Distributional regression typically requires tailoring methods to the response variable's type—continuous, discrete, or otherwise. A new arXiv preprint proposes a single framework that works for arbitrary response variables, using ReLU neural networks to model conditional event probabilities directly.

Instead of relying on scalar thresholds or assuming Euclidean structure in the response space, the approach indexes conditional event probabilities in a way that adapts to the data's own geometry. This removes a common design constraint and could simplify applying regression to unusual or complex outcomes.

Because the paper's abstract is brief, details on architecture and experiments are not yet available. The contribution at this stage is conceptual: a unified formulation that treats distributional regression as a probability-modeling problem rather than a thresholding problem.