Improving Conditional Coverage in Time-Series Foundation Models via Direct Volatility Modeling

Jason B. Cho, David S. Matteson

Presentation Attendance: No, we cannot present in-person

Keywords: Uncertainty Quantification; Volatility; Conformal Inference

TL;DR: Time-series foundation models achieve reasonable marginal coverage but are conditionally miscalibrated under persistent volatility regimes; modeling second-order variance dynamics with GARCH substantially improves local uncertainty calibration.

Abstract: Time-series foundation models (TSFMs) provide probabilistic forecasts and are often reported to be well calibrated when evaluated via marginal coverage. However, forecast errors in time series are typically serially dependent and heteroskedastic, so marginal calibration can mask substantial regime-dependent miscalibration. We show empirically that prediction intervals from several TSFMs systematically under-cover during high-volatility periods, despite achieving near-nominal coverage on average. We propose a post-hoc multiplicative volatility correction that uses the uncertainty quantification produced by the TSFM as a baseline scale and dynamically adjusts it through a GARCH model. We evaluate this proposed method against adaptive conformal prediction as a benchmark. Across four real-world datasets, the proposed correction yields markedly improved conditional calibration over both the native TSFM intervals and adaptive conformal prediction. These findings highlight the importance of treating predictive variance as a dynamic process in TSFM uncertainty quantification.

Track: Research Track (max 4 pages)

Submission Number: 76