Beyond Point Forecasts: A Survey on Probabilistic Forecasting for Time Series and Spatiotemporal Data
Donia Besher, Rajdeep Pathak, Madhurima Panja, Tanujit Chakraborty
stat.ML · cs.LG · econ.EM · stat.AP
Quick verdict
Survey unifying probabilistic forecasting from stats to foundation models; strong taxonomy and guidance, but headline empirical claims unverifiable here.
Summary
This survey organizes probabilistic forecasting methods for time series and spatiotemporal data into a two-branch taxonomy based on where uncertainty enters the pipeline: model-agnostic approaches (pre-control limits, resampling/randomization/deep ensembles, Bayesian model averaging, conformal prediction and conformalized quantile regression; Section 3) and model-intrinsic approaches (Bayesian modeling including Gaussian processes, Dirichlet processes, BART and Bayesian neural networks with MCMC, MC-dropout, variational and INLA inference; parametric predictive distributions and mixture density networks; quantile regression, engression and loss-driven interval estimation; nine generative families from copulas to diffusion models; Section 4), plus a dedicated treatment of probabilistic time series foundation models such as Chronos-2, Moirai, TimesFM-2.5, TiRex and Sundial (Section 5). Comparison tables characterize each family by coverage guarantees, assumptions, likelihood tractability and computational cost (Tables 3-8). Section 6 compiles datasets, software, evaluation metrics, significance testing and two model-selection frameworks (data-driven and domain-specific). The cross-paradigm empirical study (Section 7) reportedly shows that no single paradigm dominates and that calibration, sharpness, accuracy and efficiency yield different model preferences, with generative and zero-shot foundation models trading accuracy against cost. Open challenges cover evolving dependencies, physics-informed distributions, extremes, count/directional/continuous-time data and unified software (Section 8). Limitations: per-method depth is brief, many cited works are 2025-2026 preprints, and the empirical protocol could not be verified from the provided text.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 1
● low 3
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single taxonomy based on where/how uncertainty enters the forecasting pipeline can unify probabilistic forecasting from pre-control limits and conformal methods through generative models to foundation models (Secs 3-5).
- No existing survey jointly covers probabilistic forecasting across temporal and spatiotemporal settings and the full methodological spectrum (Table 2, 23 prior surveys coded).
- In the cross-paradigm empirical study, no uncertainty-quantification paradigm dominates; calibration, sharpness, accuracy and efficiency criteria yield different model preferences (Sec 7, as reported in the abstract).
- Probabilistic spatiotemporal foundation models do not yet exist; combining graph foundation models with temporal foundation models is the promising route (Sec 5.2).
- Key open problems are uncertainty under evolving dependencies, physics-informed predictive distributions, extreme events, count/directional/continuous-time data, and unified software (Sec 8).
Strengths
Uncertainty-centric taxonomy with argued boundary cases
Organizes the field by where/how uncertainty enters the pipeline rather than by architecture; boundary cases are explicitly reasoned — engression stays in distributional regression despite sample-based output, copulas join generative models for their joint-dependence role (Sec 4.4 intro), QRF is flagged as straddling ensembles and distributional regression (Sec 3.2.2).
Decision-ready comparative tables
Tables 3-8 encode coverage guarantees (finite-sample for CP/CQR in Table 3), parametric assumptions, exact-likelihood tractability, sampling procedures and inference cost — e.g., Table 4 separates MCMC (high fidelity/high cost) from MC-dropout (restricted applicability, low/low) and INLA (latent-Gaussian models only).
Up-to-date foundation-model treatment
Bridges taxonomy to practice
Common-protocol empirical study
Weaknesses
Breadth over depth
Empirical evidence unverifiable from the provided text
Checkmark epistemology
Citation hygiene
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a methods survey of probabilistic forecasting for univariate/multivariate time series and spatiotemporal data (arXiv 2609.13345, stat.ML). The organizing claim: prior literature fragments across architectures, UQ-in-general, or single domains — Table 2 codes 23 surveys on five axes (probabilistic, TS, ST, architecture-agnostic, general scope) and concludes none jointly covers all — so the authors propose a taxonomy by where/how uncertainty enters the pipeline: model-agnostic (Sec 3) vs model-intrinsic (Sec 4), extended by foundation models (Sec 5). Secondary claims: two practical selection frameworks (Fig 9 data-driven; domain-specific), a cross-paradigm empirical study (Sec 7) showing 'no single paradigm dominates', and a challenge list (Sec 8: evolving dependency structures, physics-informed predictive distributions, extremes, count/directional/continuous-time data, unified software). The positioning claim is plausible but self-adjudicated: the ✓/❍/✗ grid is the authors' own coding with no stated rubric.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main contribution of this probabilistic forecasting survey?
It organizes methods — from statistical tools (pre-control limits, bootstrapping, conformal prediction, Bayesian models, quantile regression) to deep generative models and time series foundation models — into one taxonomy based on where uncertainty enters the forecasting pipeline, adding comparison tables, evaluation guidance, a cross-paradigm empirical study and an open-challenges list.
Which time series foundation models does the survey cover and how do they differ?
It covers Chronos-2 (autoregressive token sampling with a 21-quantile head), Moirai (mixture-distribution heads), Lag-Llama (Student-t distribution head), TimesFM-2.5 (added quantile head), TiRex (xLSTM predicting a 9-quantile grid from 0.1 to 0.9) and Sundial (flow-matching generative sampling), compared by probabilistic output mechanism, zero-shot and multivariate support, and spatiotemporal potential (Table 8).
Does the survey identify a single best probabilistic forecasting method?
No. Per the abstract, the empirical study finds that no uncertainty-quantification paradigm dominates across univariate, multivariate and spatiotemporal settings; the preferred model changes depending on whether calibration, sharpness, point accuracy or computational efficiency is prioritized.
Is code or a practical roadmap available?
Yes — the abstract links a public repository (github.com/PyCoder913/ProbForecasting) with code and supporting materials; Section 6.1 catalogs benchmark datasets and open-source libraries, and Section 6.4 provides data-driven and domain-specific method-selection guidance.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text ends inside Section 6.1; Sections 6.1.1-6.4, the full empirical study (Sec 7, including datasets, metrics and result tables), Challenges (Sec 8) and Conclusion (Sec 9) are beyond the excerpt, so empirical claims are assessed from the abstract and section framing only.
- Sampling gaps also skipped Table 1's middle rows, Sections 3.2.4-3.3.1, the four approximate-inference subsections of 4.1.4, and 4.4.1-4.4.3/4.4.9; family-level judgments for those rely on the surrounding visible text and the Fig 1 taxonomy tree.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.