High-Frequency Data Gap Exposes TSFM Limitations
Researchers have released a new millisecond-resolution dataset capturing real-world 5G wireless and traffic conditions to address a critical gap in time series foundation model training. Current large-scale datasets focus on low-frequency data sampled at intervals of seconds to years, limiting TSFMs' ability to handle high-frequency phenomena. This dataset introduces wireless networks as a new domain for TSFM pre-training and includes forecasting tasks with prediction horizons from 1 to 96 milliseconds. Benchmarking reveals that most TSFM configurations perform poorly on this high-frequency data in both zero-shot and fine-tuned settings, highlighting the need for high-frequency datasets during pre-training to improve generalization and robustness.
TL;DR
- →New dataset provides millisecond-resolution wireless and traffic data from operational 5G deployment, filling a gap in high-frequency time series training data
- →Introduces wireless networks as a new domain for time series foundation models, complementing existing domains like energy and finance
- →Includes short-term forecasting use cases with prediction horizons from 1 millisecond to 96 milliseconds
- →Benchmarking shows most TSFM configurations underperform on high-frequency data, indicating need for architectural and fine-tuning improvements
Why it matters
Time series foundation models are increasingly central to AI applications across infrastructure, finance, and operations, but their training data has been skewed toward low-frequency signals. This work exposes a fundamental limitation: TSFMs trained on second-to-year-scale data cannot reliably handle millisecond-scale phenomena critical to wireless networks, real-time systems, and other latency-sensitive domains. The poor zero-shot and fine-tuned performance on high-frequency data suggests that current TSFM architectures may need rethinking to handle diverse temporal frequencies.
Business relevance
Operators managing 5G networks, telecom infrastructure, and real-time systems need accurate forecasting at millisecond scales to optimize performance and prevent failures. The dataset and benchmarking results provide a concrete foundation for improving TSFM capabilities in these domains, potentially unlocking better predictive maintenance, traffic management, and resource allocation. Companies building or deploying TSFMs should expect that models trained on standard datasets will not generalize to high-frequency wireless and network data without targeted pre-training or fine-tuning.
Key implications
- →TSFM pre-training strategies need to explicitly incorporate high-frequency data to improve cross-domain generalization and avoid performance cliffs when applied to millisecond-scale phenomena
- →Wireless networks represent an underexplored domain for time series modeling, with distinct characteristics and forecasting requirements that differ from energy, finance, and other established domains
- →Current TSFM architectures may require modifications to handle the temporal dynamics and noise characteristics of high-frequency data, suggesting opportunities for model innovation
What to watch
Monitor whether major TSFM developers (OpenAI, Google, Anthropic, and others) incorporate high-frequency datasets into future pre-training runs and whether this improves zero-shot performance on wireless and network tasks. Watch for follow-up work on architectural changes or fine-tuning strategies specifically designed for high-frequency time series. Also track adoption of this dataset by the research community and whether it becomes a standard benchmark for evaluating TSFM robustness across temporal frequencies.
vff Briefing
Weekly signal. No noise. Built for founders, operators, and AI-curious professionals.
No spam. Unsubscribe any time.
