A team of Finnish researchers has now built the tool that could change how the world judges its forecasts. Zhi-Song Liu of LUT University, Michael Boy of LUT University and the University of Helsinki, and Risto Makkonen of the Finnish Meteorological Institute have created TailWeather, a global dataset of climatological extreme weather events described in a preprint currently under review for the journal Earth System Science Data. The dataset covers every land area on Earth from 1981 through 2022 and into early January 2023, resolving the planet into a grid of cells 0.25 degrees on a side, roughly 25 to 28 kilometers at mid-latitudes. For each cell and each day, TailWeather labels whether the weather qualified as a heatwave, a cold wave, heavy precipitation, or extreme wind, and for each month it flags meteorological drought.

The technical insight behind TailWeather is deceptively simple. Rather than defining extremes by their consequences, which requires knowing where damage was recorded and reported, the researchers defined them statistically, relative to each location’s own climate. Every grid cell is compared against its local 1991 to 2020 climatology, so a heatwave in Helsinki is measured against Helsinki’s normal summer, not against the tropics. This means a temperature that would be unremarkable in Death Valley can be flagged as an extreme event in Lapland, because what matters for infrastructure, agriculture, and human physiology is the deviation from what a place is built to expect.

Each label in the dataset carries two layers of information. The first is an ordinal severity tier, a ranking of how far into the statistical tail an event falls. The second is a continuous intensity score expressed relative to the local climate, which allows users to shift the threshold defining what counts as extreme without ever recomputing the underlying climatology. This design choice matters enormously for reproducibility. A researcher studying moderate heat stress and another studying lethal heat can both work from the same dataset, simply choosing different cutoffs within the score’s resolution, and their results will remain directly comparable.

To validate that these statistical tails actually correspond to real-world harm, the team compared TailWeather labels against EM-DAT, one of the most widely used international disaster catalogs. The comparison produced a striking and consistent pattern: for all five hazard types, the stricter the tail definition, the more strongly the events were enriched in recorded disaster impacts. At the best-balanced threshold, heatwave tails were 9.3 times more likely to coincide with recorded disasters than average weather, while extreme wind tails showed a 1.6-fold enrichment. Cold waves, heavy precipitation, and drought fell between these bounds. Rarity, in other words, is not a perfect proxy for damage, but it is a meaningful one, and the relationship strengthens as events become more exceptional.

Yet the validation also exposed a sobering blind spot in the global disaster record. The researchers calculated that 80.1 percent of the world’s land population lives in regions where EM-DAT is too sparse to test the link between climatological extremes and recorded harm locally for at least one hazard type. Disaster catalogs are shaped by reporting capacity, media attention, and political priorities, and vast stretches of Africa, Asia, and South America remain effectively unmapped in terms of documented weather impacts. This is why the authors emphasize that TailWeather complements rather than replaces disaster catalogs: it provides a physically consistent target everywhere, including the places where human records of suffering are thinnest.

The dataset’s most viral finding, however, concerns the AI models themselves. When the researchers tested leading machine-learning forecasting systems against TailWeather’s extreme event labels, the results were humbling. Models that achieve remarkable accuracy on ordinary days, hitting that sub-kelvin temperature target at day one, reached a critical success index below 0.5 on heatwave tails, meaning they missed or falsely flagged more than half of the extreme heat events. And their performance was worst of all for extreme wind, the hazard that drives much of the structural damage in storms. The models that look brilliant on paper may be quietly failing exactly when it matters most.

This weakness is not accidental. Machine-learning weather models are trained to minimize average error across billions of examples, and the mathematics of that optimization naturally prioritizes the frequent, well-sampled conditions that dominate the training data. Rare extremes are, by definition, underrepresented, and the loss functions that drive learning offer little reward for nailing a once-in-a-decade storm. Traditional numerical models, which solve physical equations rather than learn statistical patterns, have their own biases but are not subject to this particular failure mode in the same way. The result is that the AI revolution in weather forecasting, while genuinely transformative for routine prediction, may be building its impressive averages on a foundation that crumbles under extreme conditions.

The implications ripple outward across the entire weather enterprise. Insurance companies pricing catastrophe risk, governments planning heat-action plans, energy grid operators stress-testing supply, and humanitarian organizations prepositioning relief all depend on knowing when forecasts can be trusted and when they cannot. If the leading models systematically underpredict the very events that cause the most damage, then every downstream decision built on those forecasts inherits that blind spot. TailWeather gives these users, and the model developers themselves, a standardized global benchmark for measuring tail skill, the forecasting equivalent of a crash test for the AI era.

The dataset is openly available, distributed under a Creative Commons Attribution 4.0 license, and hosted with an interactive computing environment that allows researchers to explore the labels directly in a browser. The authors frame TailWeather as a reusable physical target for studying and evaluating weather extremes, one that can serve training, validation, and benchmarking across the rapidly growing ecosystem of machine-learning forecast systems. As climate change pushes more of the planet’s weather into unprecedented territory, and as AI models increasingly become the default tool for prediction, the ability to test forecasts specifically against the rare and the extreme is no longer a luxury. It is the difference between knowing the average weather and being prepared for the days that change everything.

Subject of Research: TailWeather: from tail to extremes, a global climatological dataset for machine-learning weather forecasting

Article Title: TailWeather: from tail to extremes, a global climatological dataset for machine-learning weather forecasting

Article References: Liu, Z.-S., Boy, M., & Makkonen, R. (2026). TailWeather: from tail to extremes, a global climatological dataset for machine-learning weather forecasting. https://doi.org/10.5194/essd-2026-792

Image Credits: AI Generated

DOI: 10.5194/essd-2026-792

Keywords: TailWeather, tail, extremes, global, climatological, dataset, machine-learning, weather, forecasting, scientific research