How to Tell if a Climate Model Is Truly Reliable?

Imagine looking at a map of temperature or pollution: the highest values are like wave peaks on a rough sea. To study these peaks, statisticians use "max‑stable fields", models that describe how extreme events behave across space. But how can we check if the model fits the real data? This is where a goodness‑of‑fit test comes in. The authors built a statistic based on the Fourier transform of extreme‑value indicators. In practice, they turn the presence or absence of a peak into a signal that can be analyzed like a radio wave. When the assumed model is correct, the result follows a Gaussian distribution over the whole map. The catch is that the covariance of this distribution isn’t known in a closed form. To solve this, they introduced a "stationary bootstrap", a method that creates many synthetic versions of the spatial field to estimate the test’s critical values. Computer simulations confirm the theory works, and real‑world applications to PM2.5 and temperature data show the test can truly help pick the best model. Thus, scientists gain another tool to decide if their climate forecasts are reliable or need tweaking.

Max‑stable fields form the backbone of extreme‑value theory for spatial data. A field is max‑stable if, after taking the maximum of a finite number of independent copies of the field, the resulting distribution belongs to the same parametric family as the original field. This concept is essential for modeling phenomena such as intense rainfall or PM2.5 concentrations, where distribution tails are very heavy. To assess the adequacy of a max‑stable model, the authors consider indicators I(x)=1_{Z(x)>u}, where Z(x) is the observed value at location x and u is a high threshold. The Fourier transform of I(x) yields a complex function φ(ω)=∑_x I(x) e^{-iω·x}, capturing the spectral structure of the peaks. Under the null hypothesis of a correct model, the process φ(ω) converges, as sample size grows, to a centered Gaussian field with covariance C(ω,ω'). However, C is not available in closed form because it depends on the spatial dependence of the max‑stable field. To bypass this obstacle, a stationary bootstrap is employed: the observed field is divided into overlapping blocks, which are randomly recombined to generate new samples preserving local dependence. Repeating the test on thousands of bootstrap replicates yields an empirical distribution of the test statistic, from which critical quantiles are extracted. Simulations show that block size affects precision, but with reasonable choices (e.g., 10–20 km blocks) the error rate stays below 5%. Applying the method to Chinese city PM2.5 data and global temperature series revealed discrepancies between traditional models and max‑stable ones, suggesting that these tests can guide the selection of more robust models for extreme‑event prediction.