LOCUS: A Distribution-Free Loss-Quantile Score for Risk-Aware Predictions

A team of researchers has introduced LOCUS, a novel framework designed to make machine learning predictions more risk-aware by directly quantifying the potential cost of being wrong. The core problem LOCUS addresses is that modern AI models can be highly accurate on average but still produce occasional, catastrophic errors that dominate real-world deployment costs. Unlike traditional methods that focus on label uncertainty, LOCUS models the actual *loss* a prediction function will incur for a given input. It works as a distribution-free 'wrapper,' meaning it can be applied to any pre-trained model without assuming a specific statistical distribution for the errors. A key innovation is a split-calibration step that transforms a predictive distribution for the loss into an interpretable, comparable score across different inputs, which can be read as an upper bound for the expected loss.

The technical approach allows LOCUS to be used in two primary ways: for ranking predictions by risk or for creating transparent flagging rules that provide distribution-free control over large-loss events. In experiments spanning 13 diverse regression benchmarks, LOCUS demonstrated superior performance in risk ranking and significantly reduced the frequency of high-cost errors compared to conventional uncertainty quantification heuristics. This makes it a powerful tool for high-stakes applications like finance, healthcare, or autonomous systems, where the consequence of a single bad prediction is severe. The method's flexibility and lack of distributional assumptions mean it could be widely adopted as a safety layer for existing AI models, moving the field toward more reliable and trustworthy systems that manage not just average performance, but worst-case outcomes.