist, we scored each estimate against GDP as it was first published, not against today's value, which the model could never have known.
  • Against a fair benchmark. The benchmark is an AR(1): a simple model that estimates the quarter from the previous one, with the data available on the same date.
  • Intervals from our own errors. The intervals do not come from the model's theory, which made them two to three times too narrow, but from the distribution of the errors the prototype made in the past at each point of the quarter.

A modest edge in the internal test, which did not hold in the first quarter of 2026

These results come from an internal test run on the data available up to July 2026, over the 80 quarters evaluated (2005 to 2025, without 2020). They are not published forecasts and would have to be redone before any publication. The relative error compares the model's error with the AR(1)'s: 0.90 means an error about 10% smaller.

  • In the first month of the quarter, a combination of two models had a relative error of about 0.90.
  • In the second month, the same combination with an economic sentiment indicator had about 0.81, a provisional result. That indicator was the only survivor of a contest with 19 candidates and a selection rule fixed in advance. The raw p-value of the Diebold–Mariano test was 0.014, but because it was the best of 19, the selection-adjusted p-value is about 0.06 to 0.1: above 0.05, so it is not statistically significant at the conventional level. It was written down before the results that if the advantage did not hold in the following quarters, the indicator would go.
  • In the third month, the relative error was about 0.78, a modest advantage that is not significant at 5% (p-value about 0.23) and rests partly on the 2021 rebound: without 2021, the advantage shrinks to a relative error of about 0.9.
  • Part of the advantage is just tracking the level. Over the last decade, a simple moving average of the last 12 quarters captures most of the gain over the AR(1).
  • In the first quarter scored after the selection (2026 Q1), every model lost to the AR(1). The 80% interval covered the observed value at all three points of the quarter; in the second month, only just.

Quarterly GDP would not be the headline figure

The first GDP release is itself an estimate. Between the first release and the latest values, revisions to Portuguese GDP have a typical error (root mean squared error) of about 0.3 percentage points, which we computed from the historical versions kept in ALFRED (Federal Reserve Bank of St. Louis). No estimate of final GDP can be more certain than that. So if the section were published, it would lead with readings of official data and treat the current-quarter GDP estimate as its least authoritative figure.

The limits we accept

  • There are few quarters to evaluate (80, without those of 2020), so many honest results are "modest, not significant".
  • Without past versions of most data, the evaluation only approximates real time.
  • Estimates for the current quarter would be indicative, never authoritative.
  • What does not survive evaluation goes: models that looked good were withdrawn when their numbers did not reproduce.

Glossary

  • AR(1): a simple model that estimates the next quarter from the previous one. It is the benchmark any model has to beat.
  • Relative error: the model's root mean squared error (RMSE) divided by the AR(1)'s. Below 1, the model errs less; 0.90 means an error about 10% smaller.
  • Diebold–Mariano test and p-value: a test of whether the difference in errors between two models could be chance. A p-value below 0.05 is the conventional threshold for saying it is not.
  • First, second and third month: the point in the quarter at which the estimate is made. The later it is, the more official data is already out.
  • Economic sentiment indicator: the European Commission's index that summarises the confidence surveys.
  • 80% interval: the range in which, in the normal quarters of the past, the official value fell eight times in ten.

This page describes the prototype tested up to July 2026, scored up to 2026 Q1. The section's code is not public.