This is the optional Uncertainty Quantification (UQ) track of the time-bounded challenge used to launch the persistent ClimX benchmark. While the main track focuses on deterministic point predictions, this track evaluates a model's ability to provide reliable probabilistic projections of climate extremes.
The challenge runs from September 1 through November 30, 2026; the benchmark resources remain available afterward. The organizers plan to submit an accompanying paper to the NeurIPS 2026 workshop Tackling Climate Change with Machine Learning in Sydney. If accepted, challenge results will be presented there; both tracks proceed independently of that decision.
The UQ track is an optional research track and does not offer cash prizes or travel support. The conditional prize package described on the main-track page applies only to eligible main-track teams.
The input variables, data resolution (daily), and training constraints are identical to the main competition. The core difference is that instead of a single value, you must provide a distribution for the 15 climate extreme indices.
To evaluate the predicted probability distributions, participants are required to submit a specific set of 9 quantiles for each of the 15 extreme indices: (0.05, 0.1, 0.2, 0.35, 0.5, 0.65, 0.8, 0.9, 0.95). This format is agnostic to your model architecture—whether you use ensembles, Bayesian Neural Networks, or density estimation models.
Leaderboard files must use the full-resolution grid. A submission must contain 86 annual maps for SSP2-4.5 (2015–2100) in time–latitude–longitude order. Each index–quantile combination must be present as a column named {index}_{quantile}. For instance, index TXx requires columns such as TXx_0.05, TXx_0.1, etc., up to TXx_0.95. The first column is row_id, containing integers from 0 through 4755455. Rows are ordered by year (2015–2100), then increasing latitude, then increasing longitude.
The primary leaderboard metric for this track is the Continuous Ranked Probability Score (CRPS), which replaces the MSE used in the deterministic track. The CRPS is a strictly proper scoring rule that rewards models for being both accurate and having a "sharp" (certain) distribution. Scores are aggregated hierarchically from pixel-level calculations to regional weighting using cosine-latitude weights, and finally averaged uniformly over valid regions and all 15 indices.
A model might accurately predict the "average" temperature while being wildly over-confident (or under-confident) about the risk of a record-breaking heatwave. By participating in this track, you are helping build emulators that can communicate risk and uncertainty, which is vital for real-world climate adaptation and policy-making.