Spectral Evidence Bundling boosts time-series reliability by 13%

Traditional post-hoc calibration for time-series classifiers only remaps output scores, leaving a critical blind spot: identical confidence values can mask vastly different temporal support, and average calibration can miss false high-confidence errors. Filippo Cenacchi, Longbing Cao, and Runze Yang address these gaps in their new paper by introducing Spectral Evidence Bundling (SEB). The method keeps the backbone prediction unchanged but estimates whether it should be trusted by combining output-side cues with whole-sample spectral descriptors—including band energy, entropy, peak dominance, period support, and phase stability. A validation gate ensures spectral conditioning is applied only when correctness ranking improves without breaching FalseConf@0.9 or AURC tolerances; otherwise, it reverts to the safer output-space baseline.

Tested on eight heterogeneous UCR/UEA time-series datasets and eight backbone families (including standard recalibrators), the unconstrained SEB improves Corr-AURC from 0.693 to 0.779 on the matched evaluation subset. The validation-gated policy further pushes Corr-AURC to 0.786 while reducing FalseConf@0.9 to 0.094—a 9.4% floor on false high-confidence errors. These results demonstrate that reliability estimation for time-series classifiers benefits significantly from bundling output confidence with spectral evidence, and that validation gating prevents unsupported spectral conditioning. The work opens new avenues for trustworthy deployment in domains like finance, healthcare, and industrial IoT where time-series predictions must be auditable and safe.