How the nowcasts work
Official GDP arrives weeks after a quarter ends. Monthly data on production, spending, trade, jobs and business sentiment arrive much sooner. A nowcast reads those monthly signals every day and translates them into an estimate of this quarter's GDP growth before the official number is out.
The daily pipeline
Every morning at 07:00 UTC the same five steps run for each economy. Pick a step.
We download the latest monthly figures from official sources: statistics offices, central banks, Eurostat, the OECD and the World Bank. Each economy uses the indicators that best track its activity, such as factory output, retail sales, payrolls, exports and business surveys.
Series are pulled from source APIs (FRED/ALFRED, Eurostat JSON-stat, OECD SDMX, World Bank GEM and WDI, ONS, NBS releases). Each is transformed to stationarity (monthly log-differences for activity and prices; levels or differences for surveys and rates), standardised, and screened for length: series with fewer than 24 observations are excluded from estimation. For the US, 30+ series are read from ALFRED vintages, so every historical nowcast uses only data as published on that day.
Data arrive on different days and cover different months: surveys for September are out in early October, industrial production a few weeks later, trade later still. The model works with whatever is known today and fills the gaps with its best estimate, which is then replaced as real numbers arrive.
The panel has a ragged edge: missing observations at the end of the sample that differ by series. Missing values are handled inside the state-space model rather than by ad-hoc extrapolation. Interior gaps (e.g. series suspended during a government shutdown) are filled the same way. Publication lags are taken from real release dates where available, otherwise from each series' typical lag (pseudo real-time).
Most indicators move together because they reflect the same underlying business cycle. The model distils dozens of series into a couple of summary "factors" that capture that shared movement, which filters out the noise of any single release.
A two-step dynamic factor model (Doz, Giannone and Reichlin, 2011). Step one: principal components on the standardised panel (EM algorithm for missing data) give initial factors and loadings. Step two: factors follow a VAR(p), typically 2 factors and 2 lags; the model is cast in state space and the Kalman filter and Rauch–Tung–Striebel smoother re-estimate the factors with the ragged edge. Parameters are re-estimated once per quarter and held fixed within it, so that day-to-day changes are attributable to data, not re-estimation.
The factors are then linked to GDP using the historical relationship between them. For the US and the euro area we also build GDP up from its pieces (consumption, investment, inventories, government and trade) and average the two views, which is both more accurate and easier to explain.
Top-down: quarterly GDP growth is regressed on the factors aggregated to quarterly frequency with Mariano–Murasawa weights (1, 2, 3, 2, 1)/9, which map monthly growth rates of a flow into quarter-on-quarter growth of its quarterly average. Bottom-up (US, euro area): each expenditure component is bridged from accounting proxies built from its own source data (e.g. real goods and services spending for consumption; shipments and construction for investment; advance trade and inventory reports for net exports and the inventory swing), with months still missing filled sequentially from helper series and the factors. Components are combined with their nominal shares; any gap to the top-down level is shown as a residual. Headline: the simple average of top-down and bottom-up for the US and euro area; top-down elsewhere. Output is expressed as quarter-on-quarter growth at a seasonally adjusted annual rate (SAAR).
Each economy's nowcast is published with a range showing how wrong it has typically been at this point in the quarter. The World and regional figures add the economies together, weighted by the size of each economy.
For each economy the day's nowcast, its 70% range, the component and release attributions and the release calendar are written to the site. Aggregates combine economy nowcasts with GDP weights in current US dollars (World Bank WDI, latest year); see World and regions. A run that fails for one economy does not block the others: that economy keeps its last good nowcast and is flagged.
Models by economy
Three economies have dedicated models built around their national accounts; the others use a common template that will be upgraded over time.
| Economy | Model | Target (truth) | Main inputs | History |
|---|---|---|---|---|
| United States | Dedicated: factor model + bottom-up components | BEA advance estimate | Payrolls, industrial production, retail sales, real spending, housing, durable goods, advance trade and inventories, regional Fed and ISM-type surveys | Real time (ALFRED vintages) |
| Euro area | Dedicated: factor model + bottom-up components | Eurostat preliminary flash | Industrial production, retail trade, construction, trade, unemployment, European Commission surveys | Pseudo real time |
| China | Dedicated factor model | NBS q/q growth (seasonally adjusted) | Industrial production, retail sales, services output, trade, official PMIs, market data | Pseudo real time |
| Japan, UK, India, Brazil, Mexico, Canada, South Korea, Australia | Standard template factor model | First official estimate (via OECD) | Industrial production, trade, equity prices (World Bank GEM), UK monthly GDP (ONS), global drivers | Pseudo real time |
Pseudo real time: historical tests use the latest data, but only what would have been published by each date given usual release lags. This ignores revisions, so it slightly flatters accuracy.
What moved the nowcast
When a new figure comes out, the model compares it with what it expected. A surprise in an indicator that matters a lot for GDP moves the nowcast; a figure in line with expectations barely moves it. Each economy page shows these moves by data release, and for the US and euro area also by GDP component.
With parameters fixed within the quarter, the change in the nowcast between two days is attributed to the data that arrived in between, in the spirit of the news decomposition of Bańbura and Modugno (2014): roughly, impact ≈ weight × (actual − model expectation), where the weight reflects how informative that series is for current-quarter GDP given everything else already known. "Expected" in the release tables is the model's own forecast of that print before it was published, not a market consensus. Changes at quarter turns, when parameters are re-estimated, are not attributed to releases.
Uncertainty
Early in a quarter there is little hard data, so nowcasts are rough; they sharpen as the quarter fills in. The shaded range on each chart covers 70% of past errors made at the same point before the official release: roughly two times out of three, the official number has landed inside it.
Bands are empirical: from the backtest, absolute errors of the headline nowcast versus the target are pooled by days to the official release and the 70th percentile half-width is applied symmetrically around today's nowcast. No distributional assumption is made. For aggregates, economy half-widths are combined in quadrature with GDP weights, which assumes independent errors across economies; errors are in fact positively correlated in global shocks, so aggregate bands are a lower bound.
World and regions
The World and regional nowcasts add up economy nowcasts weighted by GDP in current US dollars (World Bank, latest year). Market-rate weights are used because they match what moves global markets and trade; purchasing-power weights would give more weight to emerging economies and a faster headline.
Not every economy is modelled yet. The remainder of each region is held at its latest annual growth rate, so it does not move from day to day. Coverage, shown on every page, is the share of the region that is actually modelled. Regions below 50% coverage are labelled partial.
Aggregate growth = Σ wᵢ gᵢ + w_rest g_rest, with wᵢ the US-dollar GDP shares within the aggregate, gᵢ each economy's nowcast (or its official print once released) and g_rest the latest annual real growth of the uncovered remainder, derived from World Bank aggregate and country growth rates. Regions follow World Bank definitions (Europe & Central Asia, East Asia & Pacific, South Asia, Latin America & Caribbean; income groups), except North America, which here is the US, Canada and Mexico. The euro area is treated as one economy and appears on the map through its 21 members.
Track record and accuracy grades
Each model is replayed over past quarters as if run live every day, and its final nowcast is compared with official GDP: the first (advance) estimate for the US, where real-time vintages exist, and the currently published figures elsewhere.
Raw error alone would be misleading: India's GDP growth swings about six times more than US growth, so any India model has larger errors in percentage points. The accuracy score corrects for this. It is the share of quarter-to-quarter GDP swings that the nowcast got right. 0% means no better than always guessing average growth; 100% would be perfect. Each score maps to a grade, like a credit rating.
Accuracy is the out-of-sample R² of the final nowcast against the official print: 1 − RMSE² / σ², where σ is the standard deviation of official SAAR growth over the scored quarters (2020Q1–2021Q1 excluded), floored at zero. It measures skill relative to the unconditional mean and is comparable across economies with very different volatility. We also report the RMSE of a naive AR(1) as a second benchmark; a model can beat the AR(1) yet score low accuracy when GDP is persistent but hard to predict (China) or very noisy (India).
| Economy | Model | Quarters | Error (RMSE) | GDP volatility (σ) | Naive AR(1) | Accuracy | Grade |
|---|---|---|---|---|---|---|---|
| Loading… | |||||||
Errors in percentage points of annualised growth. Grade scale: A+ ≥70% · A ≥55% · A− ≥45% · B+ ≥35% · B ≥25% · B− ≥15% · C+ ≥5% · C <5%.
Versions and continuous improvement
Every model carries a version number. Any change to its inputs or specification creates a new version, is backtested over the same quarters, and goes live only if its error falls. Each economy page lists its version history with the error and grade of every release, so improvements are traceable.
Limitations
- A nowcast estimates the first official print, which is itself revised later. It is not a forecast of the final number.
- Standard-template economies rely on inputs that arrive about two months late, so their nowcasts move little until late in the quarter and carry wider ranges.
- Large, unusual shocks (pandemics, wars, strikes, tariffs) can break historical relationships between indicators and GDP.
- China's official GDP is unusually smooth; its nowcast targets the official print rather than an independent activity measure.
- The unmodelled part of each aggregate is held at annual growth and cannot react to news.
- Upstream sources occasionally change formats or are unavailable; the affected economy then keeps its last nowcast until data resume.
Glossary
Nowcast
An estimate of the present or very recent past, here current-quarter GDP growth, made before official data are published.
SAAR
Seasonally adjusted annual rate: quarter-on-quarter growth compounded over four quarters. 0.5% q/q ≈ 2.0% SAAR. US convention, applied to all economies for comparability.
pp (percentage points)
The arithmetic difference between two percentages: a nowcast moving from 2.0% to 2.3% moved by 0.3 pp.
Dynamic factor model (DFM)
A statistical model that explains many time series with a few common, unobserved drivers ("factors") that evolve over time.
Bridge equation
A regression linking quarterly GDP (or a component) to indicators aggregated from monthly to quarterly frequency.
Kalman filter and smoother
Recursive algorithms that estimate unobserved states, here the factors, from noisy and incomplete data, updating as each observation arrives.
Ragged edge
The uneven end of a dataset when series are published on different dates and cover different months.
Vintage
A dataset as it was published on a given date, before later revisions.
AR(1)
Autoregressive model of order one: predicts growth from the previous quarter's growth only. Used as a naive benchmark.
Accuracy (out-of-sample R²)
The share of the variation in official GDP growth explained by the final nowcast, out of sample. 0% = no better than the average; 100% = perfect.
RMSE
Root mean squared error: the typical size of errors, penalising large misses more.
References
- Giannone, D., Reichlin, L. and Small, D. (2008). Nowcasting: the real-time informational content of macroeconomic data. Journal of Monetary Economics, 55(4).
- Doz, C., Giannone, D. and Reichlin, L. (2011). A two-step estimator for large approximate dynamic factor models based on Kalman filtering. Journal of Econometrics, 164(1).
- Mariano, R. and Murasawa, Y. (2003). A new coincident index of business cycles based on monthly and quarterly series. Journal of Applied Econometrics, 18(4).
- Bańbura, M. and Modugno, M. (2014). Maximum likelihood estimation of factor models on datasets with arbitrary pattern of missing data. Journal of Applied Econometrics, 29(1).
- Higgins, P. (2014). GDPNow: A model for GDP "nowcasting". Federal Reserve Bank of Atlanta Working Paper 2014-7.
worldgdp.org is independent and not affiliated with any central bank, statistics office or the Federal Reserve Bank of Atlanta. Source data remain the property of their publishers.