Benchmarking Time Series Foundation Models on their Accuracy and Energy Consumption

Loïc Guibert, Benjamin Pasquier, Frédéric Montet, Beat Wolf

Keywords: Machine Learning, Time Series, Energy Consumption, Sustainable AI

TL;DR: We benchmark ten time-series foundation models to evaluate the accuracy–energy trade-off in zero-shot forecasting, finding that prediction accuracy is highly dependent on the specific dataset, and energy efficiency is driven by model architecture.

Abstract: Our study presents a benchmark of ten time-series foundation models to quantify their accuracy–energy trade-off in zero-shot forecasting. Using an in-house and a public dataset (School and MeteoSwiss; univariate and multivariate variants), a fixed sliding-window protocol (context 512, horizon 64), and dual energy instrumentation (external PDU and CodeCarbon), we report sMAPE and NMAE accuracy metrics alongside runtime, energy ($Wh$), and Energy per Billion Parameters. Results show pronounced dataset dependence in accuracy, while efficiency is primarily architecture-driven: Chronos-Bolt achieves consistently low energy and latency, TimesFM attains the best MeteoSwiss accuracy at low energy cost, and Moirai-MoE exhibits substantially higher energy expenditure for comparable errors. This work informs decision-makers, developers, and end-users about the energy requirements of time-series foundation models and highlights the importance of considering energy alongside accuracy when evaluating models for adoption, while encouraging the systematic reporting of accuracy–energy trade-offs.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 9