Nowcasting is the practice of estimating what is happening right now, using incomplete or delayed data. The term, a blend of “now” and “forecasting,” originated in meteorology but has spread to economics, public health, and other fields where official data arrives too late to act on in real time. Rather than predicting the future, nowcasting fills the gap between what has already occurred and what the data currently shows.
How Nowcasting Differs From Forecasting
Traditional forecasting projects what will happen days, weeks, or months ahead. Nowcasting looks backward and inward: it tries to figure out what is happening today or what happened very recently, before complete data is available. The distinction matters because in many fields, the present is surprisingly hard to pin down. GDP figures arrive months after a quarter ends. Disease surveillance data trickles in over weeks as hospitals file reports. Even weather radar has blind spots.
Forecasting models typically rely on historical patterns and projected trends. Nowcasting models, by contrast, pull in whatever real-time or high-frequency data is available and use it to construct the most accurate possible picture of current conditions. The time horizons are different, the data inputs are different, and the goal is fundamentally different: not “what will happen?” but “what is actually happening right now that we can’t yet see?”
Weather: Where Nowcasting Started
In meteorology, nowcasting means forecasting with local detail over a period from the present to six hours ahead, including a description of current weather. That definition was formalized by the World Meteorological Organization in 2010. The six-hour window is the zone where radar and satellite observations outperform traditional numerical weather models, which need time to process atmospheric equations across a global grid.
Most precipitation nowcasting relies on ground-based radar. Algorithms track how rain, snow, or storm cells are moving and extrapolate their paths forward using optical flow techniques, essentially treating radar images like frames in a video and predicting the next few frames. Satellite imagery has expanded this capability to areas without dense radar networks. Yandex, for instance, developed a satellite-based nowcasting method to extend precipitation mapping to regions where ground radar doesn’t reach, moving toward global coverage.
More recently, deep learning has entered the picture. A team published in Nature developed a system using deep generative models that produces realistic, probabilistic rain predictions from radar data. Their approach uses a type of neural network that learns to generate plausible future radar sequences, with a built-in check (a “temporal discriminator”) that penalizes predictions that look jumpy or physically implausible. The result is a nowcast that captures the kind of fine-grained detail, like exactly where a thunderstorm cell will intensify, that conventional methods struggle with.
Economics: Estimating GDP Before It’s Released
GDP, the most watched measure of economic health, is published quarterly and arrives with a substantial delay. In the U.S., the first estimate comes roughly a month after a quarter ends, and it gets revised multiple times after that. Nowcasting in economics fills this gap by using higher-frequency data, things like weekly jobless claims, daily financial market movements, monthly retail sales, and manufacturing surveys, to estimate GDP growth while the quarter is still underway.
The Federal Reserve has found that survey data from the Federal Reserve Bank of Philadelphia has a particularly large impact on GDP and inflation nowcasts, even more than the closely watched Employment Report. When researchers account for how quickly different data sources become available, “hard” data like industrial production and retail sales also contribute substantially. Interest rate movements feed into GDP estimates as well.
Google search data has added another dimension. Research using Google Trends in a Bayesian statistical framework found that a broad set of search terms can improve GDP nowcasts early in a quarter, before traditional economic indicators are published. The most useful search terms tend to reflect economic anxiety and wealth effects, meaning people’s searches for things like unemployment benefits or mortgage rates serve as early signals of where the economy is heading.
Public Health: Seeing Outbreaks in Real Time
Epidemiological nowcasting tackles a specific and persistent problem: reporting delays. When someone is hospitalized with the flu or COVID-19, it can take days or even weeks for that event to appear in surveillance databases. The result is that the most recent data always looks artificially low, creating a misleading picture of whether an outbreak is growing or shrinking. This incomplete tail end of the data is called “right truncation.”
The CDC uses nowcasting to correct for this. Their system takes weekly data containing both the date someone was hospitalized and the date the hospitalization was reported. By comparing these two dates across many historical records, the models learn the typical delay pattern: how many events get reported within one day, two days, a week, and so on. That delay distribution is then applied to the most recent, incomplete data to estimate how many events actually occurred but haven’t been reported yet.
For respiratory viruses, the CDC combines predictions from three separate models into an ensemble. One model simply applies historical delay patterns to current data. The other two use more flexible statistical approaches that can adapt to changing conditions, like a holiday weekend that slows reporting or a surge that overwhelms data entry staff. Together, these models produce a corrected estimate that reveals changes in disease transmission before they would otherwise be visible in raw data.
The concept has been applied to dengue in Puerto Rico, influenza-like illness across the U.S., and COVID-19 during the pandemic. In each case, the core challenge is the same: adjusting for the gap between when infections happen and when the numbers show up.
How Nowcasting Models Work
Despite spanning very different fields, nowcasting models share a common logic. They take incomplete current observations, combine them with patterns learned from historical data, and produce an adjusted estimate of reality. The specifics vary, but the architecture falls into two broad categories.
Statistical models describe mathematical relationships between variables without trying to simulate the underlying system. In economics, this might mean identifying which monthly indicators correlate most strongly with quarterly GDP. In epidemiology, it means fitting a statistical curve to reporting delays. These models are flexible and relatively fast to run.
Mechanistic models, by contrast, simulate the actual process generating the data. In disease nowcasting, a mechanistic model might explicitly represent how infections spread, how long symptoms take to develop, and how the reporting pipeline works. The CDC’s Epinow2 tool takes this approach, using a renewal model of disease transmission combined with Bayesian inference to estimate both current case counts and how quickly the virus is spreading.
Key Challenges and Limitations
The biggest technical challenge in nowcasting is that reporting delays aren’t fixed. They shift with holidays, staffing changes, system outages, and surges in case volume. A model trained on “normal” delay patterns can produce misleading estimates when the delays themselves change. This non-stationarity problem has driven the development of newer tools. One recent approach, grounded in survival analysis, uses a sliding window of recent data to estimate current delay probabilities rather than relying on the full historical record. The idea is that the most recent reporting periods better reflect today’s delays than data from months ago.
Uncertainty is another inherent limitation. Because nowcasts are estimates built on incomplete information, they come with prediction intervals that can be wide, especially for the most recent time points where the least data is available. Responsible nowcasting systems communicate this uncertainty explicitly, showing not just a best estimate but a range of plausible values.
Data noise poses problems across all domains. In economics, high-frequency indicators like search trends or financial market data can be volatile and misleading during unusual events. In weather, radar returns can be contaminated by ground clutter or anomalous propagation. In epidemiology, a single large hospital batch-reporting old cases can create an artificial spike. Each field has developed its own techniques for smoothing out noise, but no nowcast is immune to garbage-in-garbage-out problems.
Day-of-week effects are a surprisingly stubborn issue in health nowcasting. Fewer cases get reported on weekends, creating artificial weekly cycles in the data. Models need to account for these patterns explicitly, or they’ll mistake a normal Monday reporting bump for a real increase in disease activity.
Why Nowcasting Matters
The practical value of nowcasting comes down to speed. Public health officials who can see a respiratory virus surge two weeks earlier can activate hospital surge plans sooner. Central bankers who can estimate GDP growth in real time can make better interest rate decisions. Meteorologists who can track a developing thunderstorm minute by minute can issue severe weather warnings that save lives.
In each case, nowcasting converts delayed, incomplete, or low-resolution data into something actionable. It doesn’t replace traditional forecasting or final official statistics. It fills the window between “something is happening” and “the numbers are in,” which is often exactly when decisions need to be made.