acceptodds
Under review as a conference paper at ICLR 2027

STORM: Simplicial Topological Recurrent Models

15 views

Abstract

Many dynamical systems, from traffic networks to water infrastructure, involve signals and interactions across nodes, edges, and higher-order groups. Existing temporal models often ignore this structure or use it only as preprocessing, while graph-based approaches remain limited to pairwise relations. We introduce Simplicial Topological Recurrent Models (STORM), recurrent architectures on simplicial complexes that embed simplicial operators into memory gates, allowing topology to shape hidden-state dynamics over time. We prove permutation equivariance, finite-horizon stability to bounded admissible topology perturbations, and a Rademacher-complexity generalization bound with explicit temporal dependence. On traffic and hydraulic datasets, STORM improves over the best baselines in our evaluation by up to 50% at long horizons. For node-only forecasting, STORM and LightSTORM are highly competitive with strong graph models, with LightSTORM using up to roughly two orders of magnitude fewer parameters.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
topological deep learning, simplicial complexes, graph neural networks, spatiotemporal forecasting, recurrent neural networks, higher-order networks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3d5DOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+2)
Read review
Paper summary

The paper introduces STORM, a family of gated recurrent architectures for multivariate time series using simplicial complexes. STORM embeds simplicial convolutional filters directly into the recurrent gates. This is different from prior work which usually uses simplicial complexes in the preprocessing step. This allows topology to shape hidden-state memory updates across multiple simplex orders.

Critical strengths

The idea to combine the simplicial complexes and a gated architecture is novel and interesting. The method is especially strong on long forecasting time horizons, which makes it especially interesting. It also proposes a lighter version with less parameters, while still maintaining performance.

Critical weaknesses

To my understanding, on the traffic datasets, the only measurements are at nodes (sensors). The method “create” edge and triangle signals by averaging the node values around them. So those higher-order signals aren’t actually measured, they’re just smoothed versions of node data. Because of that, when STORM/Light STORM improves over other models, I am not sure the gain comes from real higher-order topology. It might just be getting extra averaged features. For example, Water, which has edge measurement, the improvement of the methods are much smaller than with the traffic dataset. This is the main concern I have. I also did not understand why other methods in related work could not be compared against.

Questions for authors
  1. Why is it not possible to test the method on large-size complexes, as discussed in F.1.1?
  2. It seems that the claim for “LightSTORM using up to roughly two orders of magnitude fewer parameters than other graph baselines” is overstated? I don’t see this in table 10.
  3. As I am not well-versed in the literature, could you explain and potentially add to the appendix why you could (or not) compare against the methods mentioned in the related works?
  4. Finally, related to my weakness, how can you show that the method does work better because it accurately models higher-order topology? Can we use datasets with node/edge/face measurements for example?
5 comments
AuthorsOpenReview ↗

Author Response to Questions Q1-Q3

We thank the reviewer for taking the time to review our work and for appreciating the novelty of the architectural and the empirical results. We appreciate the questions and concerns raised and address them individually below, hoping to clarify the main points.

Q1: Testing STORM on larger complexes is possible. In Appendix F.1.1, we simply state that the datasets used in our experiments contain between 76 and 676 simplices. These are the sizes of the selected benchmarks, not a limitation imposed by the model. As discussed in Appendix D, for a fixed filter degree and bounded simplex degree, the computational cost scales linearly with the number of simplices. We apologize for any confusion and will clarify this in the revised manuscript.

Q2: Thank you for pointing this out. The statement appears in the conclusion, where we discuss the node-only forecasting experiments. We were referring to the results and parameter counts reported in Table 4, rather than Table 10, which reports the parameter counts for the joint multi-order experiments. For example, on PeMS-D4, STAEformer has 1.59–2.07 million parameters, while LightSTORM has 33.9–65.9 thousand parameters. This is the comparison behind our statement that LightSTORM uses up to roughly two orders of magnitude fewer parameters. We will make this distinction clear and add an explicit reference to Table 4 in the conclusion.

Q3: We selected the baselines to compare different ways of modeling spatial and temporal information. LSTM, xLSTM, Transformer, and iTransformer provide comparisons without topological operators. GCN and GCN-xLSTM use graph-based spatial information, while SC-VAR and SCRNN represent existing approaches to temporal modeling on simplicial complexes. In particular, SCRNN applies simplicial convolutions before recurrence, whereas STORM includes them directly inside the recurrent gates. We also compare with DCRNN, Graph WaveNet, STID, and STAEformer in the node-only forecasting task (Table 4). Regarding Chen et al. (2022b, 2023), both methods use higher-order topological information but focus on predicting node-level signals. In contrast, our main task jointly predicts signals across all simplex orders. Therefore, these methods cannot be directly evaluated on the main multi-order task without adapting their architectures. We discuss our baseline selection in Appendix F.2.1 and will add a brief explanation of why some related-work methods are not directly comparable to our main multi-order forecasting task.

AuthorsOpenReview ↗

Author Response to Question Q4 - Part 1 of 2.

Q4: Thank you for raising this important question. We investigated it through the ablation in Appendix H.3, additional control experiments, and a new dataset with native node and edge signals.

In Table 29 (Appendix H.3), we compare graph recurrence with node–edge and node–edge–triangle recurrence on the same node-forecasting task. Adding edge states improves performance in all six evaluated settings, while adding triangle states improves it in five out of six. For example, on PeMS-D8 at S=80, the error decreases from 0.3635 with graph recurrence to 0.2218 with node–edge–triangle recurrence.

To investigate whether these improvements arise simply from averaged edge inputs, we conducted an additional ablation (Table C.1). Removing edge supervision while retaining lifted edge inputs gives similar performance. Moreover, setting observed edge inputs to zero while retaining latent edge states and their boundary-based coupling reduces node NMSE compared with graph recurrence by 36.6% on PeMS-D4 and 36.2% on PeMS-D8.

Table C.1. Node–edge control at S=80S=80. Final-step node NMSE (mean ± standard deviation over 3 seeds; lower is better). Bold: best. Italics: second best.

VariantPeMS-D4PeMS-D8
Graph recurrence, nodes only0.4811 ± 0.00830.3642 ± 0.0117
Node–edge, lifted edge input + node–edge loss0.3152 ± 0.01180.2287 ± 0.0033
Node–edge, lifted edge input + node-only loss0.3061 ± 0.00810.2146 ± 0.0123
Node–edge, latent edge state + node-only loss0.3049 ± 0.01540.2324 ± 0.0085

We also evaluated three node forecasters (DCRNN, Graph WaveNet, and STID) followed by deterministic lifting to edges and triangles, as suggested by another reviewer. As shown in Tables C.2 and C.3, the best STORM/LightSTORM variant achieves lower pooled multi-order NMSE than all three lifted baselines at every horizon S≥10S \geq 10.

Table C.2. Node forecasting followed by deterministic lifting on PeMS-D4. Final-step pooled NMSE (mean ± standard deviation over 5 seeds; lower is better). Bold: best. Italics: second best.

ModelS=5S=10S=20S=40S=60S=80
DCRNN + lift0.0483±0.00020.0789±0.00280.1930±0.03110.4018±0.01810.6548±0.02120.8595±0.0635
Graph WaveNet + lift0.0500±0.00050.0756±0.00310.1407±0.00560.2899±0.01170.3969±0.03370.5179±0.0308
STID + lift0.0488±0.00020.0762±0.00350.1356±0.00240.2678±0.00430.3558±0.00480.4031±0.0079
Best STORM/LightSTORM0.0491±0.00020.0725±0.00040.1206±0.00200.2247±0.00330.2686±0.00310.3164±0.0017

Table C.3. Node forecasting followed by deterministic lifting on PeMS-D8. Final-step pooled NMSE (mean ± standard deviation over 5 seeds; lower is better). Bold: best. Italics: second best.

ModelS=5S=10S=20S=40S=60S=80
DCRNN + lift0.0408±0.00060.0619±0.00310.1326±0.00370.3173±0.03320.5309±0.03800.7433±0.0520
Graph WaveNet + lift0.0411±0.00180.0566±0.00150.1149±0.00570.2357±0.00880.3497±0.02550.4582±0.0729
STID + lift0.0375±0.00040.0545±0.00080.1009±0.00230.2057±0.00270.3116±0.00460.3838±0.0065
Best STORM/LightSTORM0.0402±0.00050.0541±0.00070.0953±0.00190.1592±0.00930.1951±0.01050.2310±0.0059

Together, these results suggest that the improvements on PeMS are not simply due to averaged features, but also reflect the benefit of coupled higher-order recurrent states.

AuthorsOpenReview ↗

Author Response to Question Q4 - Part 2 of 2.

To complement these experiments, we evaluated STORM on the public PSML power-system dataset, which has a fixed topology and provides native voltage angles on 23 buses and active-power flows on 30 transmission lines. The boundary operator is the physical bus–line incidence matrix. We used 60,000 consecutive one-minute samples, aggregated into 12,000 five-minute samples, and evaluated horizons S=20,60,80 with three seeds. As shown in Table C.4, the best STORM/LightSTORM variant achieves the lowest final-step pooled node–edge NMSE at all three horizons, reducing error relative to the strongest competing baseline by 37.4%, 27.2%, and 19.9%, respectively.

Table C.4. PSML forecasting results. Final-step pooled node–edge NMSE (mean ± standard deviation over 3 seeds; lower is better). Bold: best. Italics: second best.

ModelS=20S=60S=80
LSTM0.006768 ± 0.0006280.027643 ± 0.0006200.024004 ± 0.000583
xLSTM0.008226 ± 0.0000800.031128 ± 0.0009900.029276 ± 0.004377
Transformer0.008455 ± 0.0004270.025807 ± 0.0009080.020966 ± 0.000219
iTransformer0.004961 ± 0.0003390.024889 ± 0.0000600.025914 ± 0.001394
GCN0.051488 ± 0.0094400.067024 ± 0.0018870.074483 ± 0.014218
GCN-xLSTM0.007053 ± 0.0010210.020820 ± 0.0010180.018586 ± 0.001126
SC-VAR0.014630 ± 0.0001960.051733 ± 0.0001560.053974 ± 0.000197
Best STORM0.003108 ± 0.0000540.015149 ± 0.0010270.016218 ± 0.002434
Best LightSTORM0.004063 ± 0.0002500.015687 ± 0.0016250.014883 ± 0.002360

Finally, we agree that evaluating STORM on datasets with native node, edge, and face measurements would be valuable. However, finding established temporal forecasting benchmarks with native signals across all three simplex orders remains challenging, as discussed in Appendix I.

We hope our responses and additional experimental results have addressed the reviewer's questions and concerns. We remain happy to provide further clarification if needed.

Reviewer_dSqNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces temporal representation learning with simplicial complexes as a structural inductive bias. It proposes two variants: STORM, which incorporates higher-order structural information into temporal representations through simplicial filters, and STORM-Light, which uses a Dirac-based construction for a lightweight alternative. The models are evaluated on traffic and hydraulic network datasets.

Critical strengths

Strenghts:

1: Simplicial aware recurrent updates i.e., capable of preserving higher order structural information. 2:Numerical experiments on hydraulic network data.

Critical weaknesses

W1: Novelty is the biggest concern

Graph filtering within LSTM gates is already established, STORM simply extends this to simplicial complexes to account for the higher order information levearging the simplicial filters. Consequently, incorporating spatial filtering into recurrent gates cannot itself establish STORM’s novelty. Specifically,

(A) : Given that the simplicial convolutional filters and the integration of filtered features to recurrent networks are well studied, the proposed combination appears to offer limited methodological novelty.

(B) : Theoretical guarantees provided doesn't add any new insights. In particular, the robustness and the permutation equivariance are well established.

W2: . In the traffic experiments, edge and triangle signals are deterministic lifts of node observations. The edge and triangle features are deterministic averages, rather than independently observed measurements, a strong node forecaster followed by the same lifting operations is therefore an essential missing baseline for the main multi-order task.

W3:I encourage the authors to strengthen the methodological novelty by developing mechanisms specifically designed to capture long-range dependencies. Evaluating these mechanisms on established long-range graph benchmarks, or introducing controlled datasets that require long-range reasoning, would help demonstrate their benefits beyond existing architectures.

Questions for authors

Refer weakness

8 comments
AuthorsOpenReview ↗

Author Response to W1.A and W1.B

We thank the reviewer for taking the time to review our work. We appreciate the feedback and concerns raised and address each point individually below, hoping to clarify our contributions and provide additional evidence where needed.

W1.A: Thank you for your feedback. We agree that graph filtering inside recurrent gates has been studied, but we respectfully disagree that STORM is simply an extension of existing graph-recurrent architectures. The main distinction is that STORM maintains coupled recurrent states across multiple simplex orders, using Hodge Laplacians for within-order filtering and boundary operators for cross-order interactions directly inside the recurrent gates at every timestep. Unlike graph-recurrent models, which typically operate on node signals and pairwise connectivity, STORM jointly models the temporal evolution of signals on nodes, edges, and higher-order simplices. This also differs from existing simplicial temporal models. Chen et al. (2022b, 2023) use higher-order information for node-level forecasting, Mitchell et al. (2024) applies simplicial convolutions before recurrence, and Krishnan et al. (2024) considers linear autoregressive dynamics. In contrast, STORM jointly evolves nonlinear recurrent states across simplex orders, with topology directly controlling memory updates. To our knowledge, this is the first gated recurrent architecture with this form of simplicial coupling. The ablation in Table 29 in Appendix H.3 further supports this distinction: adding coupled edge states to graph recurrence improves node forecasting in all six evaluated settings.

W1.B: We agree that permutation equivariance and stability are established properties in graph neural networks and recurrent models. Our contribution is to characterize these properties for STORM's coupled multi-order recurrent dynamics. In particular, Theorem 1 bounds how perturbations of the simplicial operators affect the predictions over time. Under non-expansive dynamics, the bound grows at most linearly with time, O(tη)O(t\eta), and becomes uniform, O(η)O(\eta), under contraction. As discussed in Appendix E.3 (Remark E.2), this differs from the quadratic temporal dependence, O(t2η)O(t^2\eta), obtained for graph recurrent models by Ruiz et al. (2020).

Moreover, Theorem 2 provides a finite-horizon generalization bound using Rademacher complexity. It explicitly characterizes how sensitivity to the shared recurrent parameters accumulates over the sequence length through ΨT(qλ)=∑j=0T−1qλj\Psi_T(q_\lambda)=\sum_{j=0}^{T-1}q_\lambda^j. When qλ<1q_\lambda<1, this temporal factor remains bounded independently of the sequence length. This provides an additional theoretical characterization beyond equivariance and perturbation stability.

AuthorsOpenReview ↗

Author Response to W2

W2: Thank you for suggesting this experiment. We executed it, evaluating three established node forecasters (DCRNN, Graph WaveNet, and STID), followed by the same deterministic lifting operations used to construct the higher-order targets. We then evaluated their predictions using the same pooled multi-order NMSE as STORM.

The results are reported in Tables B.1 and B.2. The best STORM/LightSTORM variant achieves lower NMSE than all three lifted baselines at every horizon S≥10S \geq 10. At S=80S=80, it reduces the error relative to the strongest lifted baseline (STID) by 21.5% on PeMS-D4 and 39.8% on PeMS-D8. These results suggest that joint multi-order recurrent modeling provides a benefit over forecasting nodes and reconstructing the higher-order signals afterward, particularly at longer horizons.

Table B.1. Node forecasting followed by deterministic lifting on PeMS-D4. Final-step pooled NMSE (mean ± standard deviation over 5 seeds; lower is better). Bold: best. Italics: second best.

ModelS=5S=10S=20S=40S=60S=80
DCRNN + lift0.0483±0.00020.0789±0.00280.1930±0.03110.4018±0.01810.6548±0.02120.8595±0.0635
Graph WaveNet + lift0.0500±0.00050.0756±0.00310.1407±0.00560.2899±0.01170.3969±0.03370.5179±0.0308
STID + lift0.0488±0.00020.0762±0.00350.1356±0.00240.2678±0.00430.3558±0.00480.4031±0.0079
Best STORM/LightSTORM0.0491±0.00020.0725±0.00040.1206±0.00200.2247±0.00330.2686±0.00310.3164±0.0017

Table B.2. Node forecasting followed by deterministic lifting on PeMS-D8. Final-step pooled NMSE (mean ± standard deviation over 5 seeds; lower is better). Bold: best. Italics: second best.

ModelS=5S=10S=20S=40S=60S=80
DCRNN + lift0.0408±0.00060.0619±0.00310.1326±0.00370.3173±0.03320.5309±0.03800.7433±0.0520
Graph WaveNet + lift0.0411±0.00180.0566±0.00150.1149±0.00570.2357±0.00880.3497±0.02550.4582±0.0729
STID + lift0.0375±0.00040.0545±0.00080.1009±0.00230.2057±0.00270.3116±0.00460.3838±0.0065
Best STORM/LightSTORM0.0402±0.00050.0541±0.00070.0953±0.00190.1592±0.00930.1951±0.01050.2310±0.0059
AuthorsOpenReview ↗

Author Response to W3

W3: We thank the reviewer for these suggestions. Evaluating STORM on dedicated long-range dependency benchmarks would be an interesting direction for future work. We would like to clarify that our current work focuses on topology-aware recurrent temporal representation learning, rather than introducing a mechanism specifically designed for long-range reasoning. Nevertheless, we evaluate forecasting horizons up to S=80, corresponding to 6 hours and 40 minutes on PeMS, compared with the commonly used 12-step (one-hour) setting. Our empirical results show that STORM/LightSTORM outperforms the evaluated baselines on the multi-order forecasting task at nearly all horizons.

We hope our responses and additional experimental results have clarified the methodological and theoretical contributions and addressed the reviewer's concerns. We remain happy to provide further clarification if needed.

Reviewer_dGvMOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a family of gated recurrent architectures for multiorder signals on simplicial complexes, resulting in a Simplicial Topological Recurrent Model (STORM). The primary motivation is that many time-varying objects, including time series, exhibit various forms of higher-order dependencies that cannot be well described by the more conventional tools. As an alternative, we can describe such higher-order inter-dependencies using simplicial complexes and then apply message propagation over simplices by leveraging the discrete Hodge theory and combinatorial Laplacians. Such ideas, typically known as topological deep learning, have been successfully applied within the context of graph neural architectures. This paper brings these simplicial ideas to the interface of gated recurrent architectures. Particularly, STORM employs the LSTM-style gated recurrent cell with an exponential forget gate. I find the idea being original and new, and overall the method has a high potential in terms of flexibility of treating time-evolving objects. I also have a number of concerns though generally I think that they are addressable.

Critical strengths

STORM leverages simplicial representations within gated recurrent architectures, enabling higher-order descriptors to shape hidden-state dynamics over time. This is an important new step bringing simplicial representations to LSTMs. Overall, the STORM approach has a lot of potential in terms of versatility in modeling spatio-temporal processes and routes connecting with statistical inference and methods for time series.

The authors also present a simpler version LightSTORM based on a Dirac-operator variant, sharing parameters across simplex orders, which appears to be also quite competitive.

The paper is overall well written and I enjoyed reading it.

Critical weaknesses

I have a number of major concerns though I believe that they shall be addressable:

  1. The list of competitors is very limited. STORM and its variants must be compared with the state-of-the-art graph-based networks for spatio-temporal forecasting and spatio-temporal networls, including but not limited to those that integrate topological summaries. The current list of the considered competitors, i.e. GCN-xLSTM, SC-VAR, and SCRNN, is just super-tiny and non-convincing.

  2. The considered datasets (i.e., PeMS-D4 and PeMS-D8), except of the hydraulic network, do not convey the intuition behind the need of simplicial representation, and simplicial descriptors largely look artificial for PeMS-D4 and PeMS-D8. I would not be surprised if some newest SOTAs will outperform STORM on PeMS-D4 and PeMS-D8. I'd suggest to look at other datasets where, similar to the hydraulic network, the notion of simplices, their boundaries and interactions among simplices, is transparent. One possible source may be power flow network but anything else with the apparent role of simplices and boundaries among them shall be good. This will make STORM more convincing and hopefully allow it shine better.

  3. A much more extensive overview of the related work on spatio-temporal forecasting, including graph abstractions, is needed (see also Q1 above).

Questions for authors

In addition to Critical Weaknesses, some analysis of autocorrelations or other measures of weak dependence in time series will be of help too. Notice that STORM errors increase with SS, as expected, and the increase is pretty rapid. The competing models in contrast show different rates of deterioration. It'll be helpful to connect the STORM rate of decay over increasing SS with autocorrelations, which may assist with interpretability and also potentially help improving the performance.

6 comments
AuthorsOpenReview ↗

Author Response to W1

We thank the reviewer for taking the time to review our work and for appreciating the novelty and potential of our work. We are glad that the reviewer enjoyed reading our paper, we appreciate it.

We appreciate the constructive feedback and suggestions and address them individually below.

W1: Thank you for raising this concern. In addition to the multi-order forecasting baselines in Tables 1–3, including iTransformer, a strong multivariate time-series forecasting model, we compare against strong spatio-temporal graph architectures such as DCRNN, Graph WaveNet, STID, and STAEformer on node-only forecasting (Table 4). We consider these settings separately because our main task jointly forecasts signals across multiple simplex orders, whereas these graph-based baselines predict node-level signals.

Following a suggestion from another reviewer, we conducted additional experiments with DCRNN, Graph WaveNet, and STID, applying the same deterministic lifting operations to their node predictions to obtain edge and triangle predictions. As shown in Tables A.1 and A.2, the best STORM/LightSTORM variant achieves lower pooled multi-order NMSE than all three lifted baselines at every horizon S≥10S \geq 10. At S=80S=80, the error is reduced by 21.5% on PeMS-D4 and 39.8% on PeMS-D8 relative to the strongest lifted baseline, STID.

Table A.1. Node forecasting followed by deterministic lifting on PeMS-D4. Final-step pooled NMSE (mean ± standard deviation over 5 seeds; lower is better). Bold: best. Italics: second best.

ModelS=5S=10S=20S=40S=60S=80
DCRNN + lift0.0483±0.00020.0789±0.00280.1930±0.03110.4018±0.01810.6548±0.02120.8595±0.0635
Graph WaveNet + lift0.0500±0.00050.0756±0.00310.1407±0.00560.2899±0.01170.3969±0.03370.5179±0.0308
STID + lift0.0488±0.00020.0762±0.00350.1356±0.00240.2678±0.00430.3558±0.00480.4031±0.0079
Best STORM/LightSTORM0.0491±0.00020.0725±0.00040.1206±0.00200.2247±0.00330.2686±0.00310.3164±0.0017

Table A.2. Node forecasting followed by deterministic lifting on PeMS-D8. Final-step pooled NMSE (mean ± standard deviation over 5 seeds; lower is better). Bold: best. Italics: second best.

ModelS=5S=10S=20S=40S=60S=80
DCRNN + lift0.0408±0.00060.0619±0.00310.1326±0.00370.3173±0.03320.5309±0.03800.7433±0.0520
Graph WaveNet + lift0.0411±0.00180.0566±0.00150.1149±0.00570.2357±0.00880.3497±0.02550.4582±0.0729
STID + lift0.0375±0.00040.0545±0.00080.1009±0.00230.2057±0.00270.3116±0.00460.3838±0.0065
Best STORM/LightSTORM0.0402±0.00050.0541±0.00070.0953±0.00190.1592±0.00930.1951±0.01050.2310±0.0059

Overall, our goal is to learn spatio-temporal representations across multiple simplex orders. These results show that STORM remains competitive with strong graph-based models while providing benefits for joint multi-order forecasting, even when higher-order signals are deterministically constructed. We believe the flexibility of these representations is a valuable contribution and could also support other downstream tasks beyond forecasting.

AuthorsOpenReview ↗

Author Response to W2

W2: Thank you for this suggestion. We conducted additional experiments on the public PSML power-system dataset. It provides voltage angles on 23 buses and active power flow on 30 transmission lines and has a fixed topology. The boundary operator is defined by the physical bus-line incidence matrix. We used 60K consecutive one-minute samples, aggregated into 12K five-minute samples. We evaluated horizons S=20,60,80 (100, 300, and 400 minutes), using 3 seeds. We report the results on Table A.3. The best STORM/LightSTORM variant achieves the lowest error at all three horizons, reducing final-step NMSE relative to the strongest competing baseline by 37.4%, 27.2%, and 19.9%, respectively.

Table A.3. PSML forecasting results. Final-step pooled node–edge NMSE (mean ± standard deviation over 3 seeds; lower is better). Bold: best. Italics: second best.

ModelS=20S=60S=80
LSTM0.006768 ± 0.0006280.027643 ± 0.0006200.024004 ± 0.000583
xLSTM0.008226 ± 0.0000800.031128 ± 0.0009900.029276 ± 0.004377
Transformer0.008455 ± 0.0004270.025807 ± 0.0009080.020966 ± 0.000219
iTransformer0.004961 ± 0.0003390.024889 ± 0.0000600.025914 ± 0.001394
GCN0.051488 ± 0.0094400.067024 ± 0.0018870.074483 ± 0.014218
GCN-xLSTM0.007053 ± 0.0010210.020820 ± 0.0010180.018586 ± 0.001126
SC-VAR0.014630 ± 0.0001960.051733 ± 0.0001560.053974 ± 0.000197
Best STORM0.003108 ± 0.0000540.015149 ± 0.0010270.016218 ± 0.002434
Best LightSTORM0.004063 ± 0.0002500.015687 ± 0.0016250.014883 ± 0.002360

PSML dataset contains native node and edge signals, but no face measurements. As discussed in Appendix I, established temporal forecasting benchmarks with native signals across all three simplex orders remain difficult to find. We will clarify this limitation in the revised manuscript.

AuthorsOpenReview ↗

Author Response to W3

W3: We appreciate the reviewer's suggestion to expand the related work discussion. Appendix B already covers graph-based spatio-temporal forecasting and topological temporal models, but we agree that this discussion could be more detailed. We will expand this discussion, particularly for DCRNN, Graph WaveNet, STID, and STAEformer, explaining how they model spatial and temporal dependencies and how they relate to STORM.

Comments on the paper

1 comment

Dear Reviewers,

As we enter the final stage of the author-reviewer discussion, I would like to encourage everyone to actively engage with the authors’ responses and any additional evidence they have provided.

Please take this opportunity to reconsider your initial assessments with an open mind, particularly where the authors have offered substantive clarifications, detailed arguments, or additional experimental results. The goal of the discussion is not to defend initial positions, but to reach the fairest and most informed evaluation of the submission.

I encourage you to engage directly with the authors on any remaining concerns, clearly distinguishing between issues that have been satisfactorily addressed and those that remain unresolved. Where appropriate, please update your reviews and scores accordingly, or explain why your original assessment still stands.

Your active and constructive participation in these final exchanges is essential to reaching a fair, evidence-based decision.

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.