Beyond Earthbound Benchmarks: Evaluating Time Series Foundation Models on Satellite Telemetry Data

Geremia Pompei, Deepan Anbarasan

Presentation Attendance: No, we cannot present in-person

Keywords: Time Series Foundation Models, Zero-Shot Forecasting, Satellite Telemetry, Anomaly Detection, Space Operations, ESA Benchmark, Chronos-2, Credence

TL;DR: We present the first zero-shot evaluation of Time Series Foundation Models on ESA satellite telemetry.

Abstract: Time series foundation models (TSFMs) excel on terrestrial benchmarks but remain untested in the space domain. We present the first systematic zero-shot evaluation of four TSFMs - Credence, TimesFM-2.5, Chronos-2, and TiRex - on the ESA Anomaly Detection Benchmark. Results demonstrate generalization, with most models outperforming seasonal naive baselines. While probabilistic accuracy (CRPS) improves at lower sampling frequencies, point accuracy (MASE) degrades significantly. Overall, **Chronos-2** achieves the strongest point accuracy (leading MASE), while **Credence** provides superior probabilistic performance (leading CRPS). These findings validate TSFMs for automated mission operations and establish a baseline for future space forecasting benchmarks.

Track: Industry and Application Track (max 2 pages)

Submission Number: 112