Frequency Matters: When Time Series Foundation Models Fail Under Spectral Shift

2025-11-06

AI Summary

Time-series foundation models (TSFMs) achieve strong performance on public benchmarks but suffer significant generalization degradation in industrial settings. This work identifies spectral shift—i.e., mismatch between the dominant frequency components in downstream tasks and those in pretraining data—as a critical bottleneck. Method: We systematically validate this hypothesis through real-world industrial experiments on mobile game user engagement forecasting and controlled synthetic studies. Contribution/Results: We are the first to formally define, quantify, and empirically demonstrate the detrimental impact of spectral shift on TSFMs’ downstream performance—showing that such mismatch causes systematic degradation, even underperforming domain-adapted baselines. Our findings underscore the necessity of frequency-aware modeling and motivate a new pretraining paradigm and evaluation protocol explicitly designed for spectral diversity. Experiments confirm that explicit spectral characterization substantially improves cross-domain generalization, providing both theoretical grounding and practical guidance for deploying TSFMs in real-world applications.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Cognitive Modeling & Cognitive Systems: Other Foundations of Cognitive Modeling & Systems
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• User Modeling, Personalization and Recommendation: Practical large-scale studies of user experience
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Time series foundation models (TSFMs) have shown strong results on public benchmarks, prompting comparisons to a"BERT moment"for time series. Their effectiveness in industrial settings, however, remains uncertain. We examine why TSFMs often struggle to generalize and highlight spectral shift (a mismatch between the dominant frequency components in downstream tasks and those represented during pretraining) as a key factor. We present evidence from an industrial-scale player engagement prediction task in mobile gaming, where TSFMs underperform domain-adapted baselines. To isolate the mechanism, we design controlled synthetic experiments contrasting signals with seen versus unseen frequency bands, observing systematic degradation under spectral mismatch. These findings position frequency awareness as critical for robust TSFM deployment and motivate new pretraining and evaluation protocols that explicitly account for spectral diversity.

Problem

• TSFMs struggle with generalization due to spectral shift in time series data
• Spectral mismatch between pretraining and downstream tasks degrades model performance
• Frequency awareness is critical for robust deployment of time series foundation models

Innovation

• Identifies spectral shift as generalization failure cause
• Proposes frequency-aware pretraining protocols for robustness
• Introduces spectral diversity evaluation for time series models