Blog

Is Weather Forecasting Getting Worse

Jan Gunash · Co-Founder & CGO · July 6, 2026

The best questions are not always the comfortable ones. Like: if weather forecasting has been getting measurably better every decade for half a century, why does it suddenly feel less reliable, and are those two things actually in conflict? And when the data that feeds every forecast model quietly thins out because the instruments that collect it are being switched off, how long before the decline shows up in the forecast on your phone?

Weather forecasting is one of the great quiet successes of modern science. A five-day forecast today is about as accurate as a one-day forecast was in the early 1980s. Forecast skill has improved by roughly one day of lead time per decade, and the arrival of AI weather models over the past two years has pushed accuracy further still, with some now outperforming traditional physics-based models on the majority of standard measures. And yet the sense that forecasts are missing more often is real, widespread, and in some specific ways justified. This edition is about why both things are true at once, and about the data infrastructure quietly eroding underneath the models.

What’s Getting Better?

Start with what is genuinely improving, because it frames everything else. Modern numerical weather prediction has advanced on three fronts at once: higher-resolution models, faster data assimilation, and better physics. The result is that ordinary weather, the temperature tomorrow, the chance of rain this afternoon, is forecast today with a reliability that would have seemed impossible a generation ago. Short-term temperature forecasts are now routinely accurate to within one or two degrees (source: Climavision).

The newest leap comes from AI. Models like GraphCast, GenCast, and ECMWF's AIFS are trained on decades of atmospheric reanalysis data and can produce a global forecast in seconds rather than the hours a traditional model needs on a supercomputer. As of 2026, AI weather forecasting outperforms traditional models on roughly 90% of standard metrics, and national weather agencies now publish AI forecasts alongside their conventional products (source: ECMWF / AI forecasting analysis).

But here is the detail that matters for the rest of this story. Those AI models do not observe the weather. At the moment they run, they need a current snapshot of the global atmosphere, temperature, wind, humidity, and pressure at hundreds of locations and many altitudes, and that snapshot comes from the same observation networks that feed traditional models: weather balloons, satellites, ocean buoys, and aircraft sensors (source: AI forecasting analysis). A better model does not reduce the need for good input data. It raises it. The most advanced forecast system in the world is still only as good as the observations it starts from.

What’s Getting Worse?

The observations are where the problem lives, and it is a specific, measurable problem. Across the United States, weather balloon launches have been suspended or reduced at a series of sites. At least ten to eleven locations halted or cut back their launches, which historically happened twice daily at more than 100 sites (source: NBC News / Gizmodo).

Weather balloons, carrying instruments called radiosondes, rise through the atmosphere measuring temperature, humidity, wind, and pressure directly. They provide the finest-resolution vertical picture of the atmosphere available, and it is a picture that cannot easily be replicated by satellites. When NASA researchers compared the instruments that feed forecast models, they found radiosondes had the second-largest impact on forecast quality, behind only satellites (source: NBC News). Without any weather balloons at all, one analysis found today's forecasts would be about 15% worse.

The consequences are sharpest exactly where forecasting matters most. In May, at the start of peak tornado season for the Southern Plains, National Weather Service offices in Kansas failed to conduct three-quarters of scheduled morning weather balloon launches (source: Gizmodo). The vertical data those launches would have captured is precisely what models need to predict whether a storm system turns severe. There is precedent for the scale of the effect. When Russia cut its radiosonde launches roughly in half over a few months in 2015, European forecasters measured a decline in their model's forecast quality (source: NBC News).

This is the mechanism behind the feeling that forecasts have slipped. It is not that the models got worse. It is that in specific regions and specific moments, the data going into them thinned out, and the models can only be as good as what they are fed.

Why Does the Forecast Feel Less Accurate?

There is a second reason forecasts feel less accurate than the skill scores suggest, and it is about which forecasts fail rather than how many.

Forecast skill is an average, and averages are dominated by ordinary weather, which is now predicted very well. The forecasts people remember, though, are the dramatic misses: the storm that was supposed to hit and did not, the downpour that was not in the forecast, the heat that exceeded every projection. Those are almost always extreme, low-probability, high-impact events, and they sit at the tail of the distribution where models have always struggled most. Sudden thunderstorms, microclimates, and rapidly evolving systems remain among the hardest things to forecast even as average accuracy climbs (source: Climavision).

Climate change sharpens this problem. Forecast models and the statistics beneath them are calibrated on the historical record. As extreme events become more frequent and arrive outside their historical seasons, the models are increasingly asked to predict conditions with little precedent in the data they learned from. The average forecast keeps improving. The extreme forecast, the one that matters most for safety, is being asked to do something harder each year. Better average skill and more memorable misses are not a contradiction. They are the same distribution viewed from two ends.

The Communication Gap

The third reason is not about the atmosphere at all. It is about how a forecast reaches you. Modern forecasts are probabilistic. The model does not say it will rain. It says there is a 70% chance of rain, which is a statement about a range of possible outcomes. By the time that forecast reaches a phone screen, it has usually been flattened into a single icon: a raincloud, a sun, a number.

When a well-calibrated 70% forecast produces no rain, which it should, correctly, about three times out of ten, it is remembered as a wrong forecast. It was not. The forecast was accurate; the communication of its uncertainty was lost somewhere between the model and the screen. A forecast that never shows you its confidence level cannot help you understand that uncertainty is a feature of the science, not a failure of it. The data exists in probabilistic form. The interface throws most of it away.

Where This Fits the Pattern

This is the same structure this newsletter has traced across every domain it has covered. The models are improving. The physical system is getting harder to predict. And the data infrastructure that connects the two is fragile, underfunded, and in places actively shrinking.

In emissions, the monitoring databases disagreed and the uncertainty was hidden behind authoritative-looking numbers. In the ocean, a monitoring network was pulled from the water before its record could mature. In weather, the balloons that provide the single most valuable atmospheric data after satellites are being launched less often, right as climate change makes the extreme events they help predict more frequent and more dangerous. The pattern is consistent. The instrument layer, the unglamorous foundation of sensors and observations that everything else is built on, is treated as the easiest thing to cut, when it is in fact the thing that determines whether any of the sophistication above it actually works.

There is a growing weather inequality dimension to this as well. The regions most exposed to extreme weather are often those with the sparsest observational coverage. The World Bank estimates up to 13 billion dollars in annual benefits from improved weather services across sub-Saharan Africa alone, through better agricultural output, reduced disaster losses, and more efficient water management (source: WindBorne / World Bank). The places that need the forecast most frequently have the least data feeding it.

The Path Forward

The encouraging part is that this is one of the most fixable data gaps the series has covered, because the science is not in question. Everyone agrees radiosonde data improves forecasts. The measured impact is well documented. The instruments are comparatively cheap. Restoring the twice-daily launch cadence at the sites that reduced it is not a research problem or a technology problem. It is a funding and staffing decision.

Private networks are beginning to fill some of the gap. Startups deploying low-cost, long-duration balloons have offered their data to national services where launches stopped, and are building global constellations aimed at exactly the data-sparse regions traditional systems miss (source: WindBorne). That is a genuine contribution, but it should complement a functioning public observation network, not quietly replace it. Atmospheric observation is a public good, and the forecasts that depend on it protect everyone regardless of who collected the data.

The deeper lesson is the one this series keeps returning to. The most advanced model in the world, whether it runs on physics or on a neural network, begins with a snapshot of the atmosphere it did not take itself. Thin out that snapshot and the forecast degrades no matter how good the model is. The forecast that keeps missing is not usually a failure of prediction. It is a failure to keep watching closely enough to predict well. The instruments are the forecast. When they go quiet, so does our ability to see what is coming.

Sources:

Want to contribute your thoughts or findings in data and AI with climate, energy, or nature? Message Jan Gunash on LinkedIn or send an email to jgunash@projectcoeus.com.