An empirical evaluation of the zero-shot, long-horizon forecasting ability of leading time-series foundation models (TSFMs) on synthetic noisy periodic series, benchmarked against two cheap classical baselines: an FFT-based spectral method and a linear autoregressive model. Sweeping noise level, underlying frequency, sampling rate, and waveform complexity, the study finds a conditional result - TSFMs match or beat the statistical methods when periods are bounded and sampling rates high, but their accuracy degrades as periods lengthen, noise increases, sampling rates drop, and waveform shapes grow more complex - mapping the regime where general-purpose TSFMs are competitive with classical forecasters.
This is an empirical study of how well time-series foundation models (TSFMs) - large models pretrained on many series and applied zero-shot - forecast noisy periodic signals over long horizons. The a…