Biological Hallucinations in Time-Series Foundation Models: A Benchmark on High-Frequency Neural Waveforms

Time-Series Foundation Models (TSFMs) pretrained predominantly on financial, meteorological, and energy datasets have demonstrated impressive zero-shot generalization, yet their applicability to high-frequency biological signals remains critically under-examined. Neural waveforms such as electroencephalography (EEG) exhibit properties fundamentally distinct from economic or climate time series: strict stationarity within frequency bands, non-linear oscillatory dynamics governed by synaptic physiology, and spectral fingerprints that directly encode neurological state. We present the first systematic benchmark of two state-of-the-art TSFMs—Chronos-T5-Small and TimesFM-1.0-200M—against classical zero-shot baselines (DLinear, PatchTST) on the PhysioNet EEG Motor Imagery Database (160 Hz). We introduce Biological Hallucinations: model outputs that are statistically plausible under generic time-series priors but spectrally incoherent with neurophysiological reality, as measured by Jensen-Shannon divergence of power spectral densities. Our results show that TSFMs exhibit up to 0.6×higher spectral divergence than DLinear despite comparable MSE, exposing a fundamental gap between token-level reconstruction fidelity and biological signal integrity. We characterize performance degradation under frequency downsampling (160→40 Hz), identify three structural root causes, and outline a research agenda for neuro-aware foundation models.