# ConceptTime

ConceptTime uses the mismatch between a frozen probabilistic forecaster's next-step prediction and the realized signal as a single signal that both places time-series segment boundaries and describes how each segment behaved relative to expectation, yielding an inspectable concept vocabulary without fixed windows, expert concept annotations, or language-model calls. Across 17 public benchmarks it is competitive with black-box accuracy and beats every interpretable baseline; on Epilepsy it retains 97.91% accuracy with 1% of labels, and prototype distance doubles as a fidelity score for abstention.

ConceptTime uses the mismatch between a frozen probabilistic forecaster's next-step prediction and the realized signal as a single signal that both places time-series segment boundaries and describes how each segment behaved relative to expectation, yielding an inspectable concept vocabulary without fixed windows, expert concept annotations, or language-model calls. Across 17 public benchmarks it is competitive with black-box accuracy and beats every interpretable baseline; on Epilepsy it retains 97.91% accuracy with 1% of labels, and prototype distance doubles as a fidelity score for abstention.