acceptodds
Under review as a conference paper at ICLR 2027

Probabilistic End-to-End Hierarchical Forecasting with Conditional Flow Matching

Abstract

Probabilistic forecasting in hierarchical settings requires generating distributions over future trajectories that satisfy strict aggregation constraints while capturing complex uncertainty. We propose Hierarchical Conditional Flow Matching (HCFM), an end-to-end framework in which future trajectories are generated by integrating a continuous-time dynamical system that combines a learnable residual flow with explicit terms for past consistency and hierarchical coherence. The framework supports two complementary instantiations: projected dynamics that preserve sample-wise coherence throughout integration, and soft-guided dynamics with exact endpoint reconciliation. In both cases, HCFM produces individually coherent forecast samples without committing to a fixed parametric output family. Across six public benchmarks, a single configuration (HCFM-UNet) attains the lowest overall scaled Continuous Ranked Probability Score (sCRPS) on five, with an average relative reduction of 14% against the strongest baseline (CLOVER), with competitive point-forecast accuracy.

Primary area:
generative models
Keywords:
Hierarchical time series forecasting, probabilistic forecasting, conditional flow matching, forecast reconciliation, constrained generative modeling

Reviews

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

This paper studies probabilistic forecasting for hierarchical time series, where every forecast sample must add up correctly across levels (eg store sales summing to regional totals). Existing methods either fix forecasts after training or enforce consistency only for simple distributions like Gaussians, limiting how uncertainty is represented. The authors propose HCFM, a flow-matching generative model that treats forecasting as filling in the future of a trajectory given its observed past. Generation combines a learned component with built-in terms that anchor samples to the observed history and enforce hierarchical consistency. Across multiple benchmarks, HCFM achieves the best probabilistic accuracy.

Critical strengths
  1. Hierarchical time series is common in practice. The task is a practically important task. In addition, the small-data setting is realistic, since many organizations have a single hierarchy with limited history, so they cannot rely on large pretrained generative models.

  2. No Gaussian or other fixed-family assumption is made, so skewed or multi-modal forecasts can be represented. In addition, the constraint handling does not depend on the network, and it works with both a temporal CNN and a U-Net backbone.

Critical weaknesses
  1. The model assumes a fixed hierarchy and a fixed order of series. The U-Net backbone convolves across the series axis, so results may depend on an arbitrary ordering. Adding or removing series means retraining. Would it be possible to perform a simple test, such as shuffling the series order to show how sensitive the results are?

  2. PROFHiT models the full-hierarchy distribution with soft consistency regularization, which is very close to the soft-guided idea. It should be discussed and ideally compared. https://faculty.cc.gatech.edu/~badityap/papers/profhit-kdd23.pdf

  3. Only CLOVER's setup is described ("reproduced from its paper"). How HierE2E, ProbHardE2E and Hier-Transformer-CNF were configured or tuned is not reported.

Questions for authors

In hierarchical time-series forecasting, is linear sum the only studied aggregation? In real life are there any nonlinear aggregation cases?

2 comments

We thank the reviewer for the insightful comments and constructive suggestions. We address each point below.

The model assumes a fixed hierarchy and a fixed order of series. The U-Net backbone convolves across the series axis, so results may depend on an arbitrary ordering. Adding or removing series means retraining. Would it be possible to perform a simple test, such as shuffling the series order to show how sensitive the results are?

The U-Net learns cross-series mixing through stacked convolutions and combines temporal resolutions, following the multiscale motivation in [10]. We added an ordering test in Appendix D.7 (Table 3): four orderings on Labour, Tourism-S, Tourism-L, Traffic and Wiki, with separate retraining for each. Permutations are applied consistently to the data, hierarchy operators and evaluation.

Of the 20 ordering means, 19 fall within the corresponding Table 2 mean ± one SD, used only as a descriptive band; the exception is Labour Order 3 (0.0107). All orderings retain lower mean overall sCRPS than CLOVER on Traffic, Wiki and Tourism-L: the worst scores are 0.0280, 0.1939 and 0.1392, versus 0.0419, 0.2753 and 0.1489. This supports robustness across the tested retrained orderings. The model still assumes a fixed hierarchy; node-set changes are outside this test. We believe that a good research direction for the future would be to explore permutation-equivariant residuals, including suitable graph architectures.

PROFHiT models the full-hierarchy distribution with soft consistency regularization, which is very close to the soft-guided idea. It should be discussed and ideally compared.

We expanded Section 2 to discuss PROFHiT and added the comparison in Appendix I.2. PROFHiT produces conditional Gaussian forecasts through a functional neural process, refines them across nodes, and adds a soft distributional consistency penalty to its variational objective. Its samples don't need to be coherent. HCFM jointly generates all D series over the forecast horizon through a conditional-flow-matching ODE, with trajectory guidance during integration and tangency regularization during training. Final reconciliation ensures sample-wise coherence; projected residuals maintain it throughout integration.

The table reports test sCRPS mean ± SD across three windows after validation selection:

DatasetPROFHiTHCFM-UNetReduction
Labour0.0281 ± 0.00300.0093 ± 0.000366.9%
Traffic0.0719 ± 0.02510.0245 ± 0.012265.9%
Tourism-S0.1264 ± 0.02430.0808 ± 0.011736.1%
Tourism-L0.2246 ± 0.00980.1366 ± 0.006939.2%
Wiki0.3910 ± 0.01510.1838 ± 0.098153.0%

Only CLOVER's setup is described ("reproduced from its paper"). How HierE2E, ProbHardE2E and Hier-Transformer-CNF were configured or tuned is not reported.

We added implementation and selection details in Appendix D.9, referenced from Section 5.2. HierE2E, ProbHardE2E and PROFHiT use released code and published hyperparameters. For PROFHiT, we fixed numerical errors affecting Labour and Traffic under our protocol; without the fix, scores were 0.2594 ± 0.0106 and 0.1305 ± 0.0078, respectively. CLOVER was re-implemented with its published settings.

Hier-Transformer-CNF/FRT has no public code available. After unsuccessful requests to the authors, we re-implemented its published configuration. Using validation sCRPS, we tuned only the learning rate and encoder, decoder and RealNVP layer counts, following the authors’ sensitivity analysis.

HCFM shares loss weights, guidance-schedule parameterization and sampler across datasets; schedule values are learned. Dataset-specific choices are between two U-Net presets selected by validation sCRPS and MQCNN dilations taken from CLOVER (Appendices D.6–D.7). All methods share CLOVER’s preprocessing and features, the same three windows, and checkpoint selection on validation sCRPS.

In hierarchical time-series forecasting, is linear sum the only studied aggregation? In real life are there any nonlinear aggregation cases?

Linear aggregation is the standard benchmark setting [1–3], including our six datasets. Fixed weighted sums also cover averages and rates with known weights [6,7]. Log or log-ratio transforms make some multiplicative or compositional relations additive [8,9]. HCFM supports weighted aggregation (Eq. 30); its guarantees concern fixed linear idempotent projections under the assumptions in Appendices A–B.

Nonlinear aggregates include maxima, minima and quantiles. To our knowledge, this setting lacks established benchmarks. ProbHardE2E supports nonlinear constraints [5], although its hierarchical experiments aggregate linearly. Extending HCFM to these cases would require nonlinear constraint enforcement. We now state this scope and limitation in the conclusion.

We hope our responses have clarified your questions, and if so, we would be grateful if you would reconsider your evaluation.

References:

  1. Hyndman, R. J., Ahmed, R. A., Athanasopoulos, G., & Shang, H. L. (2011). Optimal combination forecasts for hierarchical time series. Computational Statistics & Data Analysis, 55(9), 2579–2589. doi:10.1016/j.csda.2011.03.006

  2. Wickramasuriya, S. L., Athanasopoulos, G., & Hyndman, R. J. (2019). Optimal forecast reconciliation for hierarchical and grouped time series through trace minimization. Journal of the American Statistical Association, 114(526), 804–819. doi:10.1080/01621459.2018.1448825

  3. Athanasopoulos, G., Hyndman, R. J., Kourentzes, N., & Panagiotelis, A. (2024). Forecast reconciliation: A review. International Journal of Forecasting, 40(2), 430–456. doi:10.1016/j.ijforecast.2023.10.010

  4. Utkarsh, U., Cai, P., Edelman, A., Gomez-Bombarelli, R., & Rackauckas, C. V. (2025). Physics-constrained flow matching: Sampling generative models with hard constraints. NeurIPS 2025.

  5. Utkarsh, U., Maddix, D. C., Ma, R., Mahoney, M. W., & Wang, B. (2026). End-to-end probabilistic framework for learning with hard constraints. ICLR 2026.

  6. Shang, H. L., & Hyndman, R. J. (2017). Grouped functional time series forecasting: An application to age-specific mortality rates. Journal of Computational and Graphical Statistics, 26(2), 330–343.

  7. Hubrich, K. (2005). Forecasting euro area inflation: Does aggregating forecasts by HICP component improve forecast accuracy? International Journal of Forecasting, 21(1), 119–136.

  8. Hyndman, R. J., Booth, H., & Yasmeen, F. (2013). Coherent mortality forecasting: The product-ratio method with functional time series models. Demography, 50(1), 261–283.

  9. Bergeron-Boucher, M.-P., Canudas-Romo, V., Oeppen, J., & Vaupel, J. W. (2017). Coherent forecasts of mortality with compositional data analysis. Demographic Research, 37(17), 527–566.

  10. Yu, B., Yin, H., and Zhu, Z. (2019). ST-UNet: A Spatio-Temporal U-Network for Graph-structured Time Series Modeling. arXiv:1903.05631. https://arxiv.org/abs/1903.05631

Reviewer_Kh2VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies probabilistic forecasting for hierarchical time series, where forecasts must satisfy aggregation constraints. The proposed HCFM model uses conditional flow matching to generate future trajectories, with additional terms for consistency with the observed past and the hierarchy. The authors consider two ways to enforce coherence: projecting the learned vector field during sampling, or applying soft guidance followed by a final reconciliation step. They establish coherence guarantees and evaluate the method on six datasets. HCFM-UNet obtains the lowest mean overall sCRPS on five of them. The experiments also include component ablations, calibration checks, multivariate scores, and sampling-cost comparisons.

Critical strengths
  • The paper gives a concrete construction for combining flexible probabilistic forecasts with hierarchical constraints. The projected variant preserves sample-wise coherence during integration, and the treatment of per-series normalization makes clear how the constraints should be applied in the coordinates used by the model.

  • The empirical results are promising. HCFM-UNet achieves the lowest mean overall sCRPS on five of six datasets. The evaluation also includes multivariate scores, calibration checks, and several ablations, including the same-backbone Leaves Only comparison.

Critical weaknesses
  • The contribution of continuous coherence guidance is not fully isolated. Table 13 provides a useful same-backbone Leaves Only comparison, but the full model also changes the generated state and training problem. On Wiki, Leaves Only achieves 0.1877 sCRPS versus 0.1838 for the full model. This supports an improvement from the full configuration, but does not establish how much comes specifically from continuous coherence guidance.

  • Table 2 contains a numerical discrepancy that needs clarification. For HCFM-UNet on Favorita, the four level-wise standard deviations are 0.0192, 0.0166, 0.0124, and 0.0149, while the overall standard deviation is 0.0214. If the overall score is the equal-weight average of the four level scores within each window, using the same runs and standard-deviation convention, its standard deviation cannot exceed 0.015775. A different aggregation procedure or different runs could explain this difference, the paper should specify which applies.

  • The temporal coverage of the evaluation is unclear. Section 5.4 describes three non-overlapping test windows, but does not specify the number or placement of forecast origins within each window. This matters particularly for Traffic, where the forecast horizon is one day. Clarifying whether each window contains a single forecast origin or multiple rolling origins would help assess how broadly the reported gains hold across time.

Questions for authors
  1. Have you evaluated the same trained full-hierarchy model with continuous coherence guidance disabled during sampling, keeping the solver and final reconciliation unchanged? Does the guidance improve forecast accuracy under this control?

  2. How are the overall Favorita scores and standard deviations computed in Table 2? Do the overall and level-wise entries use the same runs and standard-deviation convention? Please clarify the aggregation procedure that explains the reported difference.

  3. How many forecast origins are evaluated within each test window, and how are they selected? In particular, does Traffic use a single origin per window or multiple rolling origins, and which dates are covered?

4 comments

We thank the reviewer for the insightful comments and constructive suggestions. We address each point below.

Have you evaluated the same trained full-hierarchy model with continuous coherence guidance disabled during sampling, keeping the solver and final reconciliation unchanged? Does the guidance improve forecast accuracy under this control?

We appreciate this suggestion. We added this control to Appendix F.4 (Tables 11 and 12). For each evaluation window, we reused the same trained soft-guided HCFM-UNet checkpoint and set λ(τ)=0\lambda(\tau) = 0 only during sampling. All other settings, including the solver, step count and final reconciliation, remained unchanged.

DatasetSoft-guidedGuidance disabledRelative reduction
Labour0.0093 ± 0.00030.0101 ± 0.00067.9%
Tourism-S0.0808 ± 0.01170.0958 ± 0.026115.7%
Tourism-L0.1366 ± 0.00690.1379 ± 0.01550.9%
Traffic0.0245 ± 0.01220.0280 ± 0.023912.5%
Wiki0.1838 ± 0.09810.1919 ± 0.10104.2%

Guidance lowers mean overall sCRPS and spatial variogram score on all five datasets, and energy score on four, with Wiki the exception. These results support its benefit beyond endpoint reconciliation. Relative sCRPS reductions are computed from the displayed means.

How are the overall Favorita scores and standard deviations computed in Table 2? Do the overall and level-wise entries use the same runs and standard-deviation convention? Please clarify the aggregation procedure that explains the reported difference.

Thank you for identifying this inconsistency. We apologize: values from different metric outputs were accidentally mixed when preparing the original table. We reran the Favorita HCFM-UNet experiments and corrected Table 2:

Hierarchy levelsCRPS, mean ± SD
Overall0.3351 ± 0.0111
Country0.2465 ± 0.0095
Country/State0.3206 ± 0.0080
Country/State/City0.3344 ± 0.0160
Country/State/City/Store0.4387 ± 0.0130

All entries use the same three evaluation windows and the sample SD convention. Within each window, Favorita’s pooled overall sCRPS equals the arithmetic mean of its four level scores. We then compute the overall mean and sample SD across the three overall window scores; the overall SD is not an average of the level SDs. Appendix E.3 describes this procedure. The corrected Favorita result is comparable to CLOVER, and we have adjusted our claims accordingly.

How many forecast origins are evaluated within each test window, and how are they selected? In particular, does Traffic use a single origin per window or multiple rolling origins, and which dates are covered?

Each test window contains one forecast origin. For a series of length NN and horizon HH, the origins are fixed at N−3HN - 3H, N−2HN - 2H and N−HN - H, giving the final three consecutive, non-overlapping test blocks of length HH (Section 5.3; Appendix E.2).

Each origin has a chronological split:

  • Training: all data before the validation period.
  • Validation: the period of HH steps immediately before the test set's forecasting origin, used to select checkpoints and hyperparameters.
  • Test: the next HH steps, used only to evaluate the selected model.

A separate model is trained for each origin. Test data are never used for fitting or selection at that origin. As the origin advances by HH, the preceding test block becomes available history for the next split. Reported means and SDs describe variation across these three evaluation windows.

For Traffic (H=1H = 1 day), the main evaluation forecasts 29–31 December 2008 under the benchmark’s sequential daily index, scoring all 207 series. To broaden coverage, we added 21 daily forecasts for 11–31 December 2008, with a separately trained model for each method and origin. Appendix I.4, Table 19 reports mean overall sCRPS ± SD across these 21 origins:

MethodOverall sCRPS
HCFM-UNet0.0138 ± 0.0080
HCFM-MQCNN0.0528 ± 0.0357
CLOVER0.0315 ± 0.0175
HierE2E0.0370 ± 0.0178

HCFM-UNet achieves lower sCRPS than CLOVER on 21/21 origins and HierE2E on 20/21, supporting the favorable Traffic results over this broader evaluation period.

Thank you to the author for providing additional explanations and revising Favorita's results. What worries me the most about the original manuscript is that the errors in the main table directly affect the interpretation of the results. According to the author's summary method, the overall standard deviation after four layers of equal weight averaging cannot reach the values reported in the original table. The revised results are consistent with the summary rules, but the average sCRPS difference between HCFM UNet and CLOVER is only 0.0003. As for these three test windows, I would consider their performance on Favorita to be comparable. After reading the revised draft, I still have two doubts.

  • The manuscript does not clearly explain which method the flow baseline in Table 2 is. Section 5.2 and Table 2 describe Hier Transformer CNF (Wang et al., 2022), while Appendix D.9 describes the re implementation as Hier Transformer CNF/FRT. In related work, FRT (Wang et al., 2025) is introduced as another paper. This makes it difficult for me to determine which architecture Table 2 is actually comparing, and thus it is difficult to determine how much advantage HCFM has over newer hierarchical flow methods.
  • The existing results still have limited support for multi-step joint prediction. SCRPS measures the marginal distribution of each sequence at different prediction time points, while the ES and VS in the appendix mainly examine the dependencies between different sequences at the same time point. These results are valuable, but they are not sufficient to demonstrate that the model has learned the dependencies between different prediction time points more accurately. After expanding Traffic to 21 prediction starting points, the evidence for one-day predictions became more sufficient, but each prediction span was still one day, so this set of results cannot answer the question of multi-step trajectories.

We are grateful for your careful comments, which have helped us improve our paper. We hope the revisions below address your concerns.

1. Favorita. We really appreciate your check of the corrected values against the aggregation rule. HCFM-UNet and CLOVER differ by 0.0003 in mean overall sCRPS (0.3351 ± 0.0111 vs 0.3354 ± 0.0103), well within the variation across windows. We agree they are comparable and no longer present Favorita as a win. The Abstract reports that HCFM-UNet has the lowest mean overall sCRPS on four of six datasets. Section 5.4 and the Conclusion explicitly describe HCFM-UNet and CLOVER as comparable on Favorita.

2. Flow baseline. The baseline in Tables 2 and 13 is our reimplementation of the core architecture shared by Hier-Transformer-CNF (Wang et al., 2022, ICDMW) and FRT (Wang et al., 2025, KDD): an autoregressive Transformer produces all-level base forecasts, a conditional RealNVP model generates bottom-level samples, and bottom-up aggregation ensures coherence. Both papers, when compared between them, report identical CRPS means and standard deviations for Traffic, Tourism and Tourism-L under a different evaluation protocol from ours. As author-provided code was unavailable for our experiments, we followed FRT’s implementation details (Appendix B) and selected the learning rate and layer counts using validation sCRPS under our protocol (Appendix D.9).

We report this baseline once as Hier-Transformer-CNF/FRT, consistently labelled in Sections 2 and 5.2, Tables 2 and 13, and Appendix D.9. HCFM-UNet achieves lower mean overall sCRPS (Table 2) and relSE (Table 13) than this reimplementation on all six datasets.

3. Multi-step trajectories. Thank you for highlighting the need to evaluate dependence across forecast steps. We agree that sCRPS evaluates each step marginally, our previous ES and spatial VS compare series at a single step, and the 21-origin Traffic study uses H=1H=1. We have added the temporal variogram score (VS) and a seven-day Traffic experiment.

Temporal dependence. Appendix I.3 defines temporal VS (Eq. 88), and Table 17 reports spatial and temporal VS at the leaf level and with all hierarchy levels pooled. Following Scheuerer and Hamill (2015), we assess dependence across forecast steps using temporal VS computed from joint forecast trajectories, alongside sCRPS for marginal accuracy. Baron et al. (2025) similarly combine CRPS and temporal VS to evaluate multi-step sample paths.

Seven-day Traffic forecasts. Labour, Tourism-S, Tourism-L and Wiki already use horizons of 4–12 steps. We trained and evaluated HCFM-MQCNN, HCFM-UNet, CLOVER and HierE2E on Traffic at H=7H=7, using a common set of three rolling windows. The table combines pooled temporal VS (Table 17) and overall sCRPS (Appendix I.4, Table 19(b)).

MethodPooled temporal VS (×10³)Overall sCRPS
HCFM-UNet0.28 ± 0.100.0162 ± 0.0066
HCFM-MQCNN1.05 ± 0.640.0508 ± 0.0180
CLOVER0.91 ± 0.430.0389 ± 0.0184
HierE2E0.99 ± 0.330.0386 ± 0.0119

Mean ± standard deviation across three rolling windows; lower is better.

HCFM-UNet achieves approximately 69% and 72% lower mean pooled temporal VS than CLOVER and HierE2E, respectively, and 58% lower mean overall sCRPS than both. Its sCRPS advantage thus extends to seven days. At H=1H=1, it is 42% below CLOVER in Table 2 and 56% below CLOVER across the 21 origins (Table 19(a)).

Across datasets. Relative to the best baseline in Table 17, the best HCFM variant reduces mean pooled temporal VS by approximately 46% on Labour, 69% on Traffic and 24% on Wiki. On Tourism-S, HCFM-MQCNN and CLOVER are within 1%; on Tourism-L, CLOVER’s mean is approximately 5% below the best HCFM variant. At the leaf level, HCFM variants attain the lowest mean temporal VS on Labour, Traffic and Wiki; CLOVER leads on Tourism-S and Tourism-L (Table 17).

These complementary evaluations provide evidence of improved multi-step forecast quality. Section 5.3 describes the evaluation protocol, Section 5.4 summarises the multivariate findings, and Appendices I.3–I.4 provide the metric definitions and detailed results.

If these additions address your concerns, we would be grateful if you would reconsider your score.

References

Scheuerer, M., & Hamill, T. M. (2015). Variogram-based proper scoring rules for probabilistic forecasts of multivariate quantities. Monthly Weather Review, 143, 1321–1334. DOI

Baron, E., Oreshkin, B., Ma, R., Zhang, H., Torkkola, K., Mahoney, M. W., Wilson, A. G., & Konstantinova, T. (2025). Efficiently generating correlated sample paths from multi-step time series foundation models. arXiv. DOI

Thank you for the clarifications and additional experiments. The revision clarifies the flow baseline and addresses my concerns about multi-step evaluation through temporal variogram scores and the seven-day Traffic experiment. I appreciate these improvements. However, my main reservation remains that the methodological contribution is relatively incremental and the theoretical insight is limited. The authors have clarified how HCFM differs from prior hierarchical flow methods. Still, the contribution largely consists of combining conditional flow matching, structural constraints, and past–future co-evolution for hierarchical forecasting. I am not yet convinced that this combination represents a sufficiently substantial methodological advance over existing generative forecasting and coherence constructions. The theoretical analysis mainly establishes coherence and characterizes local properties under simplified settings. It provides limited insight into why the complete method improves probabilistic forecast quality or how its components jointly produce the reported gains. The new results strengthen the empirical evidence, but they do not sufficiently change my assessment of the contribution.

Reviewer_u2bXOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This papers tackles the problem of probabilistic hierarchical time series forecasting using conditional flow matching. One of the main contributions in this paper, is to apply the hierarchical constraints directly into the continuous-time generative dynamics itself. Specifically the mapping from random noise to the final output forecast maintains hierarchical coherence. This is achieved through guidance terms and projection in the velocity equations.

Evaluation on standard benchmark datasets show improved forecast accuracy compared to state of the art baselines.

Critical strengths

The paper presents a method for probabilistic hierarchical forecasting where the forecasts are sample-wise coherent i.e. each sample from the output distribution is coherent. The paper presents a novel CFM approach where the flow is constrained to the subspace of coherent outputs. The backbone architectures used for predicting the flow are U-Net and MQCNN, and yield strong results beating the state of the art baselines.

Critical weaknesses

A simple baseline is missing: A CFM based forecasting method where final forecast is projected instead of a projection throughout the flow evolution.

Minor: Some missing references and missing discussion

  • Wang et al., “FRT: Flow-based Reconcile Transformer for Hierarchical Time Series”, KDD 2025.
  • Wang, “NeuralReconciler for Hierarchical Time Series Forecasting”, WSDM 2024.
  • Kamarthi et al., “When Rigidity Hurts: Soft Consistency Regularization for Probabilistic Hierarchical Time Series Forecasting”, KDD 2023.
Questions for authors

Is it necessary to evolve both the past and the future time series rather than fix the past and only evolve the future time series?

1 comment

We thank the reviewer for the positive assessment and constructive suggestions. We address each point below.

A simple baseline is missing: A CFM based forecasting method where final forecast is projected instead of a projection throughout the flow evolution.

We added the requested baseline as CFM in Appendix F.4 (Tables 11 and 12). It conditions directly on the observed past, removes HCFM’s structural terms, and projects the final forecast by P. HCFM-UNet lowers mean overall sCRPS on all five datasets, by 10% (Tourism-L) to 36% (Wiki), with lower energy scores on all five and variogram scores on four. CFM has higher overall sCRPS than CLOVER on four of five datasets, so the gains do not come from flow matching alone.

To isolate guidance, Appendix F.4 also evaluates the same trained soft-guided HCFM-UNet checkpoints without it, keeping all other settings and final projection unchanged. Guidance lowers mean overall sCRPS on all five datasets, from 1% on Tourism-L to 16% on Tourism-S (Table 11). It also lowers the variogram score on all five and the energy score on four, with Wiki the exception (Table 12), supporting its benefit beyond endpoint projection.

The default HCFM (Table 2) already projects only the final forecast. The soft-guided model (Eq. 4) uses coherence guidance during integration and applies the bottom-up projector P once at the end (Section 4.3). The projected-residual variant (Eq. 5) enforces coherence throughout integration. For HCFM-UNet, Table 5 finds no clear winner: the default is better on Tourism-S, Labour and Wiki, and projected residuals on Traffic and Tourism-L.

Minor: Some missing references and missing discussion

We have expanded Section 2 to discuss FRT, NeuralReconciler and PROFHiT, and added a focused PROFHiT comparison in Appendix I.2. The discussion places HCFM alongside existing flow-based forecasters. FRT (Wang et al., KDD 2025) shares the core architecture of our Hier-Transformer-CNF baseline (Wang et al., ICDMW 2022) and reports identical results on four original datasets. Both condition a bottom-level RealNVP flow on autoregressive Transformer forecasts and obtain coherence by aggregation with S. NeuralReconciler (Wang, WSDM 2024) combines normalizing-flow forecasting with jointly trained attention-based reconciliation.

We also clarify the connection to PROFHiT’s soft hierarchy constraints. Its functional neural process produces conditional Gaussian forecasts, refined across nodes, while a consistency penalty augments variational training to align parent distributions with those of aggregated children. HCFM jointly samples all D series across the forecast horizon through a conditional-flow-matching ODE, with guidance during integration and a tangency penalty during training. Final reconciliation ensures exact sample-wise coherence; projected residuals maintain it throughout integration. The distinction concerns both joint trajectory modeling and how coherence enters generation.

To support this discussion empirically, we ran PROFHiT’s released implementation, fixing numerical errors affecting Labour and Traffic under our evaluation protocol. Without the fix, scores were 0.2594 ± 0.0106 for Labour and 0.1305 ± 0.0078 for Traffic. Models were selected using validation sCRPS; Appendix I.2 reports the corresponding test means and standard deviations across three evaluation windows. HCFM-UNet achieves lower mean overall test sCRPS on all five datasets, with reductions of 36–67%.

Is it necessary to evolve both the past and the future time series rather than fix the past and only evolve the future time series?

Co-evolving the past is not necessary: conditioning and coherence are independent choices, and HCFM supports a fixed past under either coherence mechanism (Section 3). In the Past-Frozen configuration, the observed past is a fixed input at every ODE step, the predicted past-block velocity is discarded, and only the future block is integrated. Both samplers use the same trained projected-residual HCFM-UNet weights, data splits and solver, so the comparison isolates the effect of past dynamics. Freezing introduces no mismatch in past-block inputs: the past block equals the observations throughout the training interpolant (Section 4.3). Co-evolution yields lower means in all 15 comparisons across five datasets: sCRPS (Table 11), energy and variogram scores (Table 12). The largest gap is on Tourism-S (sCRPS 0.0891 vs 0.1209; energy score 5.70e3 vs 8.76e3), and Traffic sCRPS improves from 0.0247 to 0.0209. Following Appendix F.4, we interpret this as an inductive bias. The learned residual acts on a shared past-future trajectory, while an explicit, separately weighted force enforces fidelity to observations. This avoids relying only on a learned conditioning pathway, which can underuse history when data are limited.

If our answers have resolved your questions, we would be grateful if you could note this in the discussion and reflect it in your final assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Review Process

We thank the reviewers for their detailed reviews. All revisions are highlighted in blue in the PDF. Below we summarise the revisions and additional results.

  1. The Favorita entries in Table 2 have been corrected, and all Table 2 overall sCRPS entries are consistent with the aggregation rule in Appendix E.3. HCFM-UNet and CLOVER are reported as comparable on Favorita (Kh2V). Appendix D.9 documents baseline code sources and tuning: HierE2E, ProbHardE2E and PROFHiT use the authors’ released code and published hyperparameters, while CLOVER is reimplemented using its published hyperparameters. The flow baseline, labelled Hier-Transformer-CNF/FRT throughout, is our reimplementation based on the implementation details in FRT (Wang et al., 2025, Appendix B); author-provided code was unavailable for our experiments. Its learning rate and layer counts were selected using validation sCRPS (RfAg, Kh2V). FRT, NeuralReconciler and PROFHiT are now discussed in Section 2, and PROFHiT is added as a baseline. HCFM-UNet achieves lower mean overall sCRPS than PROFHiT on all five evaluated datasets (Table 14; u2bX, RfAg).

  2. The ablations requested by u2bX and Kh2V evaluate coherence guidance and past–future co-evolution. Disabling coherence guidance in the same trained HCFM-UNet model increases mean overall sCRPS on all five datasets. A directly conditioned CFM with the same backbone and final reconciliation also has higher mean overall sCRPS on all five. For the projected-residual model, freezing the past instead of co-evolving it increases mean sCRPS, energy score and spatial variogram score on all five (App. F.4, Tables 11 and 12). These controls support the benefits of coherence guidance and past–future co-evolution beyond final reconciliation. Robustness of the evaluation. Each dataset is evaluated on three rolling windows under a common protocol (Apps. D.9 and E.2). On Traffic, an expanded evaluation over 21 daily origins confirms the ranking: HCFM-UNet achieves lower sCRPS than CLOVER on 21/21 origins and HierE2E on 20/21 (App. I.4, Table 19(a); Kh2V, RfAg). Across four separately retrained series orderings, 19 of 20 means lie within the corresponding Table 2 mean ± 1 SD. HCFM-UNet remains below CLOVER’s Table 2 mean on Traffic, Wiki and Tourism-L under every ordering (Table 3; RfAg).

  3. Table 17 now reports temporal variogram scores, comparing pairs of forecast steps within each series, on Labour, Tourism-S, Tourism-L and Wiki ((H=4)–(12)), and on Traffic with all four methods separately trained and evaluated at (H=7) (App. I.3; Kh2V). On seven-day Traffic forecasts, HCFM-UNet achieves approximately 69% and 72% lower mean overall temporal VS than CLOVER and HierE2E, respectively (Table 17), together with approximately 58% lower mean overall sCRPS than both (Table 19(b)). These results support improved multi-step forecast quality.

  4. Section 7 clarifies the limitations imposed by fixed hierarchies and linear aggregation and discusses directions for future work (RfAg). Every requested experiment and clarification is included in the revised manuscript. We would be grateful if the reviewers could reconsider their score.

We thank the reviewers again for their time and for helping us strengthen the paper. We would be grateful if they would reconsider their scores in light of these revisions and additional results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.