The concept of omniprediction has gained attention as a way to build predictors that are not tuned to a single metric but remain accurate across many. According to a new arXiv paper, prior work in this area has largely focused on binary prediction. The authors now aim to broaden that scope.
The paper introduces online quantile omniprediction for proper losses. This moves the problem into an online setting, where predictions are made sequentially, and targets quantile estimation rather than binary outcomes. Proper losses include common scoring rules like squared error and log loss, giving the framework wide applicability.
The abstract is brief, so details on the algorithm and theoretical guarantees are not fully available from the source. Still, the work signals a step toward making omniprediction practical for a broader class of forecasting tasks beyond the binary case.