How we do this — methodology

The state-of-the-economy dashboard is a fast, consolidated, honest read on where the Portuguese economy stands right now. It is not a GDP forecast and it does not compete with Banco de Portugal or INE on authority — it competes on timeliness, explanation, calibrated risk and openness. This page explains where the data comes from, which models we publish at each point of the quarter, how we score ourselves, and which limits we accept.

The governing principle: every number on the dashboard says what it is and what it is not. When a result is modest we say so; when it is provisional it is labelled provisional; when a model loses to a naive benchmark, the benchmark is what we publish.

Where the data comes from

The dashboard rests on roughly 210 public series, all collected from official or open sources through an open pipeline:

• BPstat (Banco de Portugal) — 117 series: coincident activity, credit, deposits, interest rates, financial stress, wages;
• Eurostat — 57 series: national accounts (the GDP target), industrial production, retail trade, prices, the labour market, and the DG-ECFIN confidence surveys (consumers, industry, services, construction, retail — released at the end of each month, with zero publication lag);
• FRED / ALFRED — 15 series, including the historical vintages of Portuguese GDP that let us score calls against the number as first published;
• Google Trends — 9 search series;
• ECB, REN, ENTSO-E, IMF and others — reference rates, electricity consumption, economic-policy uncertainty.

All transformations (year-on-year, quarter-on-quarter, adjustments) are deterministic and reproducible; no data is hand-edited.

What we publish — the four badges

Every number on the dashboard carries a badge stating its epistemic status. The taxonomy is fixed:

1. Official datum ("Dado oficial") — a value published by an official source (INE, Eurostat, BdP, IEFP), reproduced verbatim. It is not ours; we simply show it with the date it refers to.
2. Reading ("Leitura") — a synthesis of already-published official data (the activity pulse, the labour market, current recession risk, the contributions). It describes the present; it is not a forecast.
3. Indicative estimate ("Estimativa indicativa") — a model estimate of a value not yet published (current-quarter q/q GDP). Never authoritative: the official number is INE's, and our in-quarter estimates are not statistically distinguishable from simple benchmarks at the 5% level.
4. Forecast ("Previsão") — a simulated distribution of future outcomes (the annual outlook). We always publish it as a distribution — read the bands, not the midpoint.

Two compound badges derive from these: Risk scenario (conditional growth quantiles — risk context, not what we expect) and Official data + our calls (the track record, where official GDP outturns sit side by side with the calls we actually published).

The models, by point in the quarter

Quarter-on-quarter GDP is near-white-noise in a normal quarter — the hardest series here to nowcast. Publication is therefore horizon-aware: the model changes with the month of the quarter, and each horizon publishes only what survived evaluation.

Three mandatory honesty notes on this table:

• The M2 ESI member is provisional. It was the sole survivor of a pre-registered bake-off of 19 candidates under a frozen selection rule. The raw Diebold–Mariano p-value is 0.014, but because it was the best of 19, the selection-adjusted p is ≈0.06–0.1 — short of the conventional confidence bar. It is tracked on new quarters under a pre-committed demotion trigger: if the edge does not hold on post-selection scored quarters, the member is dropped and M2 reverts to the two-member combination. That commitment was written down before seeing the results.
• The M3 edge is modest and not significant at 5% (DM p≈0.23; partly carried by the 2021 rebound — dropping 2021, rel 0.87). We never present it as a standalone asset.
• Part of the edge over AR(1) is level-adaptation. Over the last decade, a simple rolling 12-quarter mean captures most of the gain; genuine co-movement skill is concentrated in 2005–2014 and 2021. Our evaluation harness always reports the rolling-mean comparison too.

How we score ourselves

• Pseudo-real-time evaluation. We replay history: for every quarter since 2005 the model sees only what would have been published by that date (the panel is masked by each series' publication lag). A serious, stated caveat: true vintage archives do not exist for most inputs, so we use revised values masked by date — which tends to overstate real-time skill. The partial defence: the DG-ECFIN surveys, which carry much of the in-quarter signal, are essentially never revised — for them the revised and the real-time value coincide.
• First-release scoring. Where vintages exist (ALFRED), we score calls against GDP as it was first published — not against today's revised value, which the model could never have known. The M3 edge survives first-release scoring (rel ≈ 0.81).
• The benchmark is always fair. We compare against an AR(1) using exactly the same information set, never a strawman. And we say when we lose: in the first post-selection scored quarter (2026Q1), all of our models lost to the AR(1) — it is written into the track record.

Intervals: calibrated on our own past misses

The q/q GDP uncertainty bands do not come from model theory (the native standard error was ~2–3× too narrow — our own audit says so). They come from split-conformal prediction, built on the distribution of our own past errors at each horizon:

• at M1/M2 the interval is asymmetric — the model tends to under-predict growth, so the band is wider above the point;
• at M3 the symmetric interval passes coverage and PIT uniformity tests;
• coverage is calibrated for normal quarters: a genuine crisis will pierce the band, and the tile says so.

In the first tracking quarter (2026Q1), the 80% band covered the outturn at all three horizons — at M2 by just 0.007 percentage points; the model's native band would have missed. That is precisely what the calibration is for.

The revision floor

A structural argument that frames everything: the first print of Portuguese GDP is itself an estimate. Between INE's flash estimate and the final figures, revisions have an RMSE of ≈0.3 percentage points (measured across more than a thousand historical vintages; the mean revision is +0.15pp). No nowcast of final GDP can be more authoritative than that — not ours, not anyone's. That is why the dashboard leads with readings of the state of the economy and treats the q/q number as the least authoritative of its nine tiles.

Limits we accept

• Small samples (≈80 usable ex-COVID quarters) mean low statistical power; many honest results are "modest, not significant".
• Without true vintages for most inputs, the evaluation is pseudo-real-time.
• In-quarter reads are indicative, never authoritative.
• The dashboard shrinks whatever fails audit: previously-celebrated models were demoted when their numbers did not reproduce, and the old headline figures were banned from our communication.

Last revision of this page: July 2026. The collection, modelling and evaluation code is open; evaluation artifacts (backtests, calibration, pre-registered bake-offs) are stored timestamped and never overwritten.