Introduces 'Noise Titration' to prove that current time-series foundation models often fail at structural inference, behaving instead as 'context parrots' during non-stationary shifts.
March 24, 2026
Original Paper
Noise Titration: Exact Distributional Benchmarking for Probabilistic Time Series Forecasting
arXiv · 2603.22219
The Takeaway
It moves time-series evaluation from passive trajectory matching to exact-statistical benchmarking on chaotic systems. The finding that zero-shot foundation models fail where structurally-aware models (like Fern) succeed suggests we need a major pivot in how we build 'Generalist' forecasting models.
From the abstract
Modern time series forecasting is evaluated almost entirely through passive observation of single historical trajectories, rendering claims about a model's robustness to non-stationarity fundamentally unfalsifiable. We propose a paradigm shift toward interventionist, exact-statistical benchmarking. By systematically titrating calibrated Gaussian observation noise into known chaotic and stochastic dynamical systems, we transform forecasting from a black-box sequence matching game into an exact di