Performance of Zero-Shot Time Series Foundation Models on Cloud Data
William Toner, Thomas L. Lee, Artjom Joosen, Rajkarn Singh, Martin AsenovSystems Infrastructure Lab, Edinburgh Research Centre, Huawei [email protected], [email protected] · 2025
Read Performance of Zero-Shot Time Series Foundation Models on Cloud Data by William Toner, Thomas L. Lee, Artjom Joosen, Rajkarn Singh, Martin AsenovSystems Infrastructure Lab, Edinburgh Research Centre, Huawei [email protected], [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-12944.txt.
Description
William Toner and four co-authors tested zero-shot forecasting models on cloud computing data in a 2025 paper. Time series foundation models are trained on many domains and sold as instant forecasters everywhere. Cloud load data, with its spikes and shifts, is a hard proving ground.
Well-known models failed to give meaningful or accurate forecasts on this data. Simple linear baselines beat them consistently. The team also caught odd behavior: models suddenly emitting wild, random-looking forecasts mid-stream. Their verdict is blunt: a broad failure of foundation models on cloud series, at least without tuning. The paper warns buyers to test such models on their own data before trusting the marketing.
Catalog record
| CREATOR | William Toner, Thomas L. Lee, Artjom Joosen, Rajkarn Singh, Martin AsenovSystems Infrastructure Lab, Edinburgh Research Centre, Huawei [email protected], [email protected] |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-12944