acceptodds
Under review as a conference paper at ICLR 2027

FractalTS: Time Series Dataset for Controlled Scaling Dynamics and Multifractal Processes

Abstract

A major challenge in time-series machine learning (ML) is separating genuine learning of temporal structure from successful matching of low-order statistics. This problem is particularly important for fractional, fractal, and multifractal (MF) processes, whose defining properties are expressed across multiple scales. Two sequences may exhibit similar marginal distributions and short-range correlations while following different scaling laws. Existing datasets are predominantly collected from natural or engineered systems, where the generating mechanism and its scaling parameters are only partially known. Thus, model failures cannot be simply separated from uncertainty in the target properties. We introduce FractalTS, a theory-grounded benchmark dataset for controlled time-series research in which process families, scaling regimes, and stochastic parameters are explicitly specified at generation time. The core dataset contains independent realizations of monofractal fractional Brownian motion and six multifractal constructions spanning anti-correlated, uncorrelated, correlated, , , and non-stationary regimes. The generator exposes interpretable controls, including the Hurst exponent (), multifractal intermittency, cascade parameter, spectral slope, envelope, and drift, and supports arbitrary numbers of independently seeded realizations. Rather than evaluating a particular downstream model, this paper establishes the dataset as a reproducible scientific testbed. A reproducible manifest and train/calibration/validation/test protocol is provided. Validation covers covariance, spectral scaling, Hurst behaviour, generalized scaling functions , multifractal spectra, and stationarity. The dataset is primarily designed for controlled evaluation of time-series generative models (GANs, VAEs, normalizing flows, and diffusion models), which are rarely tested on whether they preserve multiscale structure. It also supports representation learning, forecasting, anomaly detection, and estimator benchmarking under known multiscale dynamics.

Primary area:
datasets and benchmarks
Keywords:
time series, multiscale dynamics, multifractal processes, fractional Brownian motion, Hurst exponent

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4Xy2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Datasets play a pivotal role in AI development by benchmarking existing algorithms and pushing the field toward more robust models. While the theoretical groundings of FRACTALTS are well-conceived, the manuscript relies too heavily on intuitive motivations. To fully support its claims, the work requires empirical validation showing how this dataset challenges state-of-the-art models in ways that existing benchmarks cannot. I am happy to increase the score if the responses are strongly convincing.

Critical strengths

Current time-series generative models are primarily evaluated using marginal distribution metrics, train-on-synthetic scores, or short-range correlations. FRACTALTS provides a controlled testbed where scaling laws, intermittency, and long-range dependence are known by construction, turning quiet generative failure modes (such as flattening singularity spectra) into quantifiable metrics.

Critical weaknesses
  1. Lack of Empirical ML Model Benchmarking and Diagnostic Proof: While the dataset generation and validation pipeline are well-formulated, the paper includes no empirical experiments evaluating existing machine learning models (e.g., generative time-series models like TimeGAN, TimeDiffusion, or structural time-series baselines). To establish the utility of a new benchmark, it is crucial to empirically demonstrate:

    • Where current models fail: Showing that state-of-the-art models perform well on standard benchmarks but fail to preserve long-range dependencies, multiscale scaling laws, or multifractal spectrum width (Δα\Delta \alpha) on FRACTALTS. This breakdown is essential to motivate future algorithmic developments.
    • Superiority over existing synthetic/real benchmarks: Beyond intuitive motivations, the paper needs concrete empirical comparisons against standard time-series datasets. For instance, quantitatively mapping the range of autocorrelation (short-range vs. long-range) across real-world datasets and showing how FRACTALTS spans these regimes where prior benchmarks fall short.
  2. Incomplete Grounding of Real-World Realism and Omission of Related Evaluation Frameworks:

    • Real-World Fidelity: The manuscript relies heavily on the mathematical elegance of fractal theory, but lacks explicit empirical validation demonstrating how well these synthetic processes mirror complex real-world dynamics across key application domains (e.g., finance, telemetry, or physiological signals).
    • Incomplete Contextualization of Advanced Non-Marginal Metrics: The motivation criticizes standard "marginal distribution metrics" for failing to capture temporal structure. However, the paper overlooks recent non-marginal and dynamic evaluation paradigm, such as geometry-aware metrics designed to detect temporal mode collapse (e.g., A Geometry-Aware Metric for Mode Collapse in Time Series Generative Models, NeurIPS 2025). Clarifying how FRACTALTS complements or surpasses these geometric/topological metrics is necessary to properly position the benchmark's contribution.
  3. Absence of Hyperparameter Sensitivity and Robustness Analysis: The parameter generation and MFDFA rejection sampling pipeline rely on several key design choices (e.g., segment lengths, moment orders qq, sample length TT, and tolerance thresholds for H^\hat{H} and Δα\Delta \alpha). The authors do not provide a systematic sensitivity analysis evaluating how sensitive the benchmark's properties and computational efficiency are to these hyperparameters. Demonstrating how variation in sequence length (TT) or estimator parameters impacts the yield rate and theoretical guarantees of the accepted sequences is necessary to establish the robustness of the dataset generation protocol.

Questions for authors

Can you include baseline results showing how standard generative models (e.g., TimeGAN, TimeDiffusion)?

Reviewer_aQUtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper points out a major challenge in recent time series ML studies that standard evaluations of deep generative models do not directly certify whether the models capture the inherent scaling laws within data. Thereby, existing metrics fail to analyze whether the models learn important temporal structures, such as multiscale fractal behaviors. To address this, the authors introduce a controlled benchmark dataset, FRACTALTS, for fractal time series, including seven stochastic families. They further specify generation pipeline, independent finite-sample calibration, validation protocols for testing the scaling properties, and dataset quality gates. The paper also discusses the scientific uses and compares FRACTALTS with several existing datasets by analyzing their differences, but does not benchmark generative models.

Critical strengths
  1. This paper correctly identifies that current evaluation metrics for time series generation care more about marginal distributions and overlook specific properties (e.g., scaling laws) within time series, which cannot verify generative models’ capability to fully mimic the appropriate data-generating process. It also identifies multiscale fractal behaviors as a critical property, which are underexplored in the literature of learning-based time series generation.

  2. The proposed FRACTALTS benchmark dataset is a reproducible process to generate time series data with an interpretable control over various fractal temporal dynamics.

  3. The rejection sampling as a per-realization certification mechanism for a time-series benchmark is novel. This improves the overall quality of generated datasets.

Critical weaknesses
  1. Real-world time series data exhibit many more properties than the discussed fractal behaviors. Although this paper positions FRACTALTS dataset as a complementary layer, it does not provide a clear roadmap for incorporating other critical properties beyond fractals. This narrow scope limits the dataset’s immediate usefulness for model training on realistic data and for broader downstream analysis.

  2. I acknowledge that this paper provides some interesting insights regarding the controlled dataset for fractal properties in time series. However, several critical evaluations, as detailed below, remain unexplored. They should have been central to the proposed benchmark. Thus, the main contributions remain somehow conceptual.

  3. The paper motivates FRACTALTS as a foundation for evaluating time-series generative models. However, this paper does not benchmark any generative models, and leaves this as a future work. Without such empirical validation, the potential and limitations of the proposed dataset remain unproven. Applying the proposed evaluations to determine whether, and how well, existing generative models capture multiscale fractal properties is necessary and also a demanding study for the time series community.

  4. This paper does not provide any downstream ML validations as well. Some scientific uses are listed but not empirically demonstrated.

Questions for authors
  1. How can the proposed benchmark be used to test whether a generative model, which performs well under existing metrics, fails to preserve scaling fidelity?

  2. What architecture, objective, or training pipeline should a deep generative model consider to capture fractal properties? Do the controlled principles proposed in this paper offer useful guidance for such designs?

  3. What are the bottlenecks to extend the FRACTALTS generation and validation pipeline to multivariate settings with cross-channel dependence and joint multifractality?

  4. Can the FRACTALTS generated data improve the performance of existing ML models in downstream applications or in simulated scenarios? If the downstream models are not specifically designed to capture the fractal behaviors, how well would they perform when augmented with the FRACTALTS data?

  5. Can the sequence length be set to an arbitrary horizon to match datasets in specific domains, since different datasets have varying time horizons?

  6. How can we rigorously determine the exact values of each family’s acceptance criterion? If the acceptance criteria are slightly modified, how would the test results change? When a new family is added to the dataset, is there a natural acceptance range?

Reviewer_dZeJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FractalTS, a synthetic time-series benchmark dataset generated from controlled stochastic processes whose generating mechanisms and parameters are fully specified at construction time. The dataset comprises seven families: a monofractal reference family built from fractional Brownian motion, and six multifractal families constructed through multiplicative cascades, covering anti-correlated, uncorrelated, correlated, pink-noise, Brownian-noise, and non-stationary scaling regimes. Each candidate sequence is checked with multifractal detrended fluctuation analysis after generation, and is retained only if its scaling exponent and singularity-spectrum width fall within the range prescribed for its family, so the labels in the dataset are empirically certified rather than merely nominal. The paper also provides a reproducible manifest, a four-way data split, and a structural validation protocol, with the aim of offering a benchmark with known ground truth and reproducible use for time-series generative models and scaling estimators, although the paper itself does not evaluate any specific model.

Critical strengths
  1. The paper is reasonably convincing. Through controlled data, it turns a failure mode that could not previously be tested into one that can be measured. Because the ground truth of the benchmark is fixed by the construction process, any deviation between model output and target can be attributed to the model itself.
  2. The authors set a strict quality-control and reproducibility procedure. At the generation stage each sequence is verified individually, which substantially increases computational cost but largely ensures that the generated data meet the requirements. The authors also specify explicit splitting rules, thereby avoiding the leakage that arises when a single long series is cut into segments.
Critical weaknesses
  1. The authors claim that generative models can perform well on marginal distributions and short-range statistics while potentially distorting long-range dependence and spectral color without this being noticed. However, the paper does not quantify how serious this problem actually is. For instance, how SOTA generative models would behave.
  2. The paper does not evaluate any generative model. The authors state explicitly that they do not benchmark specific generative models and instead list this work as a future direction, so there are no baselines, no metric definitions, and no scores to compare against. The benchmark is sound as a dataset, but it has not yet been delivered as an executable evaluation protocol, which limits its immediate usefulness to the methods community.
Questions for authors

Which real-world scenarios do the distributions covered by the seven families correspond to?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.