Research & methodology

Each region has separate GDP and headline inflation models. GDP is real annual growth for the current and next quarter, plus the previous quarter while its official release is pending. Inflation is annual HICP growth for the current month and the following three months. Each horizon has its own direct forecast.

The model groups six monthly lags by indicator, selecting useful groups and individual lags. It uses an unpenalized intercept, training-only scaling and median filling. Unavailable or stale optional predictors are omitted.

Inflation uses a direct headline model with available component indices, weights and cost signals. Component rates are not presented as contributions.

Each run freezes its exact input observations and model state. The website's as-of clock starts when that snapshot finishes being captured. Earlier dates are never inferred from today's revised tables.

Forecast history and daily effects come from actual issued snapshots. Effects separate changed inputs, the calendar roll and model refitting, and must reconcile to the total change.

Historical diagnostics use revised data and a shared tuning sample. They do not establish historical PIT or live performance. RMSE, MAE and indicative ranges remain visible.

A quarterly two-lag BVAR links growth, inflation, activity, unemployment, interest rates and oil. Minnesota shrinkage keeps the model small enough for the available sample.

You can condition several variables together for one to eight quarters. The other paths follow the fitted joint distribution. Oil assumptions are percentage changes; rates, growth and survey balances use points.

Displayed bands include innovation uncertainty at fixed posterior-mean parameters. They exclude parameter uncertainty. These are conditional forecasts, not identified policy or causal effects.

Inputs come from the existing governed macro warehouse and are mapped to exact source keys. Missing service-volume and detailed price series remain visible as gaps; unrelated series are not silently substituted.

Euro-area source composition can differ by series and vintage. Source identities are retained in Data sources. Dated economic reports retain their own cutoffs and do not become current when the website refreshes.

Private licensed stock observations are outside this release. Every delivered bundle refers to a registered development run and an immutable content hash.