Spectral Evidence Boosts Time-Series AI Reliability Beyond Output Calibration

A new approach to improve the reliability of time-series classification models has been developed, moving beyond conventional output-space calibration. Traditional methods often remap output scores, but fail to address whether a confident prediction is genuinely supported by the underlying temporal signal. This research introduces a validation-gated fixed-label reliability policy that keeps the model's core prediction unchanged while estimating its trustworthiness. The method combines standard output-side confidence cues with comprehensive spectral descriptors of the entire time-series sample, including features like band energy, entropy, and phase stability. This "spectral evidence bundling" provides a more nuanced scalar reliability estimate and diagnostic band-level insights. When evaluated across various UCR/UEA datasets and time-series backbone families, this unconstrained method significantly improved selective-reliability metrics, and the validation-gated policy further enhanced performance and reduced false high-confidence errors, suggesting that integrating spectral evidence is crucial for robust time-series reliability.