GDP Nowcasting

GDP is released with a lag—the first estimate for a quarter typically arrives 30 days after quarter-end, with revisions trickling in for months. Nowcasting is the practice of synthesizing high-frequency data—credit card transactions, job postings, freight activity, electricity use, internet search volume—to estimate current-quarter GDP in near real-time. For policymakers, investors, and firms making bets on growth, a nowcast released weekly or monthly can matter more than the official release, which merely confirms what markets have already priced in.

Why the lag matters

The GDP release schedule is a relic of the pre-internet era. Official estimates require mountains of survey responses from households and businesses, weeks of processing, and careful quality control. The Bureau of Economic Analysis in the US releases an “advance” estimate about a month after quarter-end, a “second” estimate a month later, and a “final” estimate another month after that. By the time the official third release lands, the quarter is over and markets have already moved.

For a central banker deciding whether to raise or cut rates, waiting 30 days for an estimate while economic activity unfolds in real-time is frustrating. For a business deciding whether to hire or freeze payroll, the same lag is costly. If a slowdown is happening now, waiting until next month to learn it officially has already translated into job losses. This is where nowcasting steps in: use today’s data to infer today’s GDP, even if official accounting will follow weeks later.

The signal extraction problem

Nowcasting is not a science; it is disciplined guesswork. Statisticians face a classic signal-extraction problem: which high-frequency data actually reflect GDP growth, and which are noise or one-off shocks? A spike in electricity use might reflect stronger factory output, or it might just be a cold snap driving heating demand. A jump in credit card transactions might show consumer strength, or it might reflect one-time tax refunds or shift-work timing effects.

Nowcasters use regression models and filter techniques to tease signal from noise. They look at which indicators have historically been most correlated with GDP releases, what lead time they have, and how much measurement error they carry. A payroll report released on the first Friday of each month has high signal-to-noise ratio and a known timing; a credit card transaction count is noisier but available daily. Combining many noisy signals, each with different timing and reliability, can produce a reasonable estimate of the whole.

The Atlanta Fed GDPNow and relatives

The most famous nowcast is the Atlanta Federal Reserve’s GDPNow, launched during the 2008 financial crisis and released weekly. As new data arrives—employment figures, housing starts, PMI surveys, advance retail sales—the model updates its estimate of current-quarter GDP. In early 2020, when COVID shutdowns began, GDPNow crashed to −20% as it incorporated collapsing credit card spending and traffic data. It proved far more sensitive to the shock than official GDP estimates released a month later.

Other major financial institutions run nowcasts: the New York Federal Reserve Nowcast, Goldman Sachs’ Current Activity Indicator (CAI), JP Morgan’s Nowcast, and various academic models. They differ in methodology—some use machine learning, some Bayesian filtering, some simple regression—and data inputs. But they all aim at the same goal: turn high-frequency data into a real-time estimate of the low-frequency statistic (GDP) that markets care about.

Key data inputs and their timing

The nowcast kitchen uses a variety of ingredients, each arriving on its own schedule:

Labour market data arrives early: initial and continuing jobless claims come weekly; the employment report (payroll, unemployment rate) comes monthly on the first Friday. These are reliable and closely watched, so they heavily influence nowcasts. A surprise miss in payroll is immediately incorporated.

Purchasing Managers’ Indices (PMI) for manufacturing and services arrive monthly, usually mid-month. They reflect business confidence and order flow before money changes hands, giving them slight predictive power for production.

Retail and wholesale trade data arrive monthly: advance retail sales, auto sales, inventories. These feed directly into the consumption leg of GDP and are updated weekly if flash estimates are available.

Credit aggregates and payment flows can be tracked daily: credit card spending from processors, debit transaction volumes, and banking data. These are high-frequency but noisier than official sales figures.

Real-time proxies: traffic patterns from Google Maps or cell phones correlate with retail traffic; electricity use from grid operators correlates with industrial output; job postings correlate with hiring expectations.

The art is in weighting these signals. Early in the quarter, before much official data arrives, nowcasts lean heavily on credit and traffic proxies. As the month progresses and PMI, employment, and sales data land, nowcasts shift toward those. By month three of the quarter, the nowcast is essentially a blend of official partial data and extrapolations.

The surprise dynamic

Markets prize nowcasts most when they diverge from expectations. If the Atlanta Fed’s GDPNow is estimating 1.5% growth for the current quarter and consensus expects 2%, the gap signals potential for surprise (downside) when the official release lands. This can move bond yields, equity indices, and currency rates before the official data appears. In this sense, nowcasts can be more market-relevant than the official release itself.

During the 2020 pandemic, nowcasts provided crucial early signals of the collapse and recovery. GDPNow flashed −20% in April 2020 when the shock was fresh; official GDP eventually reported −31% annualized, but nowcasts had already priced in the magnitude. Conversely, in 2021–2022, nowcasts of persistent above-trend growth helped policymakers understand that inflation would be stickier than headline inflation suggested, because underlying demand was robust.

Measurement challenges and revision risk

Nowcasts are only as good as their inputs. If a data source is subsequently revised—employment figures, for instance, are revised monthly—the nowcast built on the preliminary data becomes partially obsolete. Over-reliance on a single high-frequency proxy can backfire. If credit card spending spiked because of gift card processing, not true consumption, the nowcast would overestimate growth.

Another risk: structural breaks. If a pandemic shuts call centres and flips demand online, historical correlations between traffic data and retail sales weaken. The nowcast must be re-estimated, which takes time. In early 2020, all models struggled because the shock was unprecedented; nowcasts produced wide ranges rather than point estimates.

See also