[KongchangAI]
· 2 min read· 1,230 words

Sampling Frequency Invariance of Time Series Foundation Models: Quantitative Evaluation and Key Findings

Sampling Frequency Invariance of Time Series Foundation Models: Quantitative Evaluation and Key Findings

The study is the first to evaluate four time series foundation models with sampling frequency invariance as the core, proposing FSI, FDI, and CDI metrics.

This study addresses sampling frequency (SF) variation faced by time series foundation models (TSFMs) in sensor applications, stress-testing MOMENT, Chronos-2, Moirai-MoE, and TimesFM. For the first time it makes SF invariance the core of evaluation and proposes three quantitative metrics: FSI (embedding shift), FDI (forecast degradation), and CDI (downstream task degradation). It finds that SF invariance is not a fixed model property but depends on the margin between the signal's native sampling frequency and the task-critical bandwidth, with much degradation due to information loss rather than model capability. It thus redefines SF invariance as a measurable (signal × TSFM) pair property, guiding model selection and deployment in sensor scenarios.

Time series foundation models (TSFMs) are increasingly used in sensor applications, where sampling frequencies (SF) vary widely across sensors, devices, and operating modes (e.g., power-saving modes). Ideally, as long as the sampling frequency preserves the task-critical bandwidth, a TSFM should remain effective under such SF variations. But does this assumption hold? A new study stress-tests four mainstream TSFMs, making SF invariance the core of evaluation for the first time and proposing three quantitative metrics.

Study Overview

Why Sampling Frequency Invariance Matters

Sensor data is inherently multi-sampling-rate: sampling frequencies can differ significantly across devices and power modes. When deploying TSFMs, if a model is sensitive to SF changes, performance on prediction or downstream tasks may drop sharply in real-world scenarios (e.g., after a device switches to power-saving mode and lowers its sampling rate).

Traditional evaluations typically focus on model performance at a fixed sampling rate, but rarely systematically examine robustness to SF changes. This study fills that gap by defining SF invariance as a measurable, signal-aware property rather than a static, intrinsic characteristic of the model.

Time series foundation models (TSFMs) are models pretrained on large-scale, cross-domain time series data that can be transferred to downstream tasks such as forecasting, classification, and anomaly detection in a zero-shot or few-shot manner. Unlike specialized models trained for a single task, TSFMs attempt to learn general time series representations, so their inputs typically need to be unified to a fixed length or fixed sampling interval. Common representatives include MOMENT based on masked reconstruction, Chronos based on a language model architecture, Moirai-MoE with a mixture-of-experts structure, and TimesFM based on a decoder architecture. Understanding these models' pretraining data distributions and input assumptions is a prerequisite for judging their sampling frequency robustness: if the pretraining corpus is dominated by a certain sampling frequency, the model may implicitly rely on temporal patterns at that frequency.

Stress Test Design and Three New Metrics

The study selected four prominent TSFMs: MOMENT, Chronos-2, Moirai-MoE, and TimesFM. The stress test unfolds along three dimensions:

  • Embedding shift: the degree of change in the model's internal representations when SF changes, quantified by FSI (Frequency Shift Invariance).
  • Forecast performance degradation: the decline in model performance on forecasting tasks, quantified by FDI (Forecast Degradation Index).
  • Downstream task performance degradation: the decline in model performance on specific downstream applications, quantified by CDI (Cognitive Degradation Index).

Together, these three metrics characterize model behavior under different SFs and help distinguish whether performance degradation stems from information bandwidth loss or from limitations of the model itself.

Sampling frequency (SF) refers to the number of times a continuous signal is sampled per unit time, usually expressed in hertz (Hz) or "samples per minute." According to the Nyquist sampling theorem, to reconstruct a signal with bandwidth B without distortion, the sampling frequency must be at least 2B. In sensor scenarios, lowering the sampling frequency directly reduces the retainable signal bandwidth: when the new sampling frequency remains above twice the task-critical bandwidth, information loss is usually negligible; once it approaches or falls below this threshold, high-frequency components are aliased or lost, and no model can recover them. Therefore, when discussing the SF invariance of TSFMs, one must distinguish between two sources of degradation: "irreversible bandwidth loss at the information level" and "failure of the model to adapt to the new frequency at the model level."

Embedding shift measures the drift in a model's internal representations as the input sampling frequency changes. Specifically, the same signal segment is resampled at different frequencies and fed into the model; its encoder outputs (embedding vectors) are extracted, and the distance or directional change between embeddings is computed. A smaller embedding shift indicates that the model is less sensitive to frequency changes and that its internal features are more stable. This dimension is complementary to forecast performance: a model may have small prediction error while its embeddings have already drifted significantly, hinting at fragility on out-of-distribution inputs.

Key Finding: Invariance Depends on the Bandwidth Margin Between Signal and Task

The results show that SF invariance is not a fixed property of a model but depends on the margin between the signal's native SF and the bandwidth required by the task. When the native SF is far above the task-required bandwidth, lowering the SF may not significantly affect performance; but when the margin is small, even a slight decrease in SF can cause noticeable degradation.

Moreover, the choice among different TSFMs explains part of the invariance differences, but a large proportion of performance degradation reflects information loss rather than insufficient model capability. This means that practitioners deploying TSFMs need to evaluate SF invariance in light of signal characteristics and task requirements, rather than relying solely on model performance on standard benchmarks.

Task-critical bandwidth refers to the range of signal frequencies that must be preserved to complete a specific downstream task. For example, human activity recognition may require only 0–5 Hz of bandwidth, whereas mechanical vibration fault diagnosis may require hundreds of hertz. The bandwidth margin is the gap between the signal's native sampling frequency and the upper limit of the task-critical bandwidth. The larger the margin, the less likely task-relevant information is damaged when the sampling frequency is lowered, and the more stable the model's performance; the smaller the margin, the more any downsampling may directly cut off the high-frequency components required by the task. This concept turns SF invariance from an intrinsic model property into a system property jointly determined by signal, task, and model, and also explains why the same model exhibits vastly different frequency robustness across datasets.

Implications for Practitioners

This study redefines SF invariance as a measurable property of a (signal × TSFM) pair, offering a new perspective for model selection and deployment in sensor applications. Practitioners should pay attention to:

  1. The margin between the task-critical bandwidth and the signal's native SF, which is the key factor determining SF invariance.
  2. The degradation patterns of different TSFMs under SF changes, so as to choose a model that matches the signal characteristics.
  3. Incorporating metrics such as FSI, FDI, and CDI into evaluation to comprehensively quantify a model's robustness to SF changes.

As TSFMs become widespread in IoT, wearable devices, and other domains, sampling frequency invariance will become an indispensable part of model evaluation. This study lays a methodological foundation for future work and reminds developers to pay attention to changes in signal sampling conditions in real-world deployment.

One-line Summary

This study is the first to evaluate four time series foundation models with sampling frequency invariance as the core, proposing the FSI, FDI, and CDI metrics.

Paragraph Summary

Addressing the challenge of sampling frequency (SF) variation that time series foundation models (TSFMs) face in sensor applications, this study stress-tests four mainstream models—MOMENT, Chronos-2, Moirai-MoE, and TimesFM—making SF invariance the core of evaluation for the first time and proposing three quantitative metrics: FSI (embedding shift), FDI (forecast degradation), and CDI (downstream task degradation). The study finds that SF invariance is not a fixed property of a model but depends on the margin between the signal's native sampling frequency and the task-critical bandwidth; a large proportion of performance degradation stems from information loss rather than insufficient model capability. Based on this, the study redefines SF invariance as a measurable property of a (signal × TSFM) pair, providing methodological guidance for model selection and deployment in sensor scenarios.

#time series foundation models#sampling frequency invariance#TSFM#sensor data#MOMENT#Chronos-2#Moirai-MoE#TimesFM#FSI#FDI#CDI
Share:

Related articles