acceptodds
Under review as a conference paper at ICLR 2027

PanelTS: A Multi-Domain Benchmark Dataset for Panel-Based Time Series Forecasting

Abstract

Many real-world forecasting problems involve multiple related units observed over time, each with its own target variable and associated covariates - a structure known as panel data. Yet existing time series benchmarks are predominantly organised as collections of independent univariate series or single-entity multivariate recordings, leaving panel-structured forecasting systematically underserved. We introduce PanelTS, a multi-domain benchmark dataset designed to close this gap. PanelTS adopts a unified representation that natively supports univariate, multivariate, and panel-based forecasting tasks while preserving cross-unit structure, enabling rigorous evaluation of models that exploit inter-unit dependencies. The benchmark comprises 1,844 datasets, spanning synthetic data with controllable temporal dynamics and real-world datasets drawn from COVID-19 epidemiology, exchange-traded funds (ETFs), currency markets, and equity markets, thereby providing diversity in domain, granularity, and scale. To facilitate reproducibility and systematic comparison, PanelTS provides standardised data formats, task definitions, and evaluation protocols across varying input lengths, prediction horizons, and temporal resolutions. We report baseline results across representative forecasting models to establish reference performance and demonstrate compatibility with existing approaches. PanelTS is publicly released to support the development and fair evaluation of the next generation of panel-aware forecasting methods: https://huggingface.co/datasets/Multiple-Time-Series/PanelTS (data), https://anonymous.4open.science/status/PanelTS-020D (code), https://huggingface.co/spaces/Echo0623/PanelTS-ICLR-Review (demo).

Primary area:
datasets and benchmarks
Keywords:
panel-based forecasting, synthetic data, panel data, real world time series

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2JnNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PanelTS, a multi-domain benchmark dataset for panel-based time series forecasting. It adopts a unified representation that supports univariate, multivariate, and panel-based forecasting tasks from the same data source while preserving cross-unit structure. The benchmark comprises 1,844 datasets spanning synthetic data and four real-world domains (currency exchange rates, commodity ETFs, COVID-19 epidemiology, and equity markets), together with standardized data formats, task definitions, and evaluation protocols. Reference results are reported for 14 representative forecasting models on two illustrative scenarios (currency exchange rates and COVID-19), using a 100-step input, one-step-ahead, rolling-window protocol with both raw and scale-normalized RMSE/MAE metrics.

Critical strengths
  1. The problem framing is legitimate and fills a real gap. The paper observes that existing forecasting benchmarks are organized either as collections of independent univariate series or as single-entity multivariate recordings, neither of which captures the "multiple related units with shared covariate definitions" structure common in real-world forecasting.

  2. The unified data schema is a reasonable engineering contribution. Organizing targets and covariates so that the same data source can support univariate, multivariate, and panel tasks is clean and practically useful, and the scale-normalized error metrics address a real comparability problem across heterogeneous units.

Critical weaknesses
  1. The benchmark's central claim is not empirically demonstrated. The paper's stated purpose is to enable the study of cross-unit dependencies, but (a) only two of the five dataset groups are evaluated; (b) nearly all baselines run in a target-history-only mode, so the panel structure and shared covariates the benchmark is built around are essentially never exercised; and (c) there is no panel-specific baseline or controlled ablation showing that exploiting the panel structure improves forecasting.

  2. The independent contribution is thin relative to existing work. The core real-world data (exchange rates and COVID-19) overlaps substantially with the contemporaneous ESE work (Xu et al., 2026), and is largely repackaged from public sources; the synthetic generator's linear additive data-generating process is too simple to stress-test the nonlinear, time-varying cross-unit dependencies of real panel data.

The paper reads more as a repackaging/extension of the ESE data effort than as a self-contained benchmark contribution.

Questions for authors
  1. Could you provide a controlled comparison demonstrating that exploiting the panel structure improves accuracy?
  2. Please clarify the relationship between PanelTS and the data used in ESE (Xu et al., 2026).
  3. Why are the ETF and stock-market datasets absent from the reference experiments?
Reviewer_AFvaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PanelTS, a benchmark for panel-based time series forecasting that preserves unit identities and associated covariates. Experiments on real-world datasets provide extensive results across a range of forecasting methods and illustrate the practical applicability of this benchmark.

Critical strengths

S1. The organization is easy to follow. S2. Experiments on real-world datasets demonstrate the practical applicability of this benchmark.

Critical weaknesses

W1. The references contain citation mismatches and unedited template text or placeholders. W2. The evaluation results contain inconsistencies. Table 17 reports several cases where MAE exceeds RMSE. These discrepancies require clarification and verification. W3. The motivation does not sufficiently establish the need for a new benchmark. While the paper emphasizes preserving panel structure, it does not clearly identify the specific limitations of existing benchmarks or demonstrate the additional evaluation capabilities offered by PanelTS. W4. The task protocol is too broadly described to support reproducibility. Some key details are missing.

Questions for authors

See above

Reviewer_cMcjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces PanelTS, a benchmark for panel-based time-series forecasting, where multiple related units have their own target series and covariates. The authors argue that existing benchmarks mainly focus on independent time series or multivariate data from a single entity, and do not adequately preserve cross-unit structure.

The main contributions are: (1) a unified data format supporting univariate, multivariate, and panel-based forecasting; (2) a collection of 1,844 synthetic and real-world datasets covering COVID-19, currency exchange rates, ETFs, and stock markets; and (3) standardized forecasting protocols and baseline results.

For validation, the authors evaluate 14 forecasting methods on currency exchange rates and COVID-19 datasets, using a rolling-window setup with 100 input steps and a one-step-ahead prediction. Results show that different models perform differently across datasets and panel sizes. However, most baselines do not explicitly exploit cross-unit dependencies, and the experiments mainly demonstrate dataset usability rather than fully validating the benefits of panel-aware forecasting.

Critical strengths

Useful unified framework: PanelTS preserves cross-unit structure and supports univariate, multivariate, and panel-based forecasting within one framework. This may help researchers systematically study panel-aware forecasting.

Diverse datasets and reproducibility: The benchmark combines synthetic and real-world data from multiple domains, with standardized protocols, baseline results, and public code. This provides a useful starting point for future research and fairer comparisons.

Critical weaknesses

I appreciate the motivation to build a benchmark that organizes panel-based forecasting datasets under a unified framework. However, I expected the synthetic data generation to be a stronger contribution, and I am not convinced that the current design provides sufficient value beyond generating time series with basic temporal patterns.

Panel-based forecasting is a general data structure, but the underlying relationships between targets, covariates, and observational units can be highly domain-specific. For example, traffic forecasting depends on road topology and traffic propagation, economic forecasting is affected by market mechanisms and macroeconomic shocks, and semiconductor wafer quality forecasting can involve complicated physical, chemical, and process-control relationships between variables such as temperature, pressure, and impedance. A common data representation does not imply that these domains share the same underlying dynamics or that the same modeling assumptions are appropriate.

The proposed synthetic generator (Section 4.1) combines piecewise trends, Fourier-based seasonality, and stochastic residuals to construct covariates, and then generates targets through an additive covariate-driven equation. While this is useful for controlled experiments, it does not explicitly capture more complex mechanisms such as nonlinear covariate interactions, delayed effects, regime-dependent relationships, or dynamic dependencies between units. In particular, the paper describes the unit-specific term (c_i) as capturing cross-unit dynamics, but the stated equation does not clarify how dynamic dependence between different units is generated.

More importantly, the paper does not demonstrate that the synthetic data can meaningfully distinguish panel-aware forecasting methods from methods that forecast each unit independently. I would encourage the authors to clarify the intended scope of the generator and provide controlled experiments that systematically vary cross-unit dependence and target–covariate relationships. Such experiments would help establish whether the synthetic component provides a useful evaluation tool for studying the benefits and limitations of panel-aware forecasting, rather than merely providing additional datasets with basic temporal patterns.

Questions for authors
  1. What specific cross-unit dependencies does the synthetic generator produce, beyond unit-specific offsets and covariate correlations?

  2. Can the authors demonstrate controlled conditions under which exploiting cross-unit information improves forecasting, and conditions under which it does not?

  3. What aspects of the synthetic data are intended to generalize across domains, and what domain-specific mechanisms are explicitly outside the generator's scope?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.