acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Cyclic Neural ODEs

Abstract

Biological systems are inherently cyclical, operating through feedback loops where epigenetic modifications regulate transcription, driving protein synthesis and metabolic flux, which ultimately modulate the epigenome. Modeling these continuous, high-dimensional dynamics remains challenging because the underlying molecular layers are operating on disparate timescales. While standard Neural Ordinary Differential Equations (NODEs) are widely recognized for modeling continuous-time processes, they frequently experience timescale collapse and numerical stiffness when forced to map extreme multi-scale interactions. To stably learn these hierarchical oscillatory systems, the Hierarchical Cyclic Neural ODE is introduced. A Stuart-Landau limit-cycle prior is embedded into the model's latent vector field, constraining the neural network to focus primarily on the biological variations that deviate from an underlying rhythm. A cyclic small-gain analysis provides a theoretical guarantee that this architecture yields unconditionally bounded trajectories (preventing numerical instability) and exponential orbital attraction (ensuring perturbed states converge to a homeostatic orbit). Evaluated against leading continuous-time models including scNODE and Heavy Ball NODE, on a four-layer mouse liver atlas and a two-layer course in constant darkness, this architecture maintains stable trajectories through the forecast window and reduces held-out error on both. In contrast, acyclic ablations suffer from temporal mode collapse and amplify perturbations over a full cycle, confirming that closed-loop structural feedback is necessary to bound system energy. Leveraging these continuous dynamics, an analysis of the system's phase response demonstrates that a perturbation confined to a single layer produces a lasting phase shift in layers reachable only around the loop, a response that vanishes identically when the single return edge is removed.

Primary area:
learning on time series and dynamical systems
Keywords:
Neural ODEs, Dynamical Systems, Input-to-State Stability, Limit Cycles, Multi-omics, Systems Biology

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Z7sQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies rhythmic multi-omics dynamics involving feedback between molecular layers that operate on different timescales. The authors propose a Hierarchical Cyclic Neural ODE that combines a closed-loop topology, a Stuart–Landau oscillatory prior, and orbit-centered neural residuals. They use cyclic small-gain analysis to study the boundedness and regional orbital stability of the latent dynamics and to guide the design of a stability regularizer for training. Experiments evaluate on two datasets: a mouse liver dataset with four omics layers and a Neurospora crassa dataset with two layers. The method achieves the lowest mean MSE for each of the three evaluated omics layers in Table 1 and the lowest average MSE across the two layers in Table 2.

Critical strengths
  1. The architecture combines explicit feedback between omics layers, a Stuart–Landau oscillatory prior, and orbit-centered neural residuals. This provides a structured latent dynamical model for rhythmic multi-omics time series.
  2. The cyclic small-gain analysis provides theoretical support for trajectory boundedness and exponential orbital attraction, offering a stability perspective on the proposed architecture.
Critical weaknesses
  1. The exponential attraction result in item 2 of Theorem 1 requires the initial state to satisfy z0∈B ={z:V(e)≤c⋆}. However, the paper does not clarify how this regional guarantee applies to the initial states used for forecasting or the perturbed states in the phase response experiments. Figure 4 and Appendix F provide loop gain monitoring and local stability diagnostics, but do not establish whether these experimental states satisfy the required regional condition.
  2. The ablations change multiple components simultaneously, making it difficult to attribute performance gains to individual design choices. Cyclic No-Prior removes both the oscillatory prior and the stability penalty, while Acyclic Cascade replaces the oscillatory prior with linear damping and removes the return edge. These comparisons do not fully isolate the individual contributions of the oscillatory prior and the feedback topology to forecasting performance.
Questions for authors
  1. Could the authors verify whether the initial latent states used for forecasting and the perturbed states in the phase response experiments satisfy the certified regional condition V(e)≤c⋆?
  2. How does removing only the return edge, while retaining the Stuart–Landau prior, affect forecasting performance?
  3. How does replacing the Stuart–Landau prior with linear damping affect forecasting performance when the closed-loop topology is retained under comparable training settings?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Z7sQ

We thank the reviewer for feedback regarding stability guarantees and ablations.

1. Initial States, Perturbations, and the Certified Region (W1, Q1)

By retraining the mouse liver model with the stability penalty evaluated at the inner radius δk=0.75rk∗\delta_ k = 0.75 r_ k^{\ast} rather than in the δk→rk∗\delta_ k \to r_ k^{\ast} limit, the loop gain at that radius is ρ\rho = 0.991, 0.990 and 0.974 on the three seeds (mean 0.985), each below one, which certifies the invariant region B={V(e)≤c∗}\mathcal{B} = \lbrace V(\mathbf{e}) \le c^{\ast} \rbrace with c∗=min⁡kpk(rk∗−δk)c^{\ast} = \min_ k p_ k (r_ k^{\ast} - \delta_ k). The reported half-width of 5.2% (mean over the three seeds, ranging from 4.6% to 5.6%) is the largest radial deviation a single layer can carry inside B\mathcal{B}, c∗/pkc^{\ast}/p_ k, expressed as a fraction of rk∗r_ k^{\ast} and taken at the most restrictive layer. The encoded initial state z(0)\mathbf{z}(0) starts outside this region, but its trajectory enters it within 2.5 h on every seed and remains inside through the forecast window, at no cost in forecasting accuracy (Section 3.4, lines 298 to 305). The impulses of 0.30rk∗0.30 r_ k^{\ast} used in the phase response experiments place the perturbed states outside this certified region, so their return to Γ\Gamma is an empirical observation rather than a consequence of Theorem 1, and in every perturbed run the trajectory contracts back to the limit cycle.

2. Isolated Ablations (W2, Q2, Q3)

We evaluated targeted ablations across 6 seeds (for van der Pol, 5 on Neurospora and 4 on mouse, as it diverged on the remaining seeds) by testing alternative geometries (Cyclic + van der Pol) against an unpenalized control (Cyclic + SL Prior (No Penalty)), replacing the Stuart-Landau prior with linear damping (Cyclic + Linear Damping), and retaining the Stuart-Landau prior while severing the return edge (Acyclic + SL Prior). As detailed in the General Response, Tables 1 and 2, Section 4.3 (lines 430 to 472), and Table 5 (Appendix F):

  • Replacing the Stuart-Landau prior with linear damping or a van der Pol geometry eliminates orbital closure and Floquet stability, introducing expanding directions, causing gradient divergence, and degrading forecasting MSE across layers. With the closed loop, seeds, optimizer and epoch budget unchanged, linear damping reaches 2.126, 3.895 and 3.100 on the mouse metabolome, proteome and transcriptome against 1.867, 2.048 and 1.752, and 3.245 and 1.615 against 1.730 and 1.525 on Neurospora (Q3).
  • Severing the return edge eliminates all 6 cross-layer phase response paths dependent on metabolic-to-epigenetic feedback (Section 4.4, Figure 3, and Table 3 in Appendix C) and yields slightly higher forecasting MSE than the full closed-loop model on every layer of both datasets (Q2), at 1.899, 2.058 and 1.766 against 1.867, 2.048 and 1.752 on mouse and 1.737 and 1.530 against 1.730 and 1.525 on Neurospora.

Together, the limit-cycle prior provides orbital stability while the closed-loop topology enables causal cross-layer feedback and achieves the strongest forecasting accuracy.

Thanks for the responses, which have solved my major concerns. I will maintain my positive rating.

Thank you for following up. The isolated ablations you suggested, namely, removing only the return edge and replacing only the prior, now more clearly separate the contributions of the prior and the topology in both Tables 1 and 2. We have also added a third dataset, the yeast metabolic cycle, to test forecasting over two unseen cycles (Appendix H).

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

This paper proposes a computational framework for studying the dynamics (particularly periodic one) in multi-omics. The framework adopts Stuart-Landau prior to stabilize the trajectory generated by the model to an orbit. Additionally, the authors theoretically show the stability of the method. The experimental results includes comparison to various methods (with or without prior), and detailed ablation study.

Critical strengths

This paper provides a seemingly promising framework for studying the dynamics in multi-omics. The framework is theoretically justified and performs much better than the competing methods. Additionally, the paper is well-structured and easy to follow even though a reader who is not doing research in this particular field (e.g., me).

Critical weaknesses

First, I have to admit that this paper is not in the field of my expertise; therefore, I will only ask questions that much align my expertises.

  1. Small gain stability penalty uses 2∣λk∣−Lkself2|\lambda_k|-L^{self}_k for AkA_k. However, it is somewhat looser than the condition required for attraction (e.g., ρ<1\rho<1). How can this work in practice? If it is empirically, can you show some failure cases due to this?

  2. The reconstruction loss is based on the sum of per-timestamp MSE. As I know, this kind of temporal sum of MSE (i.e., integrated squared error) can be difficult to train. Is there any hidden ingredient for initialization or why the training is naturally stable?

  3. Now, this paper contains lots of variables and notations. While the authors try their best for explaining them in the place they first show up, it is high recommended to include a table to explain the variables and notations.

Questions for authors
  1. why zkz_k is partitioned into two blocks? I assume it is for the S.-L. prior. If this is correct, it is recommended to include short explanation to this.
  2. The radial projection PkP_k should be explained more. It is hard to understanding its definition based on the context now.
  3. What is the computational time for estimate LkL_k by using power iteration? Would it be a bottleneck of the training pipeline?
  4. What is failure cases of the proposed method? What if the underlying dynamics is not periodic?
3 comments
AuthorsOpenReview ↗

Response to Reviewer uMtq

We thank the reviewer for feedback clarifying architecture mechanics.

1. Looseness of the Penalty and Failure Cases (W1)

The reviewer is correct that the penalty evaluated at δk→rk∗\delta_ k \to r_ k^{\ast} is looser than the condition ρ<1\rho < 1 required for attraction to a region of nonzero size, since in that limit the certified region B\mathcal{B} shrinks to a thin tube around Γ\Gamma. To test whether the bound can be made useful in practice, we retrained the mouse model with the penalty evaluated at δk=0.75rk∗\delta_ k = 0.75 r_ k^{\ast}, which reaches ρ≈0.985\rho \approx 0.985 across three seeds and certifies a region extending 5.2% of rk∗r_ k^{\ast} on either side of Γ\Gamma with no loss in forecasting accuracy (Section 3.4, lines 291 to 305). A failure case appears when the original penalty is evaluated at that same radius after training, where it gives ρ\rho between 2.5 and 4.3, so the models trained with the looser penalty are certified only in a thin neighborhood of the orbit even though their trajectories remained bounded in every run (Appendix A, following Figure 4).

2. Training Stability (W2)

The key mechanism preventing exploding gradients during long-horizon temporal MSE training is the Stuart-Landau prior. At initialization, the output layers of each residual network are rescaled so that the stability certificate already holds at epoch 0, and because the residual vanishes on the orbit by construction, the trajectories follow the prior dynamics early in training and remain bounded from the first epoch (Section 3.5, lines 338 to 341, and Figure 4 in Appendix A). The new control without the penalty and without this rescaling isolates the prior as the mechanism, since it completed every run without a single skipped gradient update, while the matched van der Pol variant diverged during training on 3 of 12 runs.

3. Notation, Partitioning, and Radial Projection (W3, Q1, Q2)

We added a notation table to Appendix G (Table 6). To clarify the mechanics: zk∈Rdk\mathbf{z}_ k \in \mathbb{R}^{d_ k} is partitioned into dk/2d_ k/2 coordinate pairs zk,j=[xk,j,yk,j]⊤\mathbf{z}_ {k,j} = [x_ {k,j}, y_ {k,j}]^\top to form the 2D planar oscillators required by the Stuart-Landau equations. Radial projection Pk=ΠΓ(zk)\mathbf{P}_ k = \Pi_ \Gamma(\mathbf{z}_ k) normalizes each 2D oscillator pair by ∥zk,j∥2\Vert \mathbf{z}_ {k,j}\Vert _ 2 to project the state onto the homeostatic limit-cycle radius rk∗=λk/γkr_ k^{\ast} = \sqrt{\lambda_ k / \gamma_ k} (Section 3.3, Equation 3, lines 230 to 235).

4. Spectral Norm Overhead (Q3)

Because the number of macroscopic omics layers is small (K=2K=2 for Neurospora, K=4K=4 for mouse liver) and the latent dimension per layer is fixed at dk=8d_ k = 8, weight matrices in each sub-network fθkf_ {\theta_ k} are low-dimensional. Consequently, we compute each spectral norm exactly by singular value decomposition rather than by power iteration, which adds negligible overhead and does not bottleneck the ODE solver (Section 3.5, lines 336 to 338).

5. Failure Cases (Q4)

The primary failure case is a non-periodic, transient biological event (e.g., an acute shock or terminal differentiation trajectory). If underlying dynamics lack periodic behavior, the Stuart-Landau prior inappropriately biases the trajectory toward a limit cycle, increasing reconstruction error. We added this discussion to the Limitations paragraph of the Conclusion (Section 5, lines 530 to 533).

Thank the authors for addressing my concerns. I believe the responses have adequately addressed my concerns, and I appreciate the additional clarifications and experiments.

We appreciate you following up. Your questions led to the retrained certificate and the notation table in Appendix G, which are both in the revised manuscript and have made the paper more complete. We have also added an additional dataset, the yeast metabolic cycle, to test forecasting over two unseen cycles (Appendix H).

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

The paper proposes a hierarchical cyclic latent Neural ODE for multi-omics time series, combining layer-wise Stuart–Landau oscillators, orbit-centered neural residuals, and cross-layer couplings. A cyclic small-gain analysis establishes boundedness and conditional regional attraction to an invariant torus. Validation includes next-cycle forecasting on mouse liver and Neurospora datasets, architectural ablations, and model-based cross-layer phase-response analysis. Forecasting generally improves over the evaluated NODE baselines.

Critical strengths
  1. The architecture addresses a relevant failure mode of unconstrained continuous-time models: losing bounded, sustained oscillations during extrapolation.
  2. The orbit-centered residual makes the target invariant set explicit and enables an interpretable small-gain argument.
Critical weaknesses
  1. The specific benefits of the chosen prior and cyclic topology remain insufficiently isolated. The acyclic ablation changes both the prior and feedback topology, while the no-prior variant also disables stability regularization. Consequently, the gains cannot be attributed specifically to Stuart–Landau dynamics or cyclic feedback.
  2. The evaluation covers only two sparse time courses, with a single forecast cycle and only one random seed on the second dataset.
Questions for authors
  1. How does the model compare with a matched alternative bounded oscillatory prior under the same cyclic architecture and prescribed periods?
  2. What are the forecasting results when only the return edge is removed while retaining the same Stuart–Landau prior, prescribed periods, and training protocol?
  3. How does the proposed method compare with a simple periodic model, such as harmonic regression, using the same prescribed periods?
4 comments
AuthorsOpenReview ↗

Response to Reviewer SqiM

We thank the reviewer for feedback regarding architecture isolation and evaluation scope. We address the questions by presenting the targeted ablations and baselines across 6 independent random seeds.

1. Alternative Prior and Removed Return Edge (W1, Q1, Q2)

Regarding Q1, any smooth dynamical system passing through a supercritical Hopf bifurcation reduces near onset to the Stuart-Landau normal form, establishing the fundamental mathematical template for a bounded oscillator. To empirically test an alternative, we trained the matched alternative which is a van der Pol oscillator in place of the Stuart-Landau prior, with the same closed loop, the same prescribed period, the same initial weights, and the residual centered on the circle that the van der Pol orbit approaches at small μ\mu.

The structural mismatch of the van der Pol geometry causes severe training instability in which the van der Pol model suffers from gradient divergence on 3 of 12 runs (seed 0 on both datasets and seed 5 on mouse), failing to complete training. On the seeds where it survives, forecasting performance is heavily degraded (e.g., averaging an MSE of 2.315 on Neurospora transcriptome and 7.053 on mouse transcriptome, compared to 1.730 and 1.752 for our Stuart-Landau model). Because Theorem 1 does not cover a van der Pol prior, that model trains without our penalty; however, a Stuart-Landau control trained the same way (Cyclic + SL No Penalty) successfully stabilizes and converges (reaching 1.733 and 1.753 on the same two layers), confirming the performance difference is due to the prior itself, not the penalty (Tables 1 and 2, and Section 4.3, lines 466 to 472).

To address Q2, we evaluated the Acyclic + SL Prior and Cyclic + Linear Damping models. As detailed in the General Response, Tables 1 and 2, Section 4.3 (lines 461 to 466), and Table 5 in Appendix F, replacing the limit-cycle prior with linear damping or a van der Pol oscillator eliminates Floquet orbital stability. Severing only the return edge, with the same Stuart-Landau prior, prescribed period and training protocol, removes causal metabolic-to-epigenetic phase transmission (Section 4.4 and Table 3 in Appendix C) and gives slightly higher held-out MSE on every layer, at 1.899, 2.058 and 1.766 against 1.867, 2.048 and 1.752 on the mouse metabolome, proteome and transcriptome, and 1.737 and 1.530 against 1.730 and 1.525 on Neurospora.

2. Harmonic Regression and Simple Periodic Models (Q3)

We include a Cosinor (harmonic regression) baseline in Table 1 (Mouse) and Table 2 (Neurospora), described in Section 4.1, lines 377 to 378. While period-aware continuous-time models drift, our dynamic causal architecture achieves forecasting accuracy competitive with Cosinor (outperforming Cosinor on Mouse Metabolomics at 1.867 vs. 1.876 and Transcriptomics at 1.752 vs. 1.818, while Cosinor serves as a static statistical reference on Proteomics at 1.955 vs. 2.048). Unlike Cosinor, our model achieves this while retaining the capacity for structural interventions and phase response analysis.

3. Evaluation Scope (W2)

Both datasets are now evaluated over the same 6 seeds, which replaces the single seed previously reported on Neurospora. We agree that both time courses are sparsely sampled and that each provides a single held-out cycle, so the present results establish forecasting over one cycle ahead, and longer and denser series would be needed to test forecasting across several cycles.

Please see the General Response and revised PDF for the 6-seed empirical tables, Floquet spectra, and Phase Response Curves.

AuthorsOpenReview ↗

Follow-up on W2 for multi-cycle forecasting

Following up on our initial response, we would like to address W2, the single held-out cycle in the two circadian datasets, more directly.

We have evaluated the model on a third system with a longer horizon, the yeast metabolic cycle (Tu et al., 2005; 2007). Its transcriptome is measured over three consecutive cycles, so we train on the first cycle and forecast the next two. The revised PDF reports this in a new Appendix H (Table 7) and summarizes it in Section 4.2.

Across six seeds, the transcriptome error of our model rises from 0.460 in the second cycle to 0.527 in the third, while the errors of the four period-aware neural ODE baselines rise three- to fifteen-fold over the same step. Removing the limit-cycle prior raises the mean error to 29.23, as it does on the circadian datasets.

Against a cosinor fit at the known period, our model has a lower error over all held-out observations on every seed (0.726 against 0.736). It is 5% lower on the metabolome, within 1% on the second transcriptome cycle and 6% higher on the third. Removing only the return edge leaves accuracy unchanged on this dataset, so the role of the return edge remains the closed-loop phase transmission of Section 4.4.

We have also updated the supplementary code with the YMC preprocessing script, configuration and processed data. We hope this addresses your concern about the forecasting horizon, and we would be glad to answer any other questions before the discussion closes.

Thank you for the detailed responses and additional experiments, which substantially address my original concerns. I am happy to raise my score. As a future direction in real-world applications, it would be useful to discuss the relationship to existing foundation models for dynamical systems and whether learned dynamical priors could complement the prescribed oscillatory structure to accommodate a broader range of biological dynamics.

Thank you for reviewing the additional experiments and for raising your score. We agree that learned dynamical priors from foundation models could complement the prescribed oscillatory structure in our model. We have added this as future work in the conclusion of the revised version.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: 6-Seed Evaluation, New Baselines, and Isolated Ablations

We thank all reviewers for their feedback and for recognizing the theoretical soundness and relevance of the Hierarchical Cyclic Neural ODE.

To address questions regarding structural contributions, statistical baselines, and multi-seed robustness, our evaluation spans 6 independent random seeds for both the 4-layer mouse liver atlas and the 2-layer Neurospora datasets.

The revised manuscript incorporates these results alongside an updated code package, and all changes relative to the submitted version are marked in blue. We summarize the three primary empirical updates below:

1. Harmonic Regression and Baseline Comparison (Reviewer SqiM)

To ensure forecasting gains do not rely solely on recognizing the biological period, we provided known periods (24h for mouse, 22h for Neurospora) to all baselines as external time features. We also included Cosinor (harmonic regression) as a statistical reference for static periodicity.

Model (Neurospora, 6-Seed Mean)Transcriptome MSEProteome MSE
Hierarchical Cyclic (Ours)1.7301.525
scNODE (Period-Aware)2.7144.496
Jacobian Reg. (Period-Aware)9.16914.040
Vanilla NODE (Period-Aware)9.92814.640
Heavy Ball NODE (Period-Aware)11.04019.850
Cosinor (Static Harmonic Reference)1.6501.459

While standard continuous-time models drift despite knowing the period, our architecture approaches the harmonic performance of Cosinor. Unlike static Cosinor models, our architecture retains the capacity for structural interventions and cross-layer phase response analysis (Table 1, Section 4.2 lines 386 to 387, and Table 2 in Appendix B).

2. Isolating Topology vs. the Stuart-Landau Prior (Reviewers SqiM, Z7sQ)

To isolate the contributions of feedback topology and the oscillatory prior, we evaluate four targeted ablations:

  • Cyclic + van der Pol: Replaces the Stuart-Landau prior with a van der Pol oscillator to test an alternative limit-cycle geometry.
  • Cyclic + SL Prior (No Penalty): A control model that retains the Stuart-Landau prior but disables the stability penalty (since Theorem 1 does not cover van der Pol, this ensures a fair comparison).
  • Acyclic + SL Prior: Retains the Stuart-Landau prior but removes the return edge.
  • Cyclic + Linear Damping: Retains the cyclic feedback topology but replaces the Stuart-Landau prior with linear damping.
Ablation (Neurospora, 6-Seed)Trans. MSEProt. MSEFloquet Stability
Hierarchical Cyclic (Ours)1.7301.525Stable (8 neutral dirs)
Cyclic + SL Prior (No Penalty)1.7331.528Stable (8 neutral dirs)
Acyclic + SL Prior1.7371.530Stable (8 neutral dirs)
Cyclic + van der Pol (5 seeds, 1 diverged)2.3154.570Unstable (1 expanding dir)
Cyclic + Linear Damping3.2451.615Unstable (3 expanding dirs)
Cyclic + No Prior6.96613.670Unstable (6 expanding dirs)

Among the priors we tested, only the Stuart-Landau prior trains stably on every seed and keeps the closed feedback loop orbitally stable. Alternative geometries (van der Pol) or dissipative kinetics (linear damping) introduce expanding Floquet directions or cause severe gradient divergence. Meanwhile, the cyclic topology enables causal biological feedback and achieves the lowest forecasting error across all layers. Specifically, our Phase Response Curve (PRC) analysis demonstrates that cross-layer transient perturbations yield lasting phase shifts in downstream targets, a biological feedback signal that vanishes to 0.0000 when the return edge is severed (see Section 4.4, Figure 3, and Table 3 in Appendix C).

3. Theorem 1 Bounding Region (Reviewers Z7sQ, uMtq)

We clarify how the regional guarantee V(e)≤c∗V(\mathbf{e}) \le c^{\ast} applies to forecasting states. To certify a region of stated size, we retrained the mouse model with the penalty evaluated at δk=0.75rk∗\delta_ k = 0.75 r_ k^{\ast}. Across three seeds, the loop gain at that radius is ρ≈0.985\rho \approx 0.985 and the certified invariant region B\mathcal{B} extends 5.2% of rk∗r_k^{\ast} on either side of Γ\Gamma. The trajectory from the encoded initial state enters B\mathcal{B} within 2.5 h and remains inside for the rest of the 48 h window, so the guarantee covers every forecast timepoint. The held-out errors for this certified model are 1.860, 2.046, and 1.750 on the metabolome, proteome and transcriptome against 1.867, 2.048 and 1.752 for the primary model of Table 1, so certifying the larger region costs no forecasting accuracy (Section 3.4, lines 298 to 305).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.