acceptodds
Under review as a conference paper at ICLR 2027

Hyperbolic Time Series Forecasting with Learnable Decomposition and Manifold Dynamics

Abstract

Multiscale structure is central to time series forecasting: temporal signals contain slowly varying trends, intermediate seasonal patterns, and high-frequency residuals. Traditional forecasters model this multiscale structure in Euclidean space, without representing the ordering it induces among components. We identify that cascaded decomposition orders components by their characteristic frequencies, forming a natural frequency hierarchy. This hierarchy aligns with the exponential volume growth of hyperbolic space, which provides increasing representational capacity with radius and naturally accommodates nested hierarchies. We propose **HyperForecast**, the first long-term time series forecasting (LTSF) framework operating in hyperbolic space. HyperForecast embeds decomposed temporal segments on the Poincaré ball where radial distance captures this nested structure, from low-frequency components near the origin to high-frequency components near the boundary. We model geodesic velocity through two manifold-dynamics variants: (a) autoregressive forecasting via Poincaré residual dynamics, and (b) non-autoregressive forecasting that predicts all future segments in parallel. HyperForecast consistently outperforms existing and state-of-the-art baselines on 9 of 11 long-term forecasting benchmarks, achieving up to 58% lower mean squared error at long horizons. Further analysis reveals that the learned radial embeddings recover the frequency hierarchy and that hyperbolic geometry contributes substantially to forecasting performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Geometric Representational Learning, Time Series Forecasting

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 4 reviews, above 85% of the others.
Reviewer_7GQpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HyperForecast, a long-horizon time series forecasting framework based on hyperbolic geometry. It first decomposes time series into trends, coarse-grained periodic patterns, fine-grained periodic patterns, and residuals using learnable cascaded convolutions. These components are then mapped into Poincaré hyperbolic space, where future values are predicted by estimating geodesic velocities. Hierarchical and temporal consistency losses are introduced to preserve the structural relationships among representations. The framework supports both autoregressive (AR) and non-autoregressive (NAR) forecasting, achieving competitive results on 9 out of 11 benchmark datasets. Ablation studies further demonstrate the benefits of hyperbolic geometry. The key idea is to leverage hyperbolic space to capture multiscale hierarchical relationships and thereby improve long-horizon time series forecasting.

Critical strengths
  1. The method is based on hyperbolic geometry, combining multiscale decomposition with manifold dynamics to explicitly model hierarchical relationships among components of different frequencies in hyperbolic space, which is interpretable.
  2. Develop AR/NAR forecasting mechanisms, using hierarchical and temporal consistency losses to preserve representational structure and improve forecasting stability.
  3. The results are promising. The model achieves SOTA in 9 out of 11 benchmarks.
Critical weaknesses
  1. Different time series may vary greatly. The assumption that components of different frequencies from a hierarchical structure may not hold for all time series. So the general advantages of leveraging hyperbolic geometry thoery requires further demonstration.
  2. The NAR approach relies on a global velocity estimate to predict future values, which may limit its ability to model complex nonlinear dynamics and regime shifts.
  3. The experiments focus primarily on conventional long-horizon forecasting benchmarks, providing insufficient evidence of generalization to more complex and highly non-stationary scenarios. So I suggest authors to include results for more complex, non-stationary datasets to address concerns (1) and (3).
Questions for authors

N/A

5 comments
AuthorsOpenReview ↗

Response to Reviewer 7GQp (Part 1/3)

W1. Different time series may vary greatly. The assumption that components of different frequencies from a hierarchical structure may not hold for all time series. So the general advantages of leveraging hyperbolic geometry thoery requires further demonstration.

Answer: Thank you for the point. We agree that the strength of the frequency hierarchy differs between series and cannot be assumed for every one, so we measured it on the datasets directly, before any model is trained. For each dataset, we cut 24-step segments from the four components of the series (fixed moving-average cascade with the day, week, and two-week kernels), pooled them into one point cloud, and computed its relative Gromov-δ (Euclidean distances, scaled by their 99th percentile). As a reference, the table below also gives the δ of Gaussian noise with the same mean and covariance as the data, which has the same spread but no hierarchy. A lower δ than this reference means the components are more tree-like than their spread alone would produce:

DatasetSamplingδ (decomposed data)δ (Gaussian reference)
Weather10 min0.00470.0298
AQShunyi1 h0.01360.0284
AQWan1 h0.01290.0290
CzeLan30 min0.00930.0367
ZafNoo30 min0.01530.0402
Exchange1 day0.00180.0025
METR-LA5 min0.00530.0208
PM2.53 h0.01590.0296
Solar10 min0.00370.0180
Temp3 h0.01180.0155
Wind15 min0.01130.0238

On all eleven datasets, the pooled components are more tree-like than Gaussian noise with the same spread, so the structure is present in the data before any model is applied. The trained embeddings in Table 2 have δ between 0.002 and 0.023, a similar range. So the Poincaré model preserves this structure; it does not create it.

The benefit also comes from the exponential volume growth of the Poincaré ball, which leaves more room at larger radii for the higher-frequency components, whether or not the data is strictly from a tree. The volume of the Poincaré ball grows exponentially with the radius, because its metric factor λx=21−c∥x∥2\lambda_{\mathbf{x}} = \tfrac{2}{1 - c\lVert\mathbf{x}\rVert^2} grows toward the boundary, so the higher-frequency components at larger radii have more room than in a Euclidean space of the same dimension, whose volume grows only polynomially. The forecast uses the same geometry: the geodesic step exp⁡zN(t⋅vˉ)\exp_{\mathbf{z}_N}(t\cdot\bar{\mathbf{v}}) covers the same distance in the ball's own geometry at every step, so in the ball's coordinates the step shrinks at larger radii and the state stays inside the ball for every t, whereas a Euclidean step zN+t⋅vˉ\mathbf{z}_N + t\cdot\bar{\mathbf{v}} has the same length at every radius. In our revised manuscript, Section 5 (lines 428-430) now points to the δ of the decomposed data in Appendix J, and Section 4.5 (lines 315-319) adds the geometric argument.

AuthorsOpenReview ↗

Response to Reviewer 7GQp (Part 2/3)

W2. The NAR approach relies on a global velocity estimate to predict future values, which may limit its ability to model complex nonlinear dynamics and regime shifts.

Answer: Thank you for the point. NAR estimates one tangent velocity per component (four velocities, one per time scale) as an exponentially weighted average of the displacements in the lookback, so recent displacements dominate, and extrapolates it along a geodesic of the ball. Even this single-velocity path is not a uniform straight line in the ball's coordinates, as its steps are shorter near the boundary inside the ball. The per-horizon gate, the Poincaré FC layer, and the MLP decoder then bend the forecast further. The AR variant is the local alternative, which re-estimates the velocity after every predicted segment. If a single velocity limited the model, AR should gain on NAR as the horizon grows. As per Table 24, NAR is better on 29 of the 44 dataset-horizon pairs, and AR on 13 (2 ties), and the advantage of NAR grows with the horizon, so re-estimating the velocity step by step does not help. A likely reason is that AR feeds its own predictions back, so errors compound along the rollout, while NAR forecasts every segment from the observed lookback.

We agree that a single velocity per component cannot anticipate a regime change that begins after the forecast origin, which no input-only forecaster can, and that an abrupt level shift inside the lookback is the harder case for it. Two parts of the model control how it reacts to a shift. RevIN normalizes each input window by its own mean and standard deviation, so a level shift between the training and test periods is removed before the embedding. Within the lookback, the exponential weights use a fixed decay of 0.9 for all four components, and a smaller decay, or a shorter part of the lookback, makes the velocity follow the displacements after a shift sooner, at the cost of more noise. A regime-aware velocity is a natural extension that we leave for future work (line 523 of revised manuscript), for example by learning the decay per component, so the trend keeps a long memory and the residual a short one, or by combining a short-memory and a long-memory velocity with a gate learned from the lookback.

AuthorsOpenReview ↗

Response to Reviewer 7GQp (Part 3/3)

W3. The experiments focus primarily on conventional long-horizon forecasting benchmarks, providing insufficient evidence of generalization to more complex and highly non-stationary scenarios. So I suggest authors to include results for more complex, non-stationary datasets to address concerns (1) and (3).

Answer: Thank you for the suggestion. We address this with three analyses.

(i) Non-stationarity of the benchmarks. Task: long-term forecasting on the 11 benchmarks. For each dataset, the table below gives the share of channels that KPSS and ADF tests flag as non-stationary (1.00 = all channels), and the shift of the test-period mean from the training mean in training standard deviations (larger means more drift between training and test):

DatasetKPSS non-stationaryADF non-stationaryTest mean shift (train SD)
Exchange1.000.881.39
CzeLan0.360.180.73
Weather0.760.000.42
PM2.50.830.000.31
Wind1.000.140.27
Temp1.000.070.25
METR-LA0.470.000.21
Solar1.000.000.20
ZafNoo0.910.090.17
AQWan0.820.000.09
AQShunyi0.450.000.08

The benchmarks differ widely in how much they drift between training and test. HyperForecast stays ahead of the best baselines on most of them, including the most drifting ones, so its gains are not limited to stationary data.

(ii) Missing values. Task: forecasting from corrupted inputs. At test time only, random blocks of 24 consecutive steps are removed from each channel until 10% (or 30%) of the steps are missing and filled with the last observed value; all models are trained on clean data and evaluated under the same corruption (H = 96, L = 720, MSE). The table below gives the clean MSE and the MSE under corruption with its relative increase.

DatasetModelOriginal10% missing30% missing
AQShunyiHyperForecast0.6100.613 (+0.6%)0.621 (+1.9%)
TimeBase0.6290.633 (+0.7%)0.643 (+2.2%)
PhaseFormer0.6440.648 (+0.7%)0.660 (+2.5%)
AQWanHyperForecast0.6990.706 (+1.0%)0.708 (+1.3%)
TimeBase0.7260.730 (+0.5%)0.739 (+1.8%)
PhaseFormer0.7430.748 (+0.7%)0.762 (+2.6%)
WeatherHyperForecast0.1460.147 (+0.9%)0.150 (+3.3%)
TimeBase0.1450.146 (+1.1%)0.150 (+4.0%)
PhaseFormer0.1470.149 (+1.2%)0.154 (+4.4%)

Missing values degrade HyperForecast no more than the baselines. Every model loses at most about 1% of its MSE at 10% missing and a few percent at 30%, and at 30% the loss is smallest for HyperForecast on all three datasets. Its MSE also stays the lowest, or on par with the best baseline.

(iii) More complex benchmarks. Thank you for the suggestion. We evaluated HyperForecast on the Web/CloudOps subsets of GIFT-Eval (Aksu et al., 2024), whose workload traces are highly non-stationary, with frequent spikes and abrupt shifts in intensity. We follow the GIFT-Eval protocol, with its short-term horizon of 48 steps, its train, validation, and test split, and the MSE of the forecast on the original scale. HyperForecast is trained on each dataset with the learning rate chosen on the GIFT-Eval validation window (similar evaluation protocol as PhaseFormer’s experiments on GIFT-Eval), while the foundation models Moirai-2 (Liu et al., 2025), Moirai-base (Woo et al., 2024), and Chronos-base (Ansari et al., 2024) are evaluated zero-shot. The values of the other models are those reported by PhaseFormer.

DatasetHyperForecastPhaseFormerMoirai-2Moirai-baseChronos-base
bizitobs_l2c/5T16.2239.4320.6222.1024.10
bizitobs_l2c/H65.0799.9667.85272.00243.00
bitbrains_rnd/5T1,474,7431,606,0341,660,532180,000,0001,800,000
bitbrains_rnd/H1,780,9931,970,3302,235,5734,020,0002,210,000
bitbrains_fast_storage/5T1,725,5461,887,3982,142,12525,400,000,0001,920,000
bitbrains_fast_storage/H2,709,6912,585,1802,942,651604,000,0002,620,000

HyperForecast has the lowest MSE on five of the six subsets, including the two bizitobs_l2c subsets where Moirai-2 is the best of the other models, and on bitbrains_fast_storage/H it is close behind PhaseFormer and Chronos-base. Its gains therefore carry over to these volatile, non-stationary workloads as well. In our revised manuscript, we added the three analyses as Appendix L (Tables 21 to 23), referenced in Section 5 (lines 377-379).

References

Aksu, Taha, et al. "Gift-eval: A benchmark for general time series forecasting model evaluation." arXiv preprint arXiv:2410.10393 (2024).

Liu, Chenghao, et al. "Moirai 2.0: When less is more for time series forecasting." arXiv preprint arXiv:2511.11698 (2025).

Woo, Gerald et al. "Unified Training of Universal Time Series Forecasting Transformers." Proceedings of the 41st International Conference on Machine Learning (2024).

Ansari, Abdul Fatir, et al. "Chronos: Learning the Language of Time Series." Transactions on Machine Learning Research (2024)

Reviewer_P796OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces HyperForecast, a long-term time-series forecasting architecture motivated by the frequency hierarchy induced by multiscale decomposition and hyperbolic geometry's ability to represent such hierarchies. Rather than treating decomposed frequency components as geometrically equivalent in Euclidean space, HyperForecast uses learnable decomposition and embeds the resulting components in hyperbolic space, where frequency ordering is encoded through radial distance. Forecasting is performed either autoregressively through iterative geodesic velocity estimation and Poincaré residual updates or non-autoregressively by extrapolating a single velocity across all future segments in parallel. Experiments on several long-term forecasting benchmarks show that this design improves forecasting accuracy, achieving the lowest reported average MSE on 7 datasets with its non-autoregressive variant, and on 9 datasets when considering both variants.

Critical strengths

1. The formulation of forecasting through geodesic velocity estimation and manifold dynamics is a distinctive architectural contribution.

Rather than using hyperbolic geometry solely for representation learning, the paper models temporal evolution directly on the manifold. The autoregressive and non-autoregressive variants demonstrate the flexibility of this formulation, with the latter offering substantially faster inference and strong forecasting performance.

2. The paper provides extensive experiments and analyses beyond standard forecasting comparisons.

The evaluation covers 11 datasets, multiple forecasting horizons, geometric diagnostics, architectural ablations, efficiency measurements, and multi-seed experiments. These analyses offer useful insight into the proposed model's behavior, while the publicly available code improves transparency and facilitates further investigation.

Critical weaknesses

1. The proposed frequency hierarchy does not sufficiently justify hyperbolic geometry.

The cascaded decomposition in Eq. 4 guarantees additive reconstruction, but not the branching structure motivating hyperbolic representations. Moreover, Eqs. 1 and 18 imply dc(0,exp⁡0c(γh))=2∣γ∣∥h∥2d_c(0,\exp_0^c(\gamma h)) = 2|\gamma|\|h\|_2, making the radial constraint in Eq. 10 equivalent to ranking scaled Euclidean norms. The coherence term in Eq. 11 further favors neighboring components along a common geodesic rather than a branching hierarchy. The geometric diagnostics provide limited additional evidence for the proposed mechanism: Gromov hyperbolicity is measured on already-hyperbolic embeddings, while radial ordering is imperfect on Temp and AQShunyi despite strong forecasting performance. These observations weaken the mathematical and empirical justification for the proposed inductive bias.

2. The released Euclidean ablation implementation does not adequately isolate hyperbolic geometry.

Although Figure 5a reports improvements over a Euclidean counterpart, the released Euclidean NAR implementation introduces substantial differences beyond manifold geometry. It disables both auxiliary losses and instantiates a freshly initialized, unregistered reconstruction layer during each forward pass, whereas the hyperbolic model uses a persistent, trainable multilayer reconstruction head. These discrepancies confound the intended geometry comparison if this implementation was used to generate Figure 5a. The released Euclidean NAR constructor also fails due to a missing required argument, raising additional reproducibility concerns.

Questions for authors
  1. Given that the radial hierarchy loss reduces to Euclidean norm ranking, what aspect of the proposed frequency hierarchy specifically benefits from negative curvature rather than an ordered Euclidean representation? Could the authors clarify whether the frequency hierarchy exhibits branching structure beyond the ordering of four components?

  2. Was the released Euclidean NAR implementation used to generate Figure 5a? If not, could the authors identify the implementation used? If so, how do the authors account for the unregistered reconstruction layer and differences in regularization and architectural capacity?

  3. Equation 8 produces a velocity in the tangent space at zN\mathrm{z}_N, but Section 4.5 subsequently passes it through the Poincaré FC layer in Eq. 2, which operates on manifold points. How is this velocity mapped through the layer and converted into the tangent vector v used in Eq. 9?

5 comments
AuthorsOpenReview ↗

Response to Reviewer P796 (Part 1/4)

W1. The proposed frequency hierarchy does not sufficiently justify hyperbolic geometry. The cascaded decomposition in Eq. 4 guarantees additive reconstruction, but not the branching structure motivating hyperbolic representations. Moreover, Eqs. 1 and 18 imply dc(0,exp⁡0c(γh))=2∣γ∣∥h∥2d_c(0,\exp_0^c(\gamma h)) = 2|\gamma|\|h\|_2, making the radial constraint in Eq. 10 equivalent to ranking scaled Euclidean norms. The coherence term in Eq. 11 further favors neighboring components along a common geodesic rather than a branching hierarchy. The geometric diagnostics provide limited additional evidence for the proposed mechanism: Gromov hyperbolicity is measured on already-hyperbolic embeddings, while radial ordering is imperfect on Temp and AQShunyi despite strong forecasting performance. These observations weaken the mathematical and empirical justification for the proposed inductive bias.

Answer: Thank you for the careful analysis. Your algebra is correct. With exp⁡o\exp_{\mathbf{o}} as the lift, d(o,exp⁡o(γh))=2∣γ∣,∥h∥d(\mathbf{o}, \exp_{\mathbf{o}}(\gamma\mathbf{h})) = 2|\gamma|\\,\lVert\mathbf{h}\rVert, so the radial term of the hierarchy loss is a ranking loss on scaled Euclidean norms of the pre-projection features. The coherence term is different in kind, since it uses pairwise hyperbolic distances between components, but it pulls consecutive components together along a path rather than apart into branches, so we agree it does not encode branching either. We updated Section 4.6 (lines 337-339) in our revised manuscript to mention this explicitly. The contribution of the hierarchy regularizer can be noted in the auxiliary loss ablations (Appendix E, Table 13). For each dataset, we cut 24-step segments from the four components of the series (fixed moving-average cascade with the day, week, and two-week kernels), pooled them into one point cloud, and computed its relative Gromov-δ (Euclidean distances, scaled by their 99th percentile). As a reference, the table below also gives the δ of Gaussian noise with the same mean and covariance as the data, which has the same spread but no hierarchy. A lower δ than this reference means the components are more tree-like than their spread alone would produce:

DatasetSamplingδ (decomposed data)δ (Gaussian reference)
Weather10 min0.00470.0298
AQShunyi1 h0.01360.0284
AQWan1 h0.01290.0290
CzeLan30 min0.00930.0367
ZafNoo30 min0.01530.0402
Exchange1 day0.00180.0025
METR-LA5 min0.00530.0208
PM2.53 h0.01590.0296
Solar10 min0.00370.0180
Temp3 h0.01180.0155
Wind15 min0.01130.0238

On all eleven datasets, the pooled components are more tree-like than Gaussian noise with the same spread, so the structure is present in the data before any model is applied. The trained embeddings in Table 2 have δ between 0.002 and 0.023, a similar range. So the Poincaré model preserves this structure; it does not create it. For a comparison with an unconstrained representation, we also measured δ on the embeddings of the matched Euclidean model (same test segments, H = 96, each model in its own metric: Euclidean distance for the Euclidean model, geodesic distance for the Poincaré model as in Table 2):

DatasetEuclidean-model embeddingsPoincaré-model embeddings
AQShunyi0.0220.012
AQWan0.0200.0085
Weather0.0180.018
Temp0.0310.022

We added Appendix J in our revised manuscript to report this δ analysis. The Euclidean embeddings are also low, however, Poincaré embeddings seem to have even lower δ than the Euclidean counterpart and generally lower than the Gaussian noise level. Taken together, the δ of raw data and the embeddings indicate that the tree-like structure comes from the data and not from our hierarchical constraints, and that the Poincaré embeddings are more tree-like than the Euclidean embeddings. The radial order is indeed imperfect on Temp and AQShunyi (Table 2), and both still rank first by MSE (Table 1), so the accuracy does not hinge on a perfect radial order. The benefit also comes from the exponential volume growth of the Poincaré ball, which leaves more room at larger radii for the higher-frequency components, as explained in our response to Q1.

AuthorsOpenReview ↗

Response to Reviewer P796 (Part 2/4)

W2. The released Euclidean ablation implementation does not adequately isolate hyperbolic geometry.Although Figure 5a reports improvements over a Euclidean counterpart, the released Euclidean NAR implementation introduces substantial differences beyond manifold geometry. It disables both auxiliary losses and instantiates a freshly initialized, unregistered reconstruction layer during each forward pass, whereas the hyperbolic model uses a persistent, trainable multilayer reconstruction head. These discrepancies confound the intended geometry comparison if this implementation was used to generate Figure 5a. The released Euclidean NAR constructor also fails due to a missing required argument, raising additional reproducibility concerns. (Q2)

Answer: Thank you for the question. We have answered this in our response to Q2


Q1. Given that the radial hierarchy loss reduces to Euclidean norm ranking, what aspect of the proposed frequency hierarchy specifically benefits from negative curvature rather than an ordered Euclidean representation? Could the authors clarify whether the frequency hierarchy exhibits branching structure beyond the ordering of four components?

Answer: Thank you for the question. As explained in our response to W1, the radial loss is a norm-ranking regularizer. The benefit comes from the exponential volume growth of the Poincaré ball, which leaves more room at larger radii for the higher-frequency components, whether or not the data is strictly from a tree. The volume of the Poincaré ball grows exponentially with the radius, because its metric factor λx=21−c∥x∥2\lambda_{\mathbf{x}} = \tfrac{2}{1 - c\lVert\mathbf{x}\rVert^2} grows toward the boundary, so the higher-frequency components at larger radii have more room than in a Euclidean space of the same dimension, whose volume grows only polynomially. The forecast uses the same geometry: the geodesic step exp⁡zN(t⋅vˉ)\exp_{\mathbf{z}_N}(t\cdot\bar{\mathbf{v}}) covers the same distance in the ball's own geometry at every step, so in the ball's coordinates the step shrinks at larger radii and the state stays inside the ball for every t, whereas a Euclidean step zN+t⋅vˉ\mathbf{z}_N + t\cdot\bar{\mathbf{v}} has the same length at every radius. As for branching, the four components of one time segment and one channel form a chain. Across segments and channels, these chains start in the region near the origin, where the trend components cluster, and fan out toward the boundary, so the pooled embedding branches at that shared region. In our revised manuscript, we added this argument to Section 4.5 (lines 315-319), and Section 5 (lines 449-451) now describes how the chains fan out from the origin in Figure 4.

AuthorsOpenReview ↗

Response to Reviewer P796 (Part 3/4)

Q2. Was the released Euclidean NAR implementation used to generate Figure 5a? If not, could the authors identify the implementation used? If so, how do the authors account for the unregistered reconstruction layer and differences in regularization and architectural capacity? (W2)

Answer: Thank you for reviewing our code so carefully. Figure 5a was produced with the Euclidean NAR implementation in the repository at the time of submission. We agree with you that the reconstruction layer should not be unregistered for the Euclidean version of our model, and that the auxiliary losses were switched off in that version. We have now updated the repository where the reconstruction layer is registered, the Euclidean versions of the hierarchy and temporal-consistency losses are used, and the constructor error is fixed. We are focused on isolating the performance difference between Euclidean geometry and Poincaré geometry. The MLP has only one hidden layer, and it is used to deal with the nonlinearity that comes from hyperbolic space and to lift the predictions back to the global Euclidean space from the local tangent space. The Euclidean ablation does not require this, since its predictions are already in the global Euclidean space. The architecture capacity of HyperForecast is therefore not the same as that of its Euclidean ablation, because the operations are not one-to-one comparable. A tangent space is a flat linear approximation of a curved manifold anchored strictly at a single basepoint. While tangent space is isomorphic to an Rn\mathbb{R}^n vector space locally, it only reflects the local geometry around the basepoint. So the difference in layers and parameters between HyperForecast and its Euclidean ablation comes from what the geometry requires, not from an arbitrary increase in capacity. The two comparisons below check that the gain does not come from this difference.

We know that there is a difference in model performance when we change from the Euclidean FC layer to the Poincaré FC layer; both had the same MLP. Here are the differences in performance on the AR model with Euclidean FC Layer vs Poincaré FC average over the four horizons (96, 192, 336, 720):

DatasetEuclidean FC (MSE)Poincaré FC (MSE)
Weather0.2270.221
Temp0.1760.145
AQWan0.7680.762
AQShunyi0.6750.665

Here are the results of the corrected Euclidean NAR model, with the registered reconstruction layer, and of the Poincaré model with its hidden dimension reduced from 256 to 64 and its ReLU activation removed, so that the decoder has no nonlinear activation, to check that the gain does not come from the size or the nonlinearity of the decoder (NAR, L = 720, test MSE):

DatasetModel96192336720Average
AQWanEuclidean0.7100.7450.7760.8280.765
Poincaré0.6920.7340.7600.8210.752
AQShunyiEuclidean0.6200.6520.6760.7220.667
Poincaré0.6060.6430.6720.7160.659
WeatherEuclidean0.1590.2090.2540.3180.235
Poincaré0.1450.1890.2350.3050.218
TempEuclidean0.1720.1660.1750.1910.176
Poincaré0.1330.1310.1270.1280.130

With the smaller decoder, the Poincaré model still has a lower MSE than the corrected Euclidean model at every horizon on all four datasets. In our revised manuscript, we replaced Figure 5a with this comparison and updated its description in Section 5 (lines 484-490), and Appendix E.4 includes the AR comparison with a Euclidean vs. Poincaré FC layer.

Reviewer_nPX4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HyperForecast for long-term time series forecasting. The method first decomposes the input series into trend, coarse seasonal, fine seasonal, and residual components using learnable convolutions. Each component is then projected into a hyperbolic latent space through a linear encoder and an exponential map. The method estimates future latent states by computing an exponentially weighted average of historical latent differences and extrapolating the estimated velocity along geodesics. The predicted components are fused through standard Möbius addition then with a MLP decoder to generate the final forecasts. The authors evaluate the proposed framework on 11 standard long-term forecasting benchmarks and provide ablation studies, geometric analysis, and efficiency comparisons.

Critical strengths

1 The paper is generally organized.

2 The authors combine the familiar decomposition, weighted exponential average methods with the new mathematical framework.

Critical weaknesses

Weakness:

1 The methodological novelty and theoretical contribution appear limited to me. The proposed framework largely follows the familiar decomposition-based forecasting pipeline, as explored in methods such as DLinear and TimeMixer, where time series are decomposed into multiple components, processed separately, and subsequently recombined for forecasting. The main distinction appears to be the reformulation of these operations in a hyperbolic latent space using established geometric tools, including Poincaré embeddings, logarithmic/exponential maps, parallel transport, and Möbius addition. Despite the extensive geometric formulation, the paper provides limited theoretical insight into why hyperbolic geometry is appropriate for this forecasting task.

2 The authors assume that frequency ordering naturally corresponds to a tree-like hierarchy, although an ordering of frequency components does not necessarily imply a hierarchical structure. The paper does not sufficiently justify why such geodesic dynamics should better characterize temporal evolution than simpler Euclidean alternatives.

3 The empirical improvements over strong baselines are relatively marginal on several datasets, which makes the claimed SOTA performance less convincing. For example, HyperForecast NAR achieves an average MSE of 0.218 compared with 0.219 for TimeBase on Weather, and 0.930 compared with 0.940 on Wind. Besides, the context window length is fixed to 720, which is a different settings with these proposed methods. It raises concerns if the proposed methods are statistically significant then these baselines.

4 The authors compare HyperForecast with transformers baselines, however, the evaluation lacks comparisons with more directly related multi-resolution frequency-decomposition methods, such as WaveletMixer and WPMixer. Since the central motivation of HyperForecast is to model the hierarchical structure induced by frequency decomposition, comparisons with these methods are particularly relevant. Without such comparisons, it remains unclear whether the proposed hyperbolic representation provides meaningful advantages over existing frequency-decomposition approaches.

5 Since the proposed velocity estimation relies on aggregating historical latent differences, the effectiveness of the method may depend heavily on the available historical context. What is the method performance with different context length like 336 or 96.

Questions for authors

1 The paper establishes a complex idea in geometry operations to decomposition, weighted aggregation, and recombination. Do you have any theoretical evidence to show your method is better in such a hyperbolic space.

2 Why does frequency ordering necessarily imply a tree-like hierarchy? Besides, what is the motivation for assuming that future latent states follow geodesic trajectories, rather than simpler Euclidean extrapolation? Such hierarchy seems to be datasets specific instead of universal.

3 Are the marginal improvements over strong baselines statistically significant? Why are the multi-seed results only reported for HyperForecast without corresponding variances for the strongest baseline.

4 How does HyperForecast compare with other frequency-decomposition methods like WaveletMixer and WPMixer? Since these methods also model multi-scale frequency structures, it is unclear whether the hyperbolic formulation provides substantial additional benefits.

5 The main experiments fix the lookback length to 720, while the sensitivity analysis only covers two datasets. How does the method perform under shorter context windows like 96, 192, and 336 compared with strong baselines?

6 The hyperbolic assumption seems to place strong requirements on the underlying temporal structure. How does the method perform on non-stationary data or time series with missing values? Some experiments on broader benchmarks like GIFT-EVAL could further demonstrate its generalizability.

[1] Zhang, Zichi, et al. “WaveletMixer: a multi-resolution wavelets based MLP-mixer for multivariate long-term time series forecasting.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 21. 2025. [2] Murad, Md Mahmuddun Nabi, Mehmet Aktukmak, and Yasin Yilmaz. “Wpmixer: Efficient multi-resolution mixing for long-term time series forecasting.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 18. 2025.

7 comments
AuthorsOpenReview ↗

Response to Reviewer nPX4 (Part 1/5)

W1. The methodological novelty and theoretical contribution appear limited to me. The proposed framework largely follows the familiar decomposition-based forecasting pipeline, as explored in methods such as DLinear and TimeMixer, where time series are decomposed into multiple components, processed separately, and subsequently recombined for forecasting. The main distinction appears to be the reformulation of these operations in a hyperbolic latent space using established geometric tools, including Poincaré embeddings, logarithmic/exponential maps, parallel transport, and Möbius addition. Despite the extensive geometric formulation, the paper provides limited theoretical insight into why hyperbolic geometry is appropriate for this forecasting task. (Q1)

Answer: Thank you for the comment. Decomposing a series, processing the parts separately, and recombining them is an established design (DLinear, TimeMixer), and we do not claim it as new. What we change is how each step is done. DLinear and TimeMixer split the series with a fixed moving average (kernel size 25) into two parts, a trend and what remains of the series, and add the processed parts back. TimeMixer repeats this on several average-pooled scales. HyperForecast differs in three ways: (i) the decomposition is a learned cascade of three filters, each acting on what the previous one leaves, with kernels set to a day, a week and two weeks and weights trained jointly with the forecaster, and it yields four components (trend, coarse seasonal, fine seasonal, residual); (ii) the four components are embedded in a Poincaré ball and fused by Möbius addition, which depends on the curvature and on the order of its terms, instead of a plain sum; (iii) the forecast is a geodesic extrapolation of a velocity estimated from the lookback, which stays inside the ball for every horizon, instead of a direct linear or MLP map from input to output. The paper tests the parts that carry the contribution by removing the decomposition (Table 3) and removing the hierarchy loss (Table 13). We added a paragraph to Section 2 of our revised manuscript (lines 117-123) that positions HyperForecast against these decomposition methods along these differences.

Our motivation is geometric. Hyperbolic space represents hierarchies in few dimensions because its volume grows exponentially with the radius. For example, a 5-dimensional Poincaré embedding of WordNet is more accurate than a 200-dimensional Euclidean one (Nickel & Kiela, 2017), and two dimensions suffice for a near-exact reconstruction (Sala et al., 2018). The cascaded decomposition gives a series a frequency hierarchy of its own, since the components are ordered by frequency and each is extracted from what the slower ones leave. The higher-frequency components settle at larger radii of the Poincaré disk, where the larger volume leaves them more room than a Euclidean space of the same dimension would. Section 5 measures the frequency order of the decomposed components and the radial order of their embeddings (Table 2 and Figure 3).

References

Nickel, Maximilian, and Douwe Kiela. "Poincaré embeddings for learning hierarchical representations." Advances in Neural Information Processing Systems 30 (2017).

Sala, Frederic, et al. "Representation tradeoffs for hyperbolic embeddings." Proceedings of the 35th International Conference on Machine Learning (2018).

AuthorsOpenReview ↗

Response to Reviewer nPX4 (Part 2/5)

W2. The authors assume that frequency ordering naturally corresponds to a tree-like hierarchy, although an ordering of frequency components does not necessarily imply a hierarchical structure. The paper does not sufficiently justify why such geodesic dynamics should better characterize temporal evolution than simpler Euclidean alternatives. (Q2)

Answer: Thank you for the point. We agree that an ordering of four frequency bands is not by itself a tree, and we do not use it as a proof. What we use is the nesting that the cascade creates: T=Conv1Dtrend(X)\mathbf{T} = \mathrm{Conv1D}_{\mathrm{trend}}(\mathbf{X}), Sc=Conv1Dcoarse(X−T)\mathbf{S}_c = \mathrm{Conv1D}_{\mathrm{coarse}}(\mathbf{X} - \mathbf{T}), Sf=Conv1Dfine(X−T−Sc)\mathbf{S}_f = \mathrm{Conv1D}_{\mathrm{fine}}(\mathbf{X} - \mathbf{T} - \mathbf{S}_c), R=X−T−Sc−Sf\mathbf{R} = \mathbf{X} - \mathbf{T} - \mathbf{S}_c - \mathbf{S}_f, so each finer component is computed from what the coarser ones left over. For one segment of one channel, this gives a chain T,Sc,Sf,R\mathbf{T}, \mathbf{S}_c, \mathbf{S}_f, \mathbf{R}. We do not impose tree edges anywhere in the model, and our claim is about a nested multiscale structure. The motivation is that components of increasing frequency need increasing room, which the volume growth of the ball provides, and we measure this instead of assuming it. The paper measures the nesting independently of the model geometry: the spectral centroids increase from the trend to the coarse, fine, and residual components (Figure 3, and additional datasets in Appendix D). Whether the data show this structure can also be measured without any model, so we measured it on all eleven datasets.

For each dataset, we cut 24-step segments from the four components of the series (fixed moving-average cascade with the day, week, and two-week kernels), pooled them into one point cloud, and computed its relative Gromov-δ (Euclidean distances, scaled by their 99th percentile). As a reference, the table below also gives the δ of Gaussian noise with the same mean and covariance as the data, which has the same spread but no hierarchy. A lower δ than this reference means the components are more tree-like than their spread alone would produce:

DatasetSamplingδ (decomposed data)δ (Gaussian reference)
Weather10 min0.00470.0298
AQShunyi1 h0.01360.0284
AQWan1 h0.01290.0290
CzeLan30 min0.00930.0367
ZafNoo30 min0.01530.0402
Exchange1 day0.00180.0025
METR-LA5 min0.00530.0208
PM2.53 h0.01590.0296
Solar10 min0.00370.0180
Temp3 h0.01180.0155
Wind15 min0.01130.0238

On all eleven datasets, the pooled components are more tree-like than Gaussian noise with the same spread, so the structure is not specific to a few datasets.

Our theoretical motivation builds upon the volume of the Poincaré ball, which grows exponentially with the radius, because its metric factor λx=21−c∥x∥2\lambda_{\mathbf{x}} = \tfrac{2}{1 - c\lVert\mathbf{x}\rVert^2} grows toward the boundary. The geodesic step exp⁡zN(t⋅vˉ)\exp_{\mathbf{z}_N}(t\cdot\bar{\mathbf{v}}) is the constant-velocity motion in this metric. The state covers the same distance in the ball’s own geometry at every step, so in the ball’s coordinates the step shrinks at larger radii, where there is more room, and the higher-frequency components lie, and the state stays inside the ball for every t. For a short step, it is close to zN+t⋅vˉ\mathbf{z}_N + t\cdot\bar{\mathbf{v}}. The velocity dynamics is thus matched to the exponential volume of the ball. In Euclidean space, the volume grows only polynomially, and the step zN+t⋅vˉ\mathbf{z}_N + t\cdot\bar{\mathbf{v}} has the same length at every radius, so nothing adapts it to where the state is. We added this argument to Section 4.5 of our revised manuscript (lines 315-319) and the δ of the decomposed data in Appendix J.

AuthorsOpenReview ↗

Response to Reviewer nPX4 (Part 3/5)

W3. The empirical improvements over strong baselines are relatively marginal on several datasets, which makes the claimed SOTA performance less convincing. For example, HyperForecast NAR achieves an average MSE of 0.218 compared with 0.219 for TimeBase on Weather, and 0.930 compared with 0.940 on Wind. Besides, the context window length is fixed to 720, which is a different settings with these proposed methods. It raises concerns if the proposed methods are statistically significant then these baselines. (Q3)

Answer: Thank you for the point. We now report 5 seeds for TimeBase, PatchTST, and PhaseFormer as well, each run at L = 720 with the hyperparameters of its original repository. Task: long-term forecasting on the five datasets with 5-seed HyperForecast results. Metric: MSE averaged over the four horizons, reported as mean ± std over the 5 seeds (std computed per horizon, then averaged over the four horizons):

DatasetHyperForecastTimeBasePatchTSTPhaseFormer
Weather0.222 ± 0.0050.220 ± 0.0020.234 ± 0.0040.227 ± 0.001
Wind0.934 ± 0.0070.955 ± 0.0031.016 ± 0.0130.997 ± 0.011
AQShunyi0.663 ± 0.0040.675 ± 0.0030.698 ± 0.0050.710 ± 0.005
AQWan0.756 ± 0.0050.781 ± 0.0030.809 ± 0.0080.815 ± 0.009
ZafNoo0.491 ± 0.0070.497 ± 0.0030.576 ± 0.0090.522 ± 0.006

We test significance per dataset and horizon with a two-sided Welch t-test over the 5 seeds (p < 0.05). HyperForecast is significantly better than PatchTST and PhaseFormer at most dataset-horizon pairs and than TimeBase at about half of them. The concern about the fixed lookback of 720 is addressed in our response to W5, where all four methods are compared at L = 96, 192, 336, and 720. We added the 5-seed baseline results and the Welch test to Appendix F of our revised manuscript (Table 17), referenced in Section 5 (line 377).


W4. The authors compare HyperForecast with transformers baselines, however, the evaluation lacks comparisons with more directly related multi-resolution frequency-decomposition methods, such as WaveletMixer and WPMixer. Since the central motivation of HyperForecast is to model the hierarchical structure induced by frequency decomposition, comparisons with these methods are particularly relevant. Without such comparisons, it remains unclear whether the proposed hyperbolic representation provides meaningful advantages over existing frequency-decomposition approaches. (Q4)

Answer: Thank you for the suggestion. WPMixer (Murad et al., 2025) has public code, so we ran it with the per-horizon configuration of its repository, at its own lookback value in its original repository (L=512) and at our protocol (L = 720), on a few datasets. WaveletMixer has no public implementation that we could find, so we could not include it. Task: long-term forecasting; metric: MSE averaged over the four horizons:

DatasetHyperForecastWPMixer (L = 512)WPMixer (L = 720)
Weather0.2180.2220.220
Wind0.9300.9660.965
AQShunyi0.6570.6740.678
AQWan0.7510.7740.781
ZafNoo0.4850.5140.504
Temp0.1190.208n/a
PM2.50.4100.428n/a

WPMixer at L = 720 and H = 720 ran out of memory on Temp and PM2.5 (184 channels). HyperForecast has the lower four-horizon MSE than WPMixer on every dataset. We added the WPMixer comparison to our revised manuscript as Appendix K (Table 20).

References

Murad, Md Mahmuddun Nabi, Mehmet Aktukmak, and Yasin Yilmaz. "WPMixer: Efficient multi-resolution mixing for long-term time series forecasting." Proceedings of the AAAI Conference on Artificial Intelligence 39 (2025).

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

This paper proposes HyperForecast, a hyperbolic framework for long-term time series forecasting that exploits the frequency hierarchy induced by multiscale decomposition. It combines learnable temporal decomposition with manifold-based dynamics and introduces both autoregressive and non-autoregressive forecasting variants. Experiments on 11 benchmarks demonstrate competitive performance, while geometric analyses and ablation studies investigate the learned frequency structure and the benefits of hyperbolic representations.

Critical strengths

S1. Interesting geometric perspective on multiscale forecasting. The paper explores a relatively underinvestigated direction by connecting multiscale temporal decomposition with hyperbolic representation learning. Modeling temporal evolution through manifold-based dynamics offers an interesting alternative to conventional Euclidean forecasting and provides a coherent geometric perspective on multiscale representations.

S2. Extensive empirical evaluation and geometric analysis. The paper evaluates both forecasting variants on 11 datasets and reports strong overall performance against diverse baselines. Beyond standard forecasting comparisons, it includes Euclidean-versus-hyperbolic ablations, frequency-centroid analyses, radial-ordering diagnostics, and Gromov hyperbolicity measurements. These analyses provide useful empirical evidence for the proposed framework and make the evaluation more informative than forecasting accuracy alone.

Critical weaknesses

W1. Insufficient justification for the assumed correspondence between frequency ordering and hyperbolic hierarchy. The central motivation relies on treating the frequency ordering induced by multiscale decomposition as a nested, tree-like hierarchy that naturally warrants hyperbolic geometry. However, an ordering from low to high frequencies does not necessarily imply hierarchical containment or branching relationships. Although the paper reports spectral-centroid ordering, radial separation, and low Gromov hyperbolicity values, these observations do not establish that the underlying temporal components possess an intrinsic tree-like structure.

W2. Insufficient justification and validation of the learnable decomposition. The proposed decomposition relies on learnable 1D convolutions to extract trend, coarse seasonal, fine seasonal, and residual components. However, unconstrained convolutional filters do not necessarily preserve the intended frequency characteristics, and the resulting components may be sensitive to learned parameters. Moreover, the choice of exactly two seasonal scales and the boundary between coarse and fine seasonality lack sufficient justification. Although spectral-centroid analyses demonstrate an average frequency ordering, they do not establish that the learned components faithfully represent their intended temporal structures. This is particularly concerning because the subsequent hyperbolic hierarchy is built upon these decomposed components.

Questions for authors

Q1. Why should the frequency ordering induced by multiscale decomposition be interpreted as a tree-like hierarchy rather than a simple ordering of temporal scales? Could the authors provide theoretical or empirical evidence of hierarchical containment or branching relationships among the decomposed components?

Q2. The reported Gromov hyperbolicity measurements are computed on learned embeddings. How can the authors establish that the observed hyperbolicity reflects intrinsic properties of the original time series rather than the imposed representation geometry? Could they provide geometry-independent diagnostics or comparisons with unconstrained representations?

Q3. How do the authors ensure that the learnable convolutional filters extract genuine trend and seasonal components rather than arbitrary predictive representations?

Q4. What motivates the division into exactly two seasonal components, and how is the boundary between coarse and fine seasonality determined? How sensitive are the decomposition and forecasting results to the number of seasonal scales and convolutional kernel sizes?

4 comments
AuthorsOpenReview ↗

Response to Reviewer AGUx (Part 1/4)

W1. Insufficient justification for the assumed correspondence between frequency ordering and hyperbolic hierarchy. The central motivation relies on treating the frequency ordering induced by multiscale decomposition as a nested, tree-like hierarchy that naturally warrants hyperbolic geometry. However, an ordering from low to high frequencies does not necessarily imply hierarchical containment or branching relationships. Although the paper reports spectral-centroid ordering, radial separation, and low Gromov hyperbolicity values, these observations do not establish that the underlying temporal components possess an intrinsic tree-like structure. (Q1, Q2)

Answer: Thank you for raising this. We agree that an ordering of four frequency bands is not by itself a tree, and we do not use it as a proof. What we use is the nesting that the cascade creates: T=Conv1Dtrend(X)\mathbf{T} = \mathrm{Conv1D}_{\mathrm{trend}}(\mathbf{X}), Sc=Conv1Dcoarse(X−T)\mathbf{S}_c = \mathrm{Conv1D}_{\mathrm{coarse}}(\mathbf{X} - \mathbf{T}), Sf=Conv1Dfine(X−T−Sc)\mathbf{S}_f = \mathrm{Conv1D}_{\mathrm{fine}}(\mathbf{X} - \mathbf{T} - \mathbf{S}_c), R=X−T−Sc−Sf\mathbf{R} = \mathbf{X} - \mathbf{T} - \mathbf{S}_c - \mathbf{S}_f, so each finer component is computed from what the coarser ones left over. For one segment of one channel, this gives a chain T,Sc,Sf,R\mathbf{T}, \mathbf{S}_c, \mathbf{S}_f, \mathbf{R}. We do not impose tree edges anywhere in the model, and our claim is about a nested multiscale structure. The motivation is that components of increasing frequency need increasing room, which the volume growth of the ball provides, and we measure this instead of assuming it. The paper measures the nesting independently of the model geometry, as the spectral centroids ordering demonstrates trend < coarse < fine < residual (Figure 3, and additional datasets in Appendix D). A geometry-independent Gromov δ computed directly on the raw segments is given in our response to Q2. In our revised manuscript, we updated Section 1 (lines 50-51) to describe the cascade as a nested multiscale structure, and added Appendix J, which reports Gromov δ on the decomposed data.


W2. Insufficient justification and validation of the learnable decomposition. The proposed decomposition relies on learnable 1D convolutions to extract trend, coarse seasonal, fine seasonal, and residual components. However, unconstrained convolutional filters do not necessarily preserve the intended frequency characteristics, and the resulting components may be sensitive to learned parameters. Moreover, the choice of exactly two seasonal scales and the boundary between coarse and fine seasonality lack sufficient justification. Although spectral-centroid analyses demonstrate an average frequency ordering, they do not establish that the learned components faithfully represent their intended temporal structures. This is particularly concerning because the subsequent hyperbolic hierarchy is built upon these decomposed components. (Q3, Q4)

Answer: Thank you for the comment. The decomposition is an exact split (R=X−T−Sc−Sf\mathbf{R} = \mathbf{X} - \mathbf{T} - \mathbf{S}_c - \mathbf{S}_f) whose scales are tied to the sampling frequency of each dataset. The kernel sizes are set from the sampling frequency as fine = 1 day, coarse = 1 week, and trend = 2 × coarse (two weeks), expressed in samples (the filter weights are what is trained):

DatasetSamplingFine (1 day)Coarse (1 week)Trend (2 weeks)
METR-LA5 min28820164032
Weather, Solar10 min14410082016
Wind15 min966721344
CzeLan, ZafNoo30 min48336672
AQShunyi, AQWan1 hour24168336
PM2.5, Temp3 hours856112
Exchange1 day1714

A larger kernel averages over a wider window, so each stage isolates a slower scale. The resulting frequency ordering trend < coarse < fine < residual is measured on the decomposed signals in Figure 3 and also in Appendix D. Our decomposition ablation in Table 3 isolates the impact of learnable decomposition. Whether the trained filters do more than a fixed split, and how sensitive the results are to the number of seasonal scales and to the kernel sizes, is measured in our answers to Q3 and Q4. We updated Section 4.2 of our revised manuscript (lines 215-217) to state that the kernels span one day, one week, and two weeks and that only the filter weights are learned, and Table 4 in Appendix E.2 lists the kernel sizes of every dataset.


Q1. Why should the frequency ordering induced by multiscale decomposition be interpreted as a tree-like hierarchy rather than a simple ordering of temporal scales? Could the authors provide theoretical or empirical evidence of hierarchical containment or branching relationships among the decomposed components? (W1)

Answer: Thank you for the question. We have answered this in our response to W1.

AuthorsOpenReview ↗

Response to Reviewer AGUx (Part 2/4)

Q2. The reported Gromov hyperbolicity measurements are computed on learned embeddings. How can the authors establish that the observed hyperbolicity reflects intrinsic properties of the original time series rather than the imposed representation geometry? Could they provide geometry-independent diagnostics or comparisons with unconstrained representations? (W1)

Answer: Thank you for the question. We agree that Gromov δ computed on the learned embeddings alone cannot separate the data from the representation, so we also computed it directly on the raw data, before any model or manifold is involved. For each dataset, we cut 24-step segments from the four components of the series (fixed moving-average cascade with the day, week, and two-week kernels), pooled them into one point cloud, and computed its relative Gromov-δ (Euclidean distances, scaled by their 99th percentile). As a reference, the table below also gives the δ of Gaussian noise with the same mean and covariance as the data, which has the same spread but no hierarchy. A lower δ than this reference means the components are more tree-like than their spread alone would produce:

DatasetSamplingδ (decomposed data)δ (Gaussian reference)
Weather10 min0.00470.0298
AQShunyi1 h0.01360.0284
AQWan1 h0.01290.0290
CzeLan30 min0.00930.0367
ZafNoo30 min0.01530.0402
Exchange1 day0.00180.0025
METR-LA5 min0.00530.0208
PM2.53 h0.01590.0296
Solar10 min0.00370.0180
Temp3 h0.01180.0155
Wind15 min0.01130.0238

On all eleven datasets, the pooled components are more tree-like than Gaussian noise with the same spread, so the structure is present in the data before any model is applied. The trained embeddings in Table 2 have δ between 0.002 and 0.023, a similar range. So the Poincaré model preserves this structure; it does not create it. For a comparison with an unconstrained representation, we also measured δ on the embeddings of the matched Euclidean model (same test segments, H = 96, each model in its own metric: Euclidean distance for the Euclidean model, geodesic distance for the Poincaré model as in Table 2):

DatasetEuclidean-model embeddingsPoincaré-model embeddings
AQShunyi0.0220.012
AQWan0.0200.0085
Weather0.0180.018
Temp0.0310.022

The Euclidean embeddings also have a low δ, which is what we expect if the structure comes from the data rather than from the geometry of the model. The Poincaré embeddings have a lower δ than the Euclidean ones on AQShunyi, AQWan, and Temp (by about 45%, 58%, and 29%), so the Poincaré ball represents this structure with less distortion. Taken together, the δ of the raw data and of the embeddings indicates that the tree-like structure comes from the data and not from our hierarchical constraints, and that the Poincaré ball fits it better than a Euclidean space. As for the performance, the benefit may also come from the exponential volume growth of the Poincaré ball, which leaves more room at larger radii for the higher-frequency components, whether or not the data is strictly from a tree. We added Appendix J to our revised manuscript with the δ of the decomposed data, the Gaussian reference, and the δ of the Euclidean model (Tables 18 and 19), with a pointer in Section 5 (lines 428-430).

AuthorsOpenReview ↗

Response to Reviewer AGUx (Part 3/4)

Q3. How do the authors ensure that the learnable convolutional filters extract genuine trend and seasonal components rather than arbitrary predictive representations? (W2)

Answer: Thank you for the question. Real series have no ground-truth trend or seasonality, so we do not claim to recover one. What constrains the filters is the exact split R=X−T−Sc−Sf\mathbf{R} = \mathbf{X} - \mathbf{T} - \mathbf{S}_c - \mathbf{S}_f and the kernel sizes, which tie each stage to a day, a week, or two weeks (our response to W2). To check whether the gain comes from this day, week, and two-week split or from the filters acting as arbitrary trainable maps, we compare the learned filters with fixed ones that have the same kernel sizes but cannot learn anything. We ran HyperForecast (NAR, L = 720) with its three decomposition filters set up in two ways; everything else (kernel sizes, hyperparameters) is unchanged:

  • Learned (default): each filter has the kernel size of its stage, and its weights are trained with the model, starting from the default PyTorch random initialization.
  • Fixed (moving average): each filter has the same kernel size, but its weights are all equal (a plain moving average) and are never trained; the rest of the model trains as usual. This is a fixed decomposition.

Task: long-term forecasting at H = 96 and H = 336; metric: test MSE.

DatasetHLearned (default)Fixed (moving average)
AQShunyi960.6070.616
3360.6700.679
AQWan960.6920.710
3360.7600.778
Temp960.1230.129
3360.1170.121
Weather960.1450.176
3360.2370.259
ZafNoo960.4180.472
3360.5080.538

The learned filters have a lower MSE than the fixed ones on all ten dataset-horizon pairs, so training the filters adds to the gain of the split. A fixed split with the same kernel sizes (a plain moving average) already improves on the model without decomposition (Table 3: AQShunyi, AQWan, Weather), so part of the gain comes from the day, week, and two-week split itself. We added this comparison to our revised manuscript as Appendix E.2 (Table 5), referenced in Section 4.2 (lines 220-222).

AuthorsOpenReview ↗

Response to Reviewer AGUx (Part 4/4)

Q4. What motivates the division into exactly two seasonal components, and how is the boundary between coarse and fine seasonality determined? How sensitive are the decomposition and forecasting results to the number of seasonal scales and convolutional kernel sizes? (W2)

Answer: Thank you for the question. The two seasonal scales are not tuned per dataset. They are the two dominant cycles of these data, a day (fine) and a week (coarse), converted into samples with each dataset's sampling frequency, and the trend kernel is twice the coarse kernel. The boundary between coarse and fine is therefore a physical period that the sampling frequency only decides how many samples that is. To measure the sensitivity, we ran HyperForecast (NAR, L = 720) on AQShunyi, AQWan, and ZafNoo at H = 96 and H = 336, ablating over seasonal scales and kernel size; metric: test MSE. The runs are with the same hyperparameters as the main results.

Number of seasonal scales: the trend (T\mathbf{T}) and the residual (R\mathbf{R}) are always kept; the weekly seasonal component (Sc\mathbf{S}_c) and the daily seasonal component (Sf\mathbf{S}_f) are the two seasonal scales, and a dropped one is set to zero, so the residual absorbs it.

  • 0 scales: T\mathbf{T} and R\mathbf{R} (the weekly and daily cycles stay in R\mathbf{R}).
  • 1 scale: T\mathbf{T}, Sc\mathbf{S}_c and R\mathbf{R} (the daily cycle stays in R\mathbf{R}).
  • 2 scales (default): T\mathbf{T}, Sc\mathbf{S}_c, Sf\mathbf{S}_f and R\mathbf{R}.
DatasetH0 scales1 scale2 scales (default)
AQShunyi960.6200.6060.607
3360.6880.6710.670
AQWan960.7020.7010.692
3360.7640.7630.760
ZafNoo960.4230.4200.418
3360.5140.5190.508

Which seasonal scale helps depends on the dataset; for example, on AQShunyi, the weekly scale gives nearly all of the gain. On AQWan and ZafNoo, the gain comes mostly from the daily scale. Keeping both scales is never meaningfully worse than keeping fewer, so we keep both. We read this as each seasonal stage helping where the series has a cycle at that period and costing little where it does not.

Kernel sizes: the daily and weekly kernels are both scaled by 0.5 or by 2; the trend kernel, which is twice the weekly kernel, scales with them.

DatasetHKernels × 0.5Kernels × 1 (default)Kernels × 2
AQShunyi960.6100.6070.614
3360.6780.6700.682
AQWan960.6970.6920.720
3360.7750.7600.764
ZafNoo960.4230.4180.430
3360.5090.5080.519

Changing the kernels away from one day, one week, and two weeks raises the error in every case, in either direction. We attribute this to what a window can resolve. A kernel as long as the cycle sees one full cycle and can pick it out, a shorter kernel sees only part of a cycle, so some of it leaks into the next stage, and a longer kernel spans several cycles and overlaps the scale of the neighboring stage, so the stages become less distinct. This supports tying the kernels to the physical periods (daily, weekly, bi-weekly) instead of tuning them per dataset. We do not claim that two scales or these exact kernel sizes are optimal for every dataset. We added both studies to our revised manuscript as Appendix E.3 (Tables 6 and 7), referenced in Section 4.2 (lines 220-222).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of the Revised Manuscript

We thank all reviewers for their careful and constructive reviews. We have uploaded a revised manuscript with color-tracked changes and responded to each reviewer.

Revised PDF. The main PDF is the revised manuscript, with all changes colored by reviewer, red for AGUx, blue for nPX4, green for P796, purple for 7GQp, and teal for changes shared across reviewers. A clean version of the same manuscript without colors is provided as supplementary material. All line numbers in our responses refer to the colored PDF.

Main changes.

  • Sections 4.2, 4.5, and 4.6 clarify the decomposition kernels, how the velocity passes through the Poincaré FC layer, the geometric argument for geodesic extrapolation, and what the radial loss encodes.
  • Figure 5a now compares HyperForecast with the corrected Euclidean variant, and the code repository has been updated accordingly.
  • Appendix E adds learned vs. fixed decomposition filters, the sensitivity to the number of seasonal scales and to the kernel sizes, a Euclidean vs. Poincaré FC layer comparison, and a lookback study against the baselines.
  • Appendix F adds 5-seed results for the strongest baselines with Welch t-tests.
  • Appendix J reports Gromov δ on the decomposed data and on the embeddings of a Euclidean model.
  • Appendix K compares HyperForecast with WPMixer.
  • Appendix L analyzes the non-stationarity of the benchmarks, the robustness to missing values, and results on the highly non-stationary Web/CloudOps subsets of GIFT-Eval.

We hope these revisions address the reviewers' concerns, and we would be grateful if they could be taken into account in the final assessment. We are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.