Time series models cut carbon waste with pruning and reuse
Sustainability of TSFM Deployment and Pretraining: From Carbon Footprint Analysis to Mitigation via Pruning
Machine Learning
Summary
Large time series models used for forecasting need a lot of energy to train at first, which raises concerns about their environmental impact. The authors show that if these models are used a lot or for big tasks, the initial energy cost pays off over time compared to smaller models that need retraining often. They also find a way to reduce energy use during forecasting by removing a small part of the model without hurting accuracy. This means it’s possible to keep using these big models more sustainably.
What this means in practice
- •For industrial data engineers: Choose and optimize forecasting models that balance accuracy and carbon cost for large-scale, frequently updated time series deployments.
- •For machine learning platform teams: Implement pruning techniques to reduce the runtime energy consumption and latency of pretrained time series foundation models without losing performance.
Authors
Cécile Luc, Nathanaël Lemonnier, Ievgen Redko
Abstract
Time Series Foundation Models (TSFMs) offer strong zero-shot forecasting capabilities, but their high pretraining cost raises a practical sustainability question: does this upfront investment pay off over a full deployment lifecycle, compared to supervised standalone models that must be periodically retrained? We model the cumulative carbon cost of different approaches across a realistic industrial scenario: Chronos-2 amortizes its pretraining cost within months for large or frequently queried deployments, but a classical model remains cheaper indefinitely for small or rarely queried ones. For cases where reducing the TSFM's own cost matters, we show that removing a single encoder block (7.9% of parameters) via a gradient-free pruning criterion yields a consistent 6 to 9% reduction in FLOPs, CO2 emissions, and latency, at essentially no cost to zero-shot forecasting performance across a 25-task benchmark.