The climate signal hides in the averages. This week measured why

Averages lie

Four preprints posted this week make the same complaint from different corners of climate science: the global average is hiding the signal. Ocean deoxygenation cannot be robustly detected below the surface with current machine-learning reconstructions. Cold extremes are warming more slowly than the rest of the temperature distribution. Greenhouse gas concentrations swing sharply along a single coastline. And forest carbon fluxes in the Alps split by driver, biomass for carbon, climate for water, at a resolution coarse maps cannot see. The shared lesson is that climate risk work has to be local and physical, not averaged and assumed. m:…7312 m:…2873 m:…9295 m:…8514

The ocean's missing trend

The ocean study is the sharpest warning. Testing neural network configurations for reconstructing dissolved oxygen between 1993 and 2022, the study found the models were highly sensitive to which climate predictors they were given, and that including the calendar year as a predictor created artificial trends with no physical driver behind them. With that error removed, robust detection of global deoxygenation was limited to the surface ocean; in the interior, methodological uncertainty made robust detection largely unachievable. The recommendation is to build future reconstructions on physical variables, such as temperature anomalies at depth, rather than on time itself. m:…7312

Cold that stays, gas that varies

The cold-extremes study, based on equilibrium and transient climate simulations, found that winter temperature variability in the Northern Hemisphere midlatitudes falls asymmetrically as the world warms: the cold tail of the distribution shrinks more slowly than the warm tail, so cold extremes warm more slowly than the distribution's core and its hot extremes. For infrastructure operators and insurers, the practical read is that a warmer mean does not retire cold-weather risk, the hazard persists, and systems adapted to past cold keep facing it. m:…2873

The coastal study shows the same locality at sea. High-frequency ship-board measurements in the California upwelling system found greenhouse gas concentrations far higher north of Cape Blanco than south of Cape Mendocino, where local circulation flowing northward offset the upwelling and lowered surface gas concentrations. The gases had different drivers: nitrous oxide followed the upwelling of enriched subsurface water, while carbon dioxide tracked sea-air exchange and productivity. A single upwelling-flux number for the coast would mislead in both directions. m:…9295

Forests, at 30 metres

The Alpine study maps the same problem onto land. Combining process-based ecosystem models with machine learning, researchers produced 30-metre-resolution maps of gross and net primary productivity and evapotranspiration across the Italian Alps for 2020. Biomass was the primary driver of carbon uptake; climate variables drove water loss. Random Forest models gave the highest predictive performance among the machine learning algorithms tested, a reminder that mapping mountain carbon is far from a solved problem. The value is the resolution: carbon accounting that cannot see a mountainside cannot price it either. m:…8514

None of these papers disputes warming. They dispute how precisely it can be read from the wrong scale, and they converge on the same prescription: measure locally, use physical drivers, and treat model outputs as hypotheses with stated uncertainty. For anyone pricing climate risk, the week's message is to distrust the smooth global line and ask what the local measurement says. m:…7312 m:…2873 m:…9295 m:…8514