There is No "apple" in Timeseries: Rethinking TSFM through the Lens of Invariance

Arian Prabowo, Flora D. SalimarXiv 2025 · arXiv · OpenReview

TL;DR

This paper argues that current Timeseries Foundation Models (TSFMs) underperform because they're trained on datasets that lack the semantic completeness of text and image data. The authors propose a shift from "scraping everything" to principled dataset design, focusing on systematically covering the space of temporal invariances to enable true generalization and emergent reasoning.

What Problem Does This Paper Address?

The paper tackles the surprising observation that despite the rise of large-scale Timeseries Foundation Models (TSFMs), they often fail to outperform simpler supervised baselines. The authors attribute this to a fundamental difference between time-series data and the modalities (text and vision) where foundation models have excelled. While the internet is rich with text and images representing countless concepts (like "apple"), time-series data is inherently complementary, capturing dynamics and signals that aren't easily represented otherwise. This means a "scrape everything" approach, which works for text and vision by capturing a vast semantic space, fails for time-series because there's no equivalent of a universally understood "apple" concept within time-series data itself. This incompleteness prevents TSFMs from developing the common-sense reasoning and generalization capabilities seen in other domains.

Methods

The paper doesn't introduce new algorithms or models but rather proposes a new paradigm for TSFM development. The core idea is to move away from opportunistic data aggregation and towards principled dataset construction. This involves:

  • Identifying and formalizing an ontology of timeseries invariances: The authors suggest that understanding the transformations that preserve the semantic identity of time-series signals is key. They propose a framework for systematically generating these invariances from primitive transformations (e.g., time-axis shifts, value-space scaling, index-space permutations).
  • Principled Dataset Design: Based on this invariance ontology, datasets should be curated and potentially synthetically generated to systematically span the space of these invariances. This ensures representational completeness, allowing models to learn robust, generalizable representations.

Datasets

The paper discusses existing TSFM pretraining datasets like the Time Series Pile, LOTSA, and UTSD, noting they are assembled through opportunistic aggregation rather than systematic design. Examples of data sources mentioned include the Monash Forecasting Archive, Informer datasets, Google Trends, JD supply-chain sales, and telemetry data. The paper does not introduce a new dataset but advocates for the creation of datasets guided by the proposed invariance ontology.

Key Insights and Ideas

  • The "Apple" Analogy: The core insight is that the "scrape everything" paradigm works for text and vision because the internet is semantically complete for these modalities. Time-series data, however, is fundamentally different and lacks this universal semantic coverage, meaning there's no "apple" equivalent in time-series.
  • Complementary Modalities: Time-series data complements, rather than overlaps with, image and text data. This means foundation models trained solely on time-series data will miss crucial aspects of real-world understanding.
  • Invariance as a Guiding Principle: The paper posits that understanding and systematically covering the space of timeseries invariances is the principled way to achieve representational completeness. This is crucial for enabling generalization, reasoning, and emergent behaviors in TSFMs.
  • Beyond Scale: Progress in TSFMs won't come from simply scaling up datasets but from designing them to cover the diverse invariances that define temporal dynamics.

Key Results

The paper is primarily theoretical and conceptual, so it doesn't present quantitative results in the traditional sense. However, the key "result" is the argument that current TSFMs underperform due to dataset incompleteness, and the proposed solution of using an invariance ontology for principled dataset design. The authors imply that following this paradigm will lead to TSFMs with improved generalization, reasoning, and emergent behavior, though specific metrics are not provided.

Limitations and Future Work

  • Limitations: The paper acknowledges that identifying the complete set of timeseries domains that approximate a "world-complete" representation is an open research question. The proposed invariance ontology is a starting point, and its practical implementation for dataset curation and model design requires further work.
  • Future Work: The authors suggest that the next steps involve systematically generating this invariance ontology from first principles and using it to guide the construction of new, comprehensive timeseries datasets. This will involve defining primitive transformations and closing these sets under composition to form a complete invariance group for temporal data. This, in turn, will guide both dataset curation and model design for future TSFMs.

Summary updated

Abstract

Timeseries foundation models (TSFMs) have multiplied, yet lightweight supervised baselines and even classical models often match them. We argue this gap stems from the naive importation of NLP or CV pipelines. In language and vision, large web-scale corpora densely capture human concepts i.e. there are countless images and text of apples. In contrast, timeseries data is built to complement the image and text modalities. There are no timeseries dataset that contains the concept apple. As a result, the scrape-everything-online paradigm fails for TS. We posit that progress demands a shift from opportunistic aggregation to principled design: constructing datasets that systematically span the space of invariance that preserve temporal semantics. To this end, we suggest that the ontology of timeseries invariances should be built based on first principles. Only by ensuring representational completeness through invariance coverage can TSFMs achieve the aligned structure necessary for generalisation, reasoning, and truly emergent behaviour.
BibTeX
@misc{prabowo2025there,
  author = {Arian Prabowo and Flora D. Salim},
  title = {{There is No "apple" in Timeseries: Rethinking TSFM through the Lens of Invariance}},
  year = {2025},
  eprint = {2510.20119},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.20119},
}

Loading related content...

Figure