Researchers propose new metrics to fix AI's biggest forecasting flaw

Researchers have introduced two new kernel-based metrics, Sig-MMD and CSig-MMD, designed to solve critical flaws in evaluating AI forecasting models. Current methods fail to capture complex dependencies and are notoriously bad at assessing predictions for rare, high-impact 'tail events' like market crashes or disease outbreaks. The new metrics, leveraging signature kernels, specifically prioritize a model's ability to predict these crucial outliers while maintaining robust statistical properties for reliable comparison.