Beyond Accuracy: Are Time Series Foundation Models Well-Calibrated?

February 26, 2026 · View on GitHub

Official Repository for the paper "Beyond Accuracy: Are Time Series Foundation Models Well-Calibrated" Accepted at ICLR 2026

Paper | OpenReview

Authors

  • Coen Adler - UC Irvine
  • Yuxin Chang - UC Irvine
  • Felix Draxler - UC Irvine
  • Samar Abdi - Google
  • Padhraic Smyth - UC Irvine

Abstract

The recent development of foundation models for time series data has generated considerable interest in using such models across a variety of applications. Although foundation models achieve state-of-the-art predictive performance, their calibration properties remain relatively underexplored, despite the fact that calibration can be critical for many practical applications. In this paper, we investigate the calibration-related properties of five recent time series foundation models and two competitive baselines. We perform a series of systematic evaluations assessing model calibration (i.e., over- or under-confidence), effects of varying prediction heads, and calibration under long-term autoregressive forecasting. We find that time series foundation models are consistently better calibrated than baseline models and tend not to be either systematically over- or under-confident, in contrast to the overconfidence often seen in other deep learning models.

Repository Structure

├── environments/          # Environment setup scripts and requirements
├── src/                   # Source code (training, forecasting, evaluation)
└── data/                  # Datasets (see data/README.md for setup)

Setup

1. Environments

Three separate venv/conda environments are required due to incompatible dependencies between models:

EnvironmentModels / Purpose
mainTimesFM, Moirai2, Chronos-Bolt, evaluation & metrics
yinglongYinglong
tirexTiRex

To create all three environments:

cd environments
bash setup_all_envs.sh

2. Data

See data/README.md for instructions on obtaining all datasets.

Usage

See src/README.md for instructions on training heads, running forecasts, and computing metrics.

Citation

@inproceedings{
adler2026beyond,
title={Beyond Accuracy: Are Time Series Foundation Models Well-Calibrated?},
author={Adler, Coen and Chang, Yuxin and Draxler, Felix and Abdi, Samar and Smyth, Padhraic},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=nGBN7UjHcy}
}

License

MIT License