Pith. sign in

REVIEW 3 major objections 7 minor 85 references

Relational Conformal Prediction for Correlated Time Series

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A graph neural network trained on forecast residuals produces narrower prediction intervals than per-series conformal baselines for correlated time series, with no known graph required.

desk verdict New relational CP architecture with strong empirical results, but the coverage theorem assumes the very calibration it is supposed to provide. read the letter →

arxiv 2502.09443 v2 pith:KCCADWT6 submitted 2025-02-13 cs.LG cs.AI

Andrea Cini , Alexander Jenkins , Danilo Mandic , Cesare Alippi , Filippo Maria Bianchi This is my paper
classification cs.LGcs.AI MSC 62G1562M1068T07
keywords conformalpredictiontimeseriesforecastinguncertaintyquantificationgraphneuralnetworksquantileregressionintervalsspatiotemporaldatanon-exchangeable
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CoRel is a post-hoc conformal prediction method that builds confidence intervals for forecasts of many correlated time series at once. Instead of computing empirical quantiles separately for each series, it trains a graph neural network on the residuals of any already-trained point forecaster to output error quantiles for every series at every step, conditioning each interval on recent residuals at neighboring series. The paper claims this relational conditioning beats per-series conformal baselines in Winkler score (interval width plus a penalty for misses) in almost all tested settings, with coverage close to the target level, while learning the latent dependence graph from data rather than requiring it. If true, uncertainty quantification for spatiotemporal forecasting can be improved without retraining or modifying the base predictor and without a known graph.

What carries the argument

The machinery is the relational quantile predictor (RelQP), a hybrid global-local STGNN in which a shared sequence encoder processes recent residuals plus a per-series node embedding, two message-passing layers propagate information across the learned graph, and a readout maps the resulting representation together with the requested quantile level to quantile estimates. The graph structure itself is learned from residuals through a differentiable subset-sampling module: a score matrix $\Phi$ parametrizes a categorical distribution over $K$-NN graphs, sampled via Gumbel-Top-K with a straight-through gradient estimator, so no hand-specified adjacency matrix is needed. The message-passing layers let residuals at neighboring nodes shape the quantile estimates, and the node embeddings provide a cheap, per-series adaptation channel that is fine-tuned on recent data at test time.

What would settle it

Run CoRel on residuals engineered to be independent across series with per-series heteroskedastic noise: if the learned graph cannot reduce to no-relation and the intervals are wider or less calibrated than split conformal prediction's, the claim that relational structure is captured without harm is falsified.

Watch

Extended reading notes

Core claim

The central object of the paper is a relational quantile predictor: a spatiotemporal graph network with per-series learnable node embeddings, message-passing layers, and a readout that maps a quantile level $\alpha$ and the recent residual window to predicted quantiles $\hat{q}_{i,\alpha/2}^{t}$ and $\hat{q}_{i,1-\alpha/2}^{t}$ for each series $i$. The interval for series $i$ at horizon $H$ is $[\hat{x}_{i,t+H} + \hat{q}_{i,\alpha/2}^{t},\ \hat{x}_{i,t+H} + \hat{q}_{i,1-\alpha/2}^{t}]$, with an optional offset $\hat{\beta}_i$ to narrow the band. The network is trained by minimizing the pinball loss on calibration residuals, and the graph is learned end-to-end by sampling a $K$-nearest-neighbor graph per node with a straight-through Gumbel-Top-K estimator; adaptation to distribution shift is done by fine-tuning only the node embeddings. Proposition 3.1 bounds the conditional coverage gap by the total variation distance between the learned and true data-generating distributions, so the coverage guarantee is approximate rather than finite-sample. Empirically, CoRel attains the best or near-best Winkler score in nearly every combination of three real datasets (METR-LA traffic, AQI air quality, CER-E energy) and three base forecasters (RNN, Transformer, STGNN), and in the synthetic GPVAR benchmark its intervals approach the theoretical optimum width while matching the coverage of split conformal prediction.

Load-bearing premise

The argument depends on the trained quantile network being well calibrated, so that $P_{\psi}(x \in \hat{C}) = 1-\alpha$ holds by construction; if the network is misspecified or miscalibrated, Proposition 3.1's coverage bound is vacuous and CoRel has no finite-sample conformal guarantee.

Editorial extensions

If this is right

  • Prediction intervals can be attached to any pre-trained point forecaster, including models that ignore inter-series structure entirely.
  • When residuals carry spatiotemporal dependencies, CoRel produces narrower intervals than per-series conformal methods at comparable coverage, as measured by Winkler score.
  • No pre-specified graph is needed: on the GPVAR benchmark the learned structure recovers most of the benefit of the true graph, nearly matching its interval width.
  • The adaptive component updates only node embeddings at test time, so the model can track seasonality and drift without full retraining.
  • Coverage is only approximate; the guarantee in Proposition 3.1 is proportional to how well the quantile network approximates the true residual distribution.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the relational structure is learned from residuals, the same machinery could serve as a diagnostic: the recovered graph shows which series share forecast-risk factors, a by-product the paper does not develop.
  • The method is demonstrated on homogeneous univariate series; extending the decoder to joint or copula-style intervals would address cross-series tail dependence that marginal intervals cannot capture, a direction the conclusion names as future work.
  • A natural stress test not run in the paper is a sudden regime change in one region, measuring whether online embedding fine-tuning recovers coverage faster than per-series baselines retrained from scratch.
  • The optional offset $\hat{\beta}_i$ trades coverage for width; a decision-theoretic extension could tune $\beta$ online per series against the Winkler score rather than once on a validation set.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper proposes CoRel, a post-hoc uncertainty quantification method for collections of correlated time series. Given any pre-trained point forecaster, a graph neural network (a hybrid global-local spatiotemporal encoder) is trained on the prediction residuals of a calibration set to output conditional quantiles of the residual distribution; the relational graph is learned end-to-end via a Gumbel-top-K sampling module, and node embeddings provide per-series adaptability. Prediction intervals are formed directly from the predicted quantiles, optionally narrowed by optimizing a quantile offset beta, and an online variant updates only the node embeddings to track non-stationarity. The authors state a coverage bound (Prop. 3.1) and evaluate on METR-LA, CER-E, AQI, and the synthetic GPVAR benchmark against SCP, SeqCP, NexCP, SPCI, HopCPT, and an ablation (CORNN). CoRel attains the best Winkler score in seven of nine main scenarios, approaches the theoretically optimal interval width in GPVAR, and reduces the coverage gap in the adaptation study.

Significance. Strengths: the paper is, to my knowledge, the first to bring graph deep learning operators into interval construction for correlated time series; the method does not require the relational structure to be known; calibration-residual training is cheap compared with the baselines (about 5 minutes versus about 3 days for SPCI on METR-LA); the code is publicly released; the controlled GPVAR experiment reaches near-optimal width (1.33 versus the theoretical 1.315); and the authors are transparent that the intervals provide only approximate coverage (Sec. 3.1) and can exhibit arbitrarily large coverage gaps (Sec. 3.4). Weaknesses: the theoretical statement (Prop. 3.1, App. A, Eq. (22)) is a tautological total-variation bound with an unquantified and unestimated TV term, so the advertised 'sound' and 'distribution-free' framing is not supported; the coverage gaps in Table 1 are systematically negative, up to -4.99 on CER-E; and the adaptation protocol in Sec. 5.3 is ambiguous about out-of-sample evaluation. If the empirical findings are reproducible, the method is a solid engineering contribution to uncertainty quantification, but the theoretical footing is overstated as written.

major comments (3)
  1. [3.3 / App. A (Prop. 3.1, Eq. (22))] The central theoretical claim is not a conformal guarantee. Eq. (22) states that P_psi(x in C_hat) = 1 - alpha 'by construction' for the intervals in Eqs. (14) and (16); however, P_psi is defined as the distribution associated with the fitted network Q_psi, so Eq. (22) holds by definition rather than by any property enforced by the algorithm. Q_psi is trained by minimizing pinball loss on a finite calibration set and the intervals are used directly, with no conformal correction step, so under model misspecification, quantile crossing, or distribution shift the equality need not hold for the data-generating distribution. The proposition therefore reduces to coverage >= 1 - alpha - TV(P_psi, P_true), with the TV term neither quantified nor estimated; the statement in Sec. 3.3 that the TV term 'can be expected to shrink asymptotically as the size of the calibration set increases' is not backed by a theorem. The analogy with Barber et al. (2023) is loose, because that work supplies explicit non-asymptotic control of the miscoverage gap, which is absent here. This matters because Sec. 1 advertises 'a novel, sound, and effective CP method' and the abstract advertises a 'distribution-free approach.' I see two compatible fixes: (i) add a CQR-style conformal adjustment on a fresh holdout of calibration residuals and state the resulting finite-sample guarantee; or (ii) drop the 'sound'/'distribution-free' framing, position CoRel explicitly as an approximate plug-in quantile regressor, and validate empirically that the validation-set coverage gap tracks the test-time gap.
  2. [5.3 (Table 3), Sec. 3.4] The adaptation experiment does not currently establish an out-of-sample benefit. The text says the test set is split into K = 6 folds and the model is 'iteratively fine-tuning the model on each fold,' but it does not state whether evaluation on a fold occurs after fine-tuning on that same fold (which would be in-sample and optimistic) or on a subsequent unseen fold. Given App. F states that fine-tuning uses samples from the last M steps with M equal to one sixth of the test length, the in-sample reading is plausible and must be ruled out. Please specify the protocol precisely (e.g., fine-tune on the latest M steps and evaluate on the following M steps) and report the number of seeds. In addition, the with-adaptation variant is compared only against its non-adaptive counterpart; since the adaptive component is advertised as handling non-exchangeable data, it should also be compared against adaptive baselines (e.g., SPCI with retraining or ACI-style reweighting) to substantiate the Sec. 3.4 claim that the procedure 'can effectively improve the quality of the uncertainty estimates.'
  3. [Abstract / Table 1 (CER-E rows)] The abstract's claim that CoRel 'provides accurate coverage' is only partially supported. Every one of the nine COREL configurations in Table 1 shows a negative Delta Cov, i.e., under-coverage, with the largest gaps on CER-E (RNN: -3.60 +/- 0.35; TRANSF: -3.97 +/- 0.24; STGNN: -4.99 +/- 0.61), the last falling in the paper's own red zone (>4%). The paper acknowledges the CER-E exceptions and explains that model selection optimized the Winkler score, but under-coverage is the unsafe direction for interval methods and the negative gaps are systematic rather than dataset-specific. Please either qualify the abstract's coverage claim, report results under a coverage-constrained model selection (e.g., among configurations with coverage gap within one percentage point of the target, pick the smallest Winkler score), or justify why the observed gaps are acceptable for the intended risk-sensitive applications.
minor comments (7)
  1. [Eq. (15)] Typo: the second quantile term should read q_hat_{i, alpha/2 + beta_i}; the subscript i on beta is missing in the printed formula.
  2. [Table 1 and Experimental Setup] Please state the number of independent seeds or runs used to compute the reported mean +/- std. Several baselines (SCP, SeqCP, NexCP) report no error bars or +/- 0.00, which makes it impossible to assess significance for those entries.
  3. [App. F (SPCI configuration)] The SPCI baseline is run without its time-adaptive retraining procedure, which is disclosed ('we trained each SPCI model only once on the calibration data'). The main-table comparison is even-handed since CoRel is also non-adaptive there, but a sentence discussing how this restriction affects the 'state-of-the-art' claim would help, especially in relation to the adaptive claim in Sec. 5.3.
  4. [Sec. 5.1 (Results)] The text says CoRel 'outperforms the competitors in terms of Winkler score in almost all cases'; Table 1 shows HopCPT winning the two STGNN-base scenarios (METR-LA and CER-E). This is acknowledged, but stating the count (seven of nine) would make the claim precise.
  5. [Notation] The acronym appears as COREL, 'C OREL', and 'CoRel' (Fig. 1 and Sec. 3.1) across the text; please unify the notation.
  6. [App. C] A compact summary table of dataset statistics (N, W, H, series length, split sizes) would improve readability; these details are currently dispersed in prose.
  7. [Sec. 3.3] The sentence 'Prop. 3.1 can be seen as an analogous result' to Barber et al. (2023) should explicitly note the difference that Barber et al. provides a computable non-asymptotic bound, whereas the TV term in Prop. 3.1 is not estimated anywhere in the paper.

Circularity Check

1 steps flagged · score 7.0 of 10

Prop. 3.1's coverage guarantee is asserted 'by construction' in Eq. (22), so the paper's theoretical soundness claim reduces to the assumption that the learned quantile distribution is exactly calibrated; the empirical benchmark results are not themselves circular.

  1. self definitional [Appendix A, Eq. (22); used in Prop. 3.1, Sec. 3.3]
    "For ˆC α i,t in both Eq. 14 and Eq. 16 we have by construction that P c ψ xi t+H ∈ bC α i,t = 1 − α. (22)"

    The proof of Prop. 3.1 inserts Eq. (22) into the total-variation inequality Eq. (21) and obtains the advertised coverage bound. Eq. (22), however, is not derived from a conformal calibration step; it is asserted 'by construction' for the distribution Pψ associated with the learned quantile network Qψ. Since Qψ is only trained by minimizing pinball loss on a finite calibration set and the output intervals are used directly, the property 'Pψ puts mass 1−α inside the interval' is exactly the calibration assumption that the theorem is supposed to establish. The proposition therefore restates the premise that the learned quantile distribution is perfectly calibrated, and the TV term merely converts this into a statement about the true distribution.

full rationale

The clearest circular step is in the theoretical justification. Prop. 3.1, which is presented as the soundness result for CoRel, relies on Eq. (22) of Appendix A, where exact coverage under the learned distribution is declared to hold 'by construction.' That declaration makes the coverage bound self-definitional: if the learned quantile distribution is taken to be exactly calibrated, then the interval built from its quantiles has 1−α coverage under that distribution by definition, and the theorem adds only an unquantified total-variation comparison to the true distribution. No conformal adjustment step in the method enforces Eq. (22); the RelQP is fit with pinball loss on residuals, so the strong calibration property is an assumption imported into the proof rather than a consequence of the conformal framework. This does not invalidate the empirical comparisons in Table 1, which are ordinary benchmark evaluations and are not statistically forced by the theorem. The self-citations to prior graph-learning work by the same authors are also not load-bearing in the same way: the graph-sampling machinery is an implementation choice, not the source of the coverage claim. Overall, the experimental contribution has independent content, but the paper's central theoretical claim to be 'a novel, sound, and effective CP method' is partially circular because its formal guarantee reduces to an asserted calibration property rather than to a conformal construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method rests primarily on domain assumptions (sparse Granger causality, time-invariance in the base setup) and on the ad hoc assumption in Eq. (22) that the learned quantile function is exactly calibrated. The main fitted quantities are the graph score matrix and node embeddings. No new physical or ontological entities are introduced.

free parameters (5)
  • node embeddings V = trained per time series, values not reported
    Per-series learnable vectors condition the quantile network on local behavior and are the only parameters updated during adaptation (Sec. 3.1, 3.4).
  • graph score matrix Phi = N x N learned scores, not reported
    Scores parameterize the categorical distribution over K-NN graphs and are learned end-to-end from calibration residuals (Sec. 3.2).
  • neighborhood size K = 20 for real-world data, varied in GPVAR
    Number of neighbors sampled per node; set by hand rather than derived (Sec. F).
  • quantile grid = 39 equally spaced quantile levels
    The quantile function is approximated by predicting 39 quantiles and interpolating; the choice is manual (Appendix D).
  • architecture hyperparameters = hidden sizes, embedding size, learning rate, epochs tuned per dataset
    Tuned on a validation set with a small grid search (Appendix F); values are dataset-dependent and not exhaustively reported.
assumptions (4)
  • domain assumption Sparse Granger-style dependence structure
    Sec. 2.1 postulates each series depends on a small subset of others, justifying the K-NN graph learning module.
  • domain assumption Time-invariant data-generating process in the base setup
    Eq. (1) assumes a stochastic time-invariant process; non-stationarity is handled later by adaptation.
  • ad hoc to paper Learned quantile function is exactly calibrated
    Eq. (22) asserts P_psi coverage = 1-alpha by construction, the key premise of Proposition 3.1.
  • ad hoc to paper Graph neural network expressivity suffices to shrink total variation
    Sec. 3.3 informally assumes the quantile regressor is expressive enough that TV(P_psi, P_true) tends to zero with calibration size; no rate or conditions are proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Relational Conformal Prediction for Correlated Time Series." pith.science (2026). https://pith.science/paper/KCCADWT6

@misc{pith2026250209443,
  author       = {Pith},
  title        = {Pith review of: Relational Conformal Prediction for Correlated Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KCCADWT6}},
  note         = {Machine review of arXiv:2502.09443}
}
read the original abstract

We address the problem of uncertainty quantification in time series forecasting by exploiting observations at correlated sequences. Relational deep learning methods leveraging graph representations are among the most effective tools for obtaining point estimates from spatiotemporal data and correlated time series. However, the problem of exploiting relational structures to estimate the uncertainty of such predictions has been largely overlooked in the same context. To this end, we propose a novel distribution-free approach based on the conformal prediction framework and quantile regression. Despite the recent applications of conformal prediction to sequential data, existing methods operate independently on each target time series and do not account for relationships among them when constructing the prediction interval. We fill this void by introducing a novel conformal prediction method based on graph deep learning operators. Our approach, named Conformal Relational Prediction (CoRel), does not require the relational structure (graph) to be known a priori and can be applied on top of any pre-trained predictor. Additionally, CoRel includes an adaptive component to handle non-exchangeable data and changes in the input time series. Our approach provides accurate coverage and achieves state-of-the-art uncertainty quantification in relevant benchmarks.

Figures

Figures reproduced from arXiv: 2502.09443 by the authors.

Figure 1
Figure 1. Overview of COREL. Past residuals are used as input to a hybrid global-local graph-based quantile network. such representation, we can leverage GDL methods for time series to process the data. In particular, we train a STGNN on the residuals of the calibration set to predict the quantiles of the error distribution. Conditioning the prediction on the recent history of related time series allows for taking the depende… view at source ↗
Figure 2
Figure 2. Comparison between (a) the true graph and (b) the graph structure learned by COREL in GPVAR. plus additional links. Note that perfectly recovering the ground truth graph is not required to achieve optimal performance here. G.2. Optimization of the PI width Here, we show how the width of the PI can be reduced by searching for an appropriate quantile offset β. Specifically, we perform the optimization described in Eq.… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 59 canonical work pages

  1. [1]

    Conformal prediction: A gentle introduction

    Anastasios N Angelopoulos, Stephen Bates, et al. Conformal prediction: A gentle introduction . Foundations and Trends in Machine Learning , 16 0 (4): 0 494--591, 2023

    2023

  2. [2]

    Theoretical foundations of conformal prediction

    Anastasios N Angelopoulos, Rina Foygel Barber, and Stephen Bates. Theoretical foundations of conformal prediction. arXiv preprint arXiv:2411.11824, 2024

    arXiv 2024

  3. [3]

    Conformal prediction for time series with Modern Hopfield Networks

    Andreas Auer, Martin Gauch, Daniel Klotz, and Sepp Hochreiter. Conformal prediction for time series with Modern Hopfield Networks . Advances in neural information processing systems, 2023

    2023

  4. [4]

    A gentle introduction to deep learning for graphs

    Davide Bacciu, Federico Errica, Alessio Micheli, and Marco Podda. A gentle introduction to deep learning for graphs . Neural Networks, 129: 0 203--221, 2020

    2020

  5. [5]

    Conformal prediction beyond exchangeability

    Rina Foygel Barber, Emmanuel J Candes, Aaditya Ramdas, and Ryan J Tibshirani. Conformal prediction beyond exchangeability . The Annals of Statistics, 51 0 (2): 0 816--845, 2023

    2023

  6. [6]

    Estimating or propagating gradients through stochastic neurons for conditional computation

    Yoshua Bengio, Nicholas L \'e onard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation . arXiv preprint arXiv:1308.3432, 2013

    arXiv 2013

  7. [7]

    Deep Learning for Time Series Forecasting: Tutorial and Literature Survey

    Konstantinos Benidis, Syama Sundar Rangapuram, Valentin Flunkert, Yuyang Wang, Danielle Maddix, Caner Turkmen, Jan Gasthaus, Michael Bohlke-Schneider, David Salinas, Lorenzo Stella, Fran c ois-Xavier Aubet, Laurent Callot, and Tim Januschowski. Deep Learning for Time Series Forecasting: Tutorial and Literature Survey . ACM Comput. Surv., 55 0 (6), dec 202...

    doi:10.1145/3533382 2022

  8. [8]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges

    Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Veli c kovi \'c . Geometric deep learning: Grids, groups, graphs, geodesics, and gauges . arXiv preprint arXiv:2104.13478, 2021

    arXiv 2021

Show all 85 references
  1. [9]

    Cluster-GCN: An efficient algorithm for training deep and large graph convolutional networks

    Wei-Lin Chiang, Xuanqing Liu, Si Si, Yang Li, Samy Bengio, and Cho-Jui Hsieh. Cluster-GCN: An efficient algorithm for training deep and large graph convolutional networks . In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pag...

    2019

  2. [10]

    On the properties of neural machine translation: Encoder-decoder approaches

    Kyunghyun Cho, Bart Van Merri \"e nboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches . arXiv preprint arXiv:1409.1259, 2014

    2014 arXiv

  3. [11]

    Torch Spatiotemporal , 2022

    Andrea Cini and Ivan Marisca. Torch Spatiotemporal , 2022. URL https://github.com/TorchSpatiotemporal/tsl

    2022

  4. [12]

    Filling the G\_ap\_s: Multivariate Time Series Imputation by Graph Neural Networks

    Andrea Cini, Ivan Marisca, and Cesare Alippi. Filling the G\_ap\_s: Multivariate Time Series Imputation by Graph Neural Networks . In International Conference on Learning Representations, 2022

    2022

  5. [13]

    Graph Deep Learning for Time Series Forecasting

    Andrea Cini, Ivan Marisca, Daniele Zambon, and Cesare Alippi. Graph Deep Learning for Time Series Forecasting . arXiv preprint arXiv:2310.15978, 2023 a

    2023 arXiv

  6. [14]

    Taming Local Effects in Graph-based Spatiotemporal Forecasting

    Andrea Cini, Ivan Marisca, Daniele Zambon, and Cesare Alippi. Taming Local Effects in Graph-based Spatiotemporal Forecasting . Advances in Neural Information Processing Systems, 2023 b

    2023

  7. [15]

    Sparse graph learning from spatiotemporal time series

    Andrea Cini, Daniele Zambon, and Cesare Alippi. Sparse graph learning from spatiotemporal time series . Journal of Machine Learning Research, 24 0 (242): 0 1--36, 2023 c

    2023

  8. [16]

    Distribution free prediction sets for node classification

    Jase Clarkson. Distribution free prediction sets for node classification. In International Conference on Machine Learning, pages 6268--6278. PMLR, 2023

    2023

  9. [17]

    CER Smart Metering Project - Electricity Customer Behaviour Trial, 2009-2010 [dataset]

    Commission for Energy Regulation . CER Smart Metering Project - Electricity Customer Behaviour Trial, 2009-2010 [dataset] . Irish Social Science Data Archive. SN: 0012-00, 2016. URL https://www.ucd.ie/issda/data/commissionforenergyregulationcer

    2009

  10. [18]

    Implicit quantile networks for distributional reinforcement learning

    Will Dabney, Georg Ostrovski, David Silver, and R \'e mi Munos. Implicit quantile networks for distributional reinforcement learning. In International conference on machine learning, pages 1096--1105. PMLR, 2018

    2018

  11. [19]

    Valid conformal prediction for dynamic GNNs

    Ed Davis, Ian Gallagher, Daniel John Lawson, and Patrick Rubin-Delanchy. Valid conformal prediction for dynamic GNNs . arXiv preprint arXiv:2405.19230, 2024

    2024 arXiv

  12. [20]

    Normalizing kalman filters for multivariate time series analysis

    Emmanuel de B \'e zenac, Syama Sundar Rangapuram, Konstantinos Benidis, Michael Bohlke-Schneider, Richard Kurle, Lorenzo Stella, Hilaf Hasson, Patrick Gallinari, and Tim Januschowski. Normalizing kalman filters for multivariate time series analysis . Advances in Neural Informa...

    2020

  13. [21]

    PyTorch Lightning , 3 2019

    William Falcon and The PyTorch Lightning team . PyTorch Lightning , 3 2019. URL https://github.com/PyTorchLightning/pytorch-lightning

    2019

  14. [22]

    Calibrated multiple-output quantile regression with representation learning

    Shai Feldman, Stephen Bates, and Yaniv Romano. Calibrated multiple-output quantile regression with representation learning. Journal of Machine Learning Research, 24 0 (24): 0 1--48, 2023

    2023

  15. [23]

    Fast graph representation learning with PyTorch Geometric

    Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with PyTorch Geometric . arXiv preprint arXiv:1903.02428, 2019

    1903 arXiv

  16. [24]

    On the Equivalence Between Temporal and Static Equivariant Graph Representations

    Jianfei Gao and Bruno Ribeiro. On the Equivalence Between Temporal and Static Equivariant Graph Representations . In International Conference on Machine Learning, pages 7052--7076. PMLR, 2022

    2022

  17. [25]

    Deep learning for time series forecasting: The electric load case

    Alberto Gasparin, Slobodan Lukovic, and Cesare Alippi. Deep learning for time series forecasting: The electric load case . CAAI Transactions on Intelligence Technology, 7 0 (1): 0 1--25, 2022

    2022

  18. [26]

    Probabilistic forecasting with spline quantile function RNNs

    Jan Gasthaus, Konstantinos Benidis, Yuyang Wang, Syama Sundar Rangapuram, David Salinas, Valentin Flunkert, and Tim Januschowski. Probabilistic forecasting with spline quantile function RNNs . In The 22nd international conference on artificial intelligence and statistics, page...

    1901

  19. [27]

    Adaptive conformal inference under distribution shift

    Isaac Gibbs and Emmanuel Candes. Adaptive conformal inference under distribution shift . Advances in Neural Information Processing Systems, 34: 0 1660--1672, 2021

    2021

  20. [28]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry . In International Conference on Machine Learning, pages 1263--1272. PMLR, 2017

    2017

  21. [29]

    Probabilistic time series forecasting with implicit quantile networks

    Ad \`e le Gouttes, Kashif Rasul, Mateusz Koren, Johannes Stephan, and Tofigh Naghibi. Probabilistic time series forecasting with implicit quantile networks. arXiv preprint arXiv:2107.03743, 2021

    2021 arXiv

  22. [30]

    Investigating causal relations by econometric models and cross-spectral methods

    Clive WJ Granger. Investigating causal relations by econometric models and cross-spectral methods . Econometrica: journal of the Econometric Society, pages 424--438, 1969

    1969

  23. [31]

    Array programming with NumPy

    Charles R Harris, K Jarrod Millman, St \'e fan J Van Der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J Smith, et al. Array programming with NumPy . Nature, 585 0 (7825): 0 357--362, 2020

    2020

  24. [32]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  25. [33]

    Uncertainty quantification over graph with conformalized graph neural networks

    Kexin Huang, Ying Jin, Emmanuel Candes, and Jure Leskovec. Uncertainty quantification over graph with conformalized graph neural networks. Advances in Neural Information Processing Systems, 36, 2024

    2024

  26. [34]

    Forecasting: principles and practice

    Rob J Hyndman and George Athanasopoulos. Forecasting: principles and practice . OTexts, 2018

    2018

  27. [35]

    Ensemble conformalized quantile regression for probabilistic time series forecasting

    Vilde Jensen, Filippo Maria Bianchi, and Stian Normann Anfinsen. Ensemble conformalized quantile regression for probabilistic time series forecasting . IEEE Transactions on Neural Networks and Learning Systems, 2022

    2022

  28. [36]

    Spatio-temporal conformal prediction for power outage data

    Hanyang Jiang, Yao Xie, and Feng Qiu. Spatio-temporal conformal prediction for power outage data. arXiv preprint arXiv:2411.17099, 2024

    2024 arXiv

  29. [37]

    A Survey on Graph Neural Networks for Time Series: Forecasting, Classification, Imputation, and Anomaly Detection

    Ming Jin, Huan Yee Koh, Qingsong Wen, Daniele Zambon, Cesare Alippi, Geoffrey I Webb, Irwin King, and Shirui Pan. A Survey on Graph Neural Networks for Time Series: Forecasting, Classification, Imputation, and Anomaly Detection . arXiv preprint arXiv:2307.03759, 2023

    2023 arXiv

  30. [38]

    Multivariate quantile function forecaster

    Kelvin Kan, Fran c ois-Xavier Aubet, Tim Januschowski, Youngsuk Park, Konstantinos Benidis, Lars Ruthotto, and Jan Gasthaus. Multivariate quantile function forecaster. In International Conference on Artificial Intelligence and Statistics, pages 10603--10621. PMLR, 2022

    2022

  31. [39]

    Differentiable graph module (dgm) for graph convolutional networks

    Anees Kazi, Luca Cosmo, Seyed-Ahmad Ahmadi, Nassir Navab, and Michael Bronstein. Differentiable graph module (dgm) for graph convolutional networks . IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    2022

  32. [40]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization . 3nd International Conference on Learning Representations, ICLR 2015, 2015

    2015

  33. [41]

    Quantile regression

    Roger Koenker and Kevin F Hallock. Quantile regression . Journal of economic perspectives, 15 0 (4): 0 143--156, 2001

    2001

  34. [42]

    Stochastic beams and where to find them: The Gumbel-top-K trick for sampling sequences without replacement

    Wouter Kool, Herke Van Hoof, and Max Welling. Stochastic beams and where to find them: The Gumbel-top-K trick for sampling sequences without replacement . In International Conference on Machine Learning, pages 3499--3508. PMLR, 2019

    2019

  35. [43]

    Kernel-based optimally weighted conformal time-series prediction

    Jonghyeok Lee, Chen Xu, and Yao Xie. Kernel-based optimally weighted conformal time-series prediction. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=oP7arLOWix

    2025

  36. [44]

    Conformal prediction for time series with transformer

    Junghwan Lee, Chen Xu, and Yao Xie. Conformal prediction for time series with transformer. In ICML 2024 Workshop on Structured Probabilistic Inference & Generative Modeling , 2024. URL https://openreview.net/forum?id=3dDDKaSrye

    2024

  37. [45]

    Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting . In International Conference on Learning Representations, 2018

    2018

  38. [46]

    Foundation models for time series analysis: A tutorial and survey

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. Foundation models for time series analysis: A tutorial and survey . arXiv preprint arXiv:2403.14735, 2024

    2024 arXiv

  39. [47]

    Forecasting: its role and value for planning and strategy

    Spyros Makridakis. Forecasting: its role and value for planning and strategy. International journal of forecasting, 12 0 (4): 0 513--537, 1996

    1996

  40. [48]

    Forecasting, uncertainty and risk; perspectives on clinical decision-making in preventive and curative medicine

    Spyros Makridakis, Richard Kirkham, Ann Wakefield, Maria Papadaki, Joanne Kirkham, and Lisa Long. Forecasting, uncertainty and risk; perspectives on clinical decision-making in preventive and curative medicine. International Journal of Forecasting, 35 0 (2): 0 659--666, 2019. ...

    2019 doi

  41. [49]

    Learning Latent Graph Structures and their Uncertainty

    Alessandro Manenti, Daniele Zambon, and Cesare Alippi. Learning Latent Graph Structures and their Uncertainty . arXiv preprint arXiv:2405.19933, 2024

    2024 arXiv

  42. [50]

    Valid model-free spatial prediction

    Huiying Mao, Ryan Martin, and Brian J Reich. Valid model-free spatial prediction. Journal of the American Statistical Association, 119 0 (546): 0 904--914, 2024

    2024

  43. [51]

    Quantile regression forests

    Nicolai Meinshausen and Greg Ridgeway. Quantile regression forests. Journal of machine learning research, 7 0 (6), 2006

    2006

  44. [52]

    Discrete latent structure in neural networks

    Vlad Niculae, Caio F Corro, Nikita Nangia, Tsvetomila Mihaylova, and Andr \'e FT Martins. Discrete latent structure in neural networks . arXiv preprint arXiv:2301.07473, 2023

    2023

  45. [53]

    Autoregressive quantile networks for generative modeling

    Georg Ostrovski, Will Dabney, and R \'e mi Munos. Autoregressive quantile networks for generative modeling. In International Conference on Machine Learning, pages 3936--3945. PMLR, 2018

    2018

  46. [54]

    RNN with Particle Flow for Probabilistic Spatio-temporal Forecasting

    Soumyasundar Pal, Liheng Ma, Yingxue Zhang, and Mark Coates. RNN with Particle Flow for Probabilistic Spatio-temporal Forecasting . In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8336--8348...

    2021

  47. [55]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu ...

    2019

  48. [56]

    Barrow, Souhaib Ben Taieb , Christoph Bergmeir, Ricardo J

    Fotios Petropoulos, Daniele Apiletti, Vassilios Assimakopoulos, Mohamed Zied Babai, Devon K. Barrow, Souhaib Ben Taieb , Christoph Bergmeir, Ricardo J. Bessa, Jakub Bijak, John E. Boylan, Jethro Browell, Claudio Carnevale, Jennifer L. Castle, Pasquale Cirillo, Michael P. Cleme...

    2022

  49. [57]

    a fl, Johannes Lehner, Philipp Seidl, Michael Widrich, Lukas Gruber, Markus Holzleitner, Thomas Adler, David Kreil, Michael K Kopp, G \

    Hubert Ramsauer, Bernhard Sch \"a fl, Johannes Lehner, Philipp Seidl, Michael Widrich, Lukas Gruber, Markus Holzleitner, Thomas Adler, David Kreil, Michael K Kopp, G \"u nter Klambauer, Johannes Brandstetter, and Sepp Hochreiter. Hopfield networks is all you need. In Internati...

    2021

  50. [58]

    Deep state space models for time series forecasting

    Syama Sundar Rangapuram, Matthias W Seeger, Jan Gasthaus, Lorenzo Stella, Yuyang Wang, and Tim Januschowski. Deep state space models for time series forecasting . Advances in Neural Information Processing Systems, 31: 0 7785--7794, 2018

    2018

  51. [59]

    Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting

    Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland Vollgraf. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting . In International Conference on Machine Learning, pages 8857--8868. PMLR, 2021

    2021

  52. [60]

    DeepAR: Probabilistic forecasting with autoregressive recurrent networks

    David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. DeepAR: Probabilistic forecasting with autoregressive recurrent networks . International Journal of Forecasting, 36 0 (3): 0 1181--1191, 2020

    2020

  53. [61]

    Multivariate time series forecasting with latent graph inference

    Victor Garcia Satorras, Syama Sundar Rangapuram, and Tim Januschowski. Multivariate time series forecasting with latent graph inference . arXiv preprint arXiv:2203.03423, 2022

    2022 arXiv

  54. [62]

    The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains

    David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains . IEEE Signal Processing Magazine, 30 0 (3): 0 83--98, 2013

    2013

  55. [63]

    Uncertainty quantification: theory, implementation, and applications

    Ralph C Smith. Uncertainty quantification: theory, implementation, and applications. SIAM, 2024

    2024

  56. [64]

    A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting

    Slawek Smyl. A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting . International Journal of Forecasting, 36 0 (1): 0 75--85, 2020

    2020

  57. [65]

    Conformal time-series forecasting

    Kamile Stankeviciute, Ahmed M Alaa, and Mihaela van der Schaar. Conformal time-series forecasting . Advances in neural information processing systems, 34: 0 6216--6228, 2021

    2021

  58. [66]

    Copula conformal prediction for multi-step time series prediction

    Sophia Huiwen Sun and Rose Yu. Copula conformal prediction for multi-step time series prediction. In The Twelfth International Conference on Learning Representations, 2024

    2024

  59. [67]

    Conformal prediction under covariate shift

    Ryan J Tibshirani, Rina Foygel Barber, Emmanuel Candes, and Aaditya Ramdas. Conformal prediction under covariate shift. Advances in neural information processing systems, 32, 2019

    2019

  60. [68]

    Guido Van Rossum and Fred L. Drake. Python 3 Reference Manual . CreateSpace, Scotts Valley, CA, 2009. ISBN 1441412697

    2009

  61. [69]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need . In Advances in Neural Information Processing Systems, pages 5998--6008, 2017

    2017

  62. [70]

    Algorithmic learning in a random world, volume 29

    Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic learning in a random world, volume 29. Springer, 2005

    2005

  63. [71]

    Diff STG : Probabilistic spatio-temporal graph forecasting with denoising diffusion models

    Haomin Wen, Youfang Lin, Yutong Xia, Huaiyu Wan, Qingsong Wen, Roger Zimmermann, and Yuxuan Liang. Diff STG : Probabilistic spatio-temporal graph forecasting with denoising diffusion models. In Proceedings of the 31st ACM International Conference on Advances in Geographic Info...

    2023

  64. [72]

    A multi-horizon quantile recurrent forecaster

    Ruofeng Wen, Kari Torkkola, Balakrishnan Narayanaswamy, and Dhruv Madeka. A multi-horizon quantile recurrent forecaster . arXiv preprint arXiv:1711.11053, 2017

    2017 arXiv

  65. [73]

    A decision-theoretic approach to interval estimation

    Robert L Winkler. A decision-theoretic approach to interval estimation. Journal of the American Statistical Association, 67 0 (337): 0 187--191, 1972

    1972

  66. [74]

    Quantifying uncertainty in deep spatiotemporal forecasting

    Dongxia Wu, Liyao Gao, Matteo Chinazzi, Xinyue Xiong, Alessandro Vespignani, Yi-An Ma, and Rose Yu. Quantifying uncertainty in deep spatiotemporal forecasting . In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 1841--1851, 2021

    2021

  67. [75]

    Graph wavenet for deep spatial-temporal graph modeling

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for deep spatial-temporal graph modeling . In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 1907--1913, 2019

    1907

  68. [76]

    Reparameterizable subset sampling via continuous relaxations

    Sang Michael Xie and Stefano Ermon. Reparameterizable subset sampling via continuous relaxations . In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 3919--3925, 2019

    2019

  69. [77]

    Conformal prediction for time series

    Chen Xu and Yao Xie. Conformal prediction for time series. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (10): 0 11575--11587, 2023 a

    2023

  70. [78]

    Sequential predictive conformal inference for time series

    Chen Xu and Yao Xie. Sequential predictive conformal inference for time series. In International Conference on Machine Learning, pages 38707--38727. PMLR, 2023 b

    2023

  71. [79]

    Conformal prediction for multi-dimensional time series by ellipsoidal sets

    Chen Xu, Hanyang Jiang, and Yao Xie. Conformal prediction for multi-dimensional time series by ellipsoidal sets. arXiv preprint arXiv:2403.03850, 2024

    2024 arXiv

  72. [80]

    AZ-whiteness Test: A Test for Signal Uncorrelation on Spatio-Temporal Graphs

    Daniele Zambon and Cesare Alippi. AZ-whiteness Test: A Test for Signal Uncorrelation on Spatio-Temporal Graphs . In Advances in Neural Information Processing Systems, 2022

    2022

  73. [81]

    Where and How to Improve Graph-based Spatio-temporal Predictors , 2023

    Daniele Zambon and Cesare Alippi. Where and How to Improve Graph-based Spatio-temporal Predictors , 2023. URL http://arxiv.org/abs/2302.01701

    2023

  74. [82]

    Graph state-space models

    Daniele Zambon, Andrea Cini, Lorenzo Livi, and Cesare Alippi. Graph state-space models . arXiv preprint arXiv:2301.01741, 2023

    2023 arXiv

  75. [83]

    Conformal prediction sets for graph neural networks

    Soroush H Zargarbashi, Simone Antonelli, and Aleksandar Bojchevski. Conformal prediction sets for graph neural networks. In International Conference on Machine Learning, pages 12292--12318. PMLR, 2023

    2023

  76. [84]

    Conformalized link prediction on graph neural networks

    Tianyi Zhao, Jian Kang, and Lu Cheng. Conformalized link prediction on graph neural networks. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4490--4499, 2024

    2024

  77. [85]

    Forecasting fine-grained air quality based on big data

    Yu Zheng, Xiuwen Yi, Ming Li, Ruiyuan Li, Zhangqing Shan, Eric Chang, and Tianrui Li. Forecasting fine-grained air quality based on big data . In Proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, pages 2267--2276, 2015

    2015

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.