Skip to main content

Nowcasting (GDPNow / Nowcast)

Updated Jun 24, 2026 at 2:35pm

Research Draft Medium 1,345 words

Nowcasting is the practice of estimating the current state of the economy — most famously real GDP growth for the quarter that is still in progress — before the official statistic exists. Because GDP is published with a long lag (the BEA's "advance" estimate for a quarter lands roughly a month after the quarter ends, and final revisions come months later), a nowcast fills the information gap by continuously translating the higher-frequency data that does arrive (jobs, retail sales, trade, construction, ISM surveys) into a running estimate of where GDP is heading. The core tension: nowcasts give you a timely read, but they are mechanical extrapolations that swing sharply with each data release and have no judgment about one-off distortions. The two reference implementations are the Atlanta Fed's GDPNow and the New York Fed Staff Nowcast.

How it's calculated / formed

Nowcasting rests on the insight that monthly and weekly indicators co-move because they share a few underlying drivers. Two model families dominate:

  • Dynamic factor models (DFM). The intellectual foundation is Giannone, Reichlin & Small's 2008 Journal of Monetary Economics paper, which formalized extracting a small number of unobserved "factors" (the business-cycle signal) from a large, ragged, real-time dataset and discarding the idiosyncratic noise. The model is cast in state-space form so it can handle an unbalanced panel — series that report on different days and at different lags. Each new release is decomposed into "news" (the surprise versus what the model expected), and the nowcast revision is a weighted sum of those news components. The NY Fed Staff Nowcast is a pure DFM of this lineage; it updates weekly and explicitly attributes each week's change to the data that moved it.
  • GDPNow's hybrid bridge approach. Per the Atlanta Fed, GDPNow aggregates statistical forecasts of 13 subcomponents of GDP (consumption, investment, net exports, government, inventories, etc.) using "bridge equations" that map monthly source data to each subcomponent, with a dynamic factor model used to fill in monthly source data that has not yet been released. The subcomponent nowcasts are then summed up to a headline real-GDP-growth figure, mimicking the BEA's own accounting identity.

Two defining traits of GDPNow: it begins nowcasting a quarter roughly 90 days before the advance estimate is released, updates about 6–7 times per month as data arrives, and makes no subjective adjustments — the number is purely the model's output (Atlanta Fed).

How it's used in practice

Practitioners treat nowcasts as a real-time growth thermometer rather than a precise forecast:

  • Tracking the data flow. Because the nowcast jumps on each release, it tells you which report moved the growth picture and by how much — the NY Fed publishes a "news" decomposition that does exactly this. That makes it a fast way to read the economy's pulse between GDP prints.
  • Cross-checking consensus. A nowcast running far from the Blue Chip / professional-forecaster consensus is a flag that the survey crowd may be stale, since surveys are produced weeks earlier.
  • Macro/regime context, not signal. For market participants the value is directional and regime-level — is growth accelerating toward potential, stalling, or contracting — feeding broader recession-risk and cyclical/defensive positioning judgments. It is far too noisy to trade off any single update.

Adoption, debate & evidence

Nowcasting is now standard infrastructure: most major central banks run GDP nowcasts, and the Atlanta and NY Fed versions are widely cited in financial media. The empirical record is honest about both strengths and limits:

  • Accuracy. Per the Atlanta Fed, the average absolute error of the final GDPNow forecast (the last nowcast before the advance estimate) versus the BEA's first ("advance") figure was about 0.77 percentage points over 2011:Q3–2025:Q2 (with a root-mean-squared error around 1.17 pp). Accuracy improves sharply within a quarter: the Atlanta Fed reports the average absolute error is roughly 1.1 pp about 90 days out but near 0.5 pp just before the release. In the five-plus pre-pandemic years the average error was about 0.5 pp, and errors widened through the pandemic era. (A St. Louis Fed Dec-2025 review cites slightly different sub-period figures of ~0.65 pp for 2021–2025 vs ~0.51 pp pre-pandemic; those specific numbers come from that single article and are not independently corroborated here.)
  • Versus consensus. The Atlanta Fed reports GDPNow's final forecasts have had a slightly smaller average absolute error than the Blue Chip consensus over recent years — but the Blue Chip number is typically produced about three weeks earlier, so this is not an apples-to-apples win. Higgins' original work found GDPNow only "slightly inferior" to near-term Blue Chip consensus, with the optimal weight on GDPNow rising as the GDP release nears (the closer you are to the print, the more the hard data the nowcast has ingested matters).
  • Documented failures. The model is vulnerable to components it cannot see in real time. The Atlanta Fed and others noted that in Q1 2025 GDPNow's standard read fell to around -2.7% against a BEA advance print near -0.3% (later revised to about -0.5%). The gap was driven largely by a surge in non-monetary gold imports (roughly $30bn+ in early 2025, much of it gold bars moved into U.S. vaults ahead of tariff fears): imports subtract from GDP, so the model read the gold inflow as a collapse in net exports. The Atlanta Fed responded in April 2025 by adding a gold-adjusted version that stripped those flows out — a concrete reminder that the accounting-identity build-up is only as good as the source-data assumptions. During COVID the NY Fed suspended its Nowcast entirely (Sept 2021) because the data volatility broke the model, relaunching a more robust 2.0 version in Sept 2023.

The genuine controversy is over interpretation, not method: financial commentary routinely treats an early-quarter GDPNow print as if it were a forecast of the final number, when the Atlanta Fed explicitly frames it as "a running estimate based on available data," noisiest at the start of the quarter and most informative near the end.

Strengths & limitations

When it works: late in a quarter, once most hard data (jobs, retail, trade, construction) is in, the nowcast is well-anchored and genuinely informative; and as a transparent, judgment-free, news-decomposable tool it shows exactly what moved the estimate.

When it fails: early in a quarter it is dominated by model extrapolation and can swing wildly with a single report; it is blind in real time to components reported with long lags (inventories, net exports) and to one-off distortions (tariff front-running, pandemic shocks); and it nowcasts the advance estimate, which the BEA itself later revises.

The #1 misuse: treating a single early-quarter GDPNow update as a prediction of the final GDP number — or, for markets, as a tradeable signal. It is a coincident estimate of now, not a forecast of the future, and not a market-timing input.

Sources

Flags: (1) The headline 0.77 pp average absolute error is the Atlanta Fed's figure for the final GDPNow forecast over 2011:Q3–2025:Q2 — not a generic "since 2011 vs. advance" error for every nowcast (early-quarter errors run ~1.1 pp). (2) The ~0.65 pp / ~0.51 pp sub-period figures come from a single St. Louis Fed Dec-2025 article and are not independently corroborated. (3) The GDPNow-vs-Blue-Chip "slightly better" comparison is not a fair head-to-head — Blue Chip is produced ~3 weeks earlier. (4) The Q1 2025 -2.7% read reflects the standard model; the driver was specifically non-monetary gold imports, and the Atlanta Fed's later gold-adjusted figure was materially less negative.