acceptodds
Under review as a conference paper at ICLR 2027

CLASP: Cluster-Conditioned Latent Alignment with Shared Prototypes for Entity-Level Time-Series Forecasting

Abstract

Forecasting across heterogeneous entities requires sharing statistical strength while preserving entity-specific variation. We introduce CLASP, Cluster-conditioned Latent Alignment with Shared Prototypes, a probabilistic latent partial-pooling mechanism. From an entity's observed history, CLASP constructs context-dependent latent distributions and aligns them with shared Gaussian prototype priors capturing population structure. Prototype-assignment weights, obtained from provided labels or learned from history and metadata, combine independent component draws, inducing an aggregate Gaussian posterior during training and a prototype-induced aggregate Gaussian prior at deployment. This partially pooled latent augments the observed history and conditions diverse forecasting backbones, with an optional validation-fitted readout. We characterize the latent train–deployment discrepancy by deriving its exact Gaussian KL divergence, proving a data-processing upper bound by component KL divergences, and showing that under one-hot assignments it equals the uncapped CLASP pooling penalty. Under additional bounded-loss, sampling, and sample-independence assumptions, we obtain PAC-Bayes guarantees for posterior-sampled risk under a fixed prediction map and a single entity-level latent law. We evaluate CLASP on held-out WEATHER-5K stations and eICU patients across forecasting horizons and provided and inferred population structure. Controlled ablations show substantial gains from prototype alignment and prior learning beyond the optional readout, with additional improvements from probabilistic and multi-prototype representations. Within-backbone comparisons show consistent gains from the complete CLASP pipeline across the evaluated probabilistic and deterministic forecasters.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series forecasting, machine learning, probabilistic modeling, representation learning, hierarchical modeling, generalization, deep learning.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yQvxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses forecasting across heterogeneous entities, such as weather stations and ICU patients, where models must share statistical strength while preserving entity-specific dynamics and generalize to unseen entities without retraining. The authors argue that existing neural approaches share information through deterministic parameters, embeddings, or cluster vectors, lacking explicit distributional partial pooling. The paper proposes a method called CLASP, a plug-in module that aligns history-dependent Gaussian latent distributions with shared Gaussian prototype priors via a weighted KL penalty, using provided labels or a learned router. They also derive the exact KL between these training and deployment laws, a data-processing bound, and an equivalence with the pooling penalty under one-hot assignments. The authors have conducted experiments on held-out WEATHER-5K stations and eICU patients, which show error reductions over strong baselines and consistent gains across forecasting backbones.

Critical strengths
  1. This paper is addressing a practically important problem: most real forecasting involves many related time-series, how to share information across such populations without training a individual model each time is a central question.

  2. The paper shows that a distributional prior over groups along with a KL alignment penalty can be added to very different forecasters including recurrent, convolutional, state-space, transformer, diffusion and flow models. This can pave the way for adapt the method to widely used time-series foundation models.

  3. The proposed framework works with many backbones, scales to many entities; the pipeline produces full predictive distributions and the prototypes expose population structure can be easily examined.

Critical weaknesses
  1. In empirical evaluation, transferring to a new population is never tested. For example, in weather-5k dataset, test stations come from the same 5 zones and same 125 station balanced subset; in eICU, the split is by patient only, with no hospital-level holdout reported.

  2. The objective of the method is similar to the regularizer of a Gaussian-mixture VAE, several closely related works are uncited or not compared against: VaDE: https://www.ijcai.org/proceedings/2017/273 GMVAE: https://arxiv.org/abs/1611.02648 CVAEs: https://papers.nips.cc/paper_files/paper/2015/hash/8d55a249e6baa5c06772297520da2051-Abstract.html

  3. It seems that the theory (88 pages in total is too long) adds very little actionable content. The results are standard, e.g. the Gaussian KL is a textbook formula, theorem 1 is data processing inequality. And its assumptions don't hold for time series, where it assumes windows within an entity are iid.

Questions for authors

How does the method perform for entities with long histories and varying length?

4 comments
AuthorsOpenReview ↗

W1. Transfer to new populations

We thank the reviewer for the comments on population transfer and connections to variational models. All manuscript references follow the revised paper numbering.

W1. Transfer to new populations

We agree that the original headline splits alone do not establish cross-zone or cross-hospital transfer. While the original submission already included an imbalanced WEATHER-5K setting, with cluster-wise CLASP results under balanced and imbalanced training configurations across all five zones and three forecast horizons in Appendix E.3.6. To directly test transfer to new populations, we added leave-one-zone-out WEATHER-5K evaluation and a hospital-disjoint eICU evaluation.

WEATHER-5K: transfer to an unseen geographic zone.

We use a five-fold leave-one-zone-out protocol at 24 hours, excluding one zone from task-specific training and evaluating on that zone in each fold. Models are trained or fine-tuned using stations from the other four zones; Timer-XL ZS is evaluated without task-specific training. Table R1 reports MAE / MSE for each held-out zone, with the best values in bold; ZS/FT denotes zero-shot/fine-tuned evaluation.

Table R1.

ModelTropicalDryTemperateContinentalPolar
D3VAE2.906 / 20.9322.700 / 19.7842.504 / 17.7042.788 / 21.4953.445 / 33.925
TSDiff1.429 / 4.7331.585 / 6.8411.717 / 7.8941.957 / 11.8632.801 / 27.359
Timer-XL ZS0.917 / 2.5021.441 / 5.2521.552 / 7.3911.694 / 9.3621.857 / 11.157
Timer-XL FT1.134 / 3.7321.863 / 9.3612.109 / 12.2162.065 / 12.2171.873 / 10.663
NsDiff1.071 / 3.0431.555 / 6.5101.821 / 8.9331.772 / 9.1731.633 / 7.842
TSFlow0.890 / 2.3401.436 / 5.5351.520 / 6.3711.571 / 7.7131.713 / 9.403
CVAE1.570 / 3.6081.909 / 5.9461.865 / 6.9481.970 / 8.9401.969 / 16.678
CAMul1.070 / 3.2291.412 / 5.4641.595 / 6.6801.587 / 7.4051.831 / 11.172
CCM1.149 / 3.6941.809 / 8.5762.066 / 11.2601.897 / 10.4412.027 / 12.422
CLASP0.855 / 2.4311.322 / 5.0641.493 / 6.1731.521 / 7.0001.399 / 6.246

CLASP achieves the lowest MAE in all five zones and the lowest MSE in four; TSFlow has lower Tropical MSE, at 2.340 versus 2.431 for CLASP. In the Polar zone, CLASP reduces MAE/MSE by 14.3%/20.4% relative to the strongest baseline, NsDiff. These results support cross-zone transfer under the evaluated 24-hour forecasting protocol.

eICU: transfer to held-out hospitals.

Since there are patients recorded at more than one hospital, we remove those patients before splitting the remaining hospitals into mutually disjoint training, validation, and test sets. Each retained patient is assigned to their hospital’s split, so both hospital and patient identities are disjoint across splits. Table R2 reports the resulting split, comprising 208 hospitals and 91,262 unique patients.

Table R2.

Dataset splitHospitalsPatients
Train17273,000
Validation189,129
Test189,133
Total20891,262

Table R3 reports MAE / MSE on patients from the 18 held-out test hospitals, with the best values in bold.

Table R3.

Model6 hours12 hours24 hours
D3VAE6.995 / 115.5397.147 / 118.2887.512 / 135.310
TSDiff6.821 / 112.0517.105 / 115.2457.424 / 130.245
Timer-XL ZS6.234 / 101.8837.200 / 126.0938.457 / 163.371
Timer-XL FT6.552 / 106.1577.049 / 120.2387.605 / 137.121
NsDiff6.516 / 104.5337.109 / 120.8037.628 / 137.277
CVAE6.682 / 100.2937.325 / 119.4517.345 / 130.933
CAMul5.984 / 93.1346.549 / 106.6137.138 / 121.853
CCM6.324 / 102.8046.749 / 114.3357.389 / 134.362
CLASP5.557 / 88.1276.150 / 103.3426.815 / 122.051

CLASP achieves the lowest MAE at all three horizons and the lowest MSE at 6 and 12 hours. Relative to CAMul, the strongest listed baseline for every horizon and metric, CLASP reduces MAE by 7.1%/6.1%/4.5% at 6/12/24 hours. At 24 hours, CAMul has slightly lower MSE, at 121.853 versus 122.051 for CLASP. These results support transfer to held-out hospitals under the reported split.

Presentation revisions.

Added cross-zone and cross-hospital evaluations in Section 4.1, with protocols and results in Appendices E.3.7 and E.3.9, respectively, and updated the transfer discussion in Appendix F.2. The original imbalanced WEATHER-5K evaluation remains in Appendix E.3.6.

AuthorsOpenReview ↗

W2. Relationship to Gaussian-mixture VAEs and forecasting baselines

W2. Relationship to Gaussian-mixture VAEs and forecasting baselines

We thank the reviewer for highlighting these related works. We agree that Gaussian priors and KL-based latent alignment are established components of variational models, including VaDE, GMVAE, and CVAE. CLASP’s contribution lies in its entity-level pooling construction using shared Gaussian prototypes and its analysis of the associated training–deployment discrepancy. Specifically, CLASP pools independent Gaussian component draws, which differs from sampling a latent representation from a selected mixture component. This construction enables CLASP to serve as an adaptive module that augments existing deterministic and probabilistic forecasters while retaining the host forecasting model.

To address the request for additional comparisons, we added CVAE as a forecasting baseline following this reviewer’s suggestion and evaluated it both with and without CLASP. VaDE and GMVAE are closely related generative clustering methods, but their original formulations do not specify a history-to-future forecasting model. Evaluating them here would therefore require constructing forecasting extensions with additional architectural choices; we have not included such extensions in the present comparisons.

Across D3VAE, TSDiff, NsDiff, Timer-XL, TSFlow, CVAE, and CAMul, the CLASP-augmented models achieve lower MAE than their corresponding hosts at every evaluated WEATHER-5K horizon (Table 24, Appendix E.3.2). For CVAE specifically, 24-hour MAE decreases from 2.279 to 1.621, a 28.9% reduction. These results support CLASP’s usefulness across forecasting architectures, including a conditional variational model.

We also added CAMul and CCM following reviewer oEHV’s suggestions. Table R4 summarizes the comparisons reported in main-text Tables 1–2 as MAE / MSE, with the best values in bold. In these experiments, CLASP uses the five provided WEATHER-5K zones and four inferred eICU groups.

Table R4.

WEATHER-5K

Model24 hours72 hours168 hours
CVAE2.279 / 14.8982.928 / 22.0763.338 / 27.660
CAMul1.595 / 8.0192.049 / 15.2452.561 / 20.099
CCM1.568 / 5.8272.326 / 16.7662.664 / 23.866
CLASP1.291 / 5.1111.902 / 11.8482.290 / 17.193

eICU

Model6 hours12 hours24 hours
CVAE9.905 / 188.45912.572 / 282.98110.090 / 220.934
CAMul6.984 / 103.1297.549 / 116.6078.102 / 133.299
CCM8.946 / 117.5019.062 / 133.8479.179 / 153.472
CLASP5.801 / 91.2916.350 / 104.6966.936 / 120.552

Presentation revisions.

Clarified modularity and related work in Section 2 and Appendix F.1; added CVAE, CAMul, and CCM comparisons in Tables 1–2 and Appendix C.1; documented integrations in Appendix C.2 and their results in Table 24, Appendix E.3.2, complementing the backbone evaluation in Section 4.3.

AuthorsOpenReview ↗

W3. Theory scope and practical relevance

W3. Theory scope and practical relevance

Theorem 1 does not assume independent time-series windows. Instead, it holds for any fixed observed context and requires independence only across latent component draws. We agree that Gaussian KL and data processing are standard tools; we use them specifically to characterize what CLASP’s pooling objective controls. Under one-hot assignments, the uncapped pooling penalty matches the aggregate training–deployment KL exactly; under soft assignments, however, neither quantity bounds the other in general, motivating us to measure aggregate KL directly rather than using the pooling penalty as a proxy bound.

We appreciate the concern that the overall length of the manuscript can make navigation challenging. This length primarily reflects our effort to document all experimental details needed to reproduce and assess our results across several forecasting backbones and baselines. Specifically, Appendices B–D detail how CLASP is integrated into each model alongside training configurations, sampling procedures, preprocessing, and evaluation protocols, while Appendix E reports comprehensive results and ablations across settings. Meanwhile, the core theoretical presentation is self-contained in Section 3.4, supported by complete proofs and regularity conditions in Appendix A, and the submission already includes an appendix table of contents and section roadmaps to guide readers.

AuthorsOpenReview ↗

Q1. Long histories and variable-length entities

Q1. Long histories and variable-length entities

eICU already includes patient records of varying lengths, with an overall range of 25–28,117 hourly steps. We added detailed length statistics to Table 15 in Appendix D.3.2.

Table R5 reports MAE / MSE on WEATHER-5K with 48-hour and 168-hour inputs at a fixed 168-hour forecast horizon. CLASP continues to improve the flow-matching backbone with longer inputs. Flow + CLASP uses five provided zones and the validation-fitted readout.

Table R5.

Input historyPlain FlowFlow + CLASP
48 hours, default2.710 / 21.0352.290 / 17.193
168 hours2.513 / 19.1481.902 / 11.848

Presentation revisions.

Added eICU length statistics in Table 15, Appendix D.3.2, with the fixed-input protocol clarified in Appendix D.3.1.

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

This paper proposes CLASP, a latent partial-pooling module for entity-level time series forecasting. It uses Gaussian prototypes to share information between entities. Experiments on WEATHER-5K and eICU show that CLASP performs better than existing baselines.

Critical strengths
  1. CLASP uses distributions instead of fixed cluster labels, which helps capture cluster uncertainty and share information between entities. The prototype priors can also be used for prediction without retraining.

  2. CLASP can be added to existing forecasting models without changing their architectures. Experiments show consistent MAE improvements across different backbones and forecasters.

Critical weaknesses
  1. The improvement over simple label conditioning is not clear. One-hot zone labels already achieve similar results to CLASP. The authors should explain why the learned structure is needed if the metadata already provides similar information.

  2. CLASP uses an additional persistence and climatology readout fitted on the validation set, while the baselines do not. This makes the comparison less fair, and it is hard to know how much improvement comes from CLASP itself.

  3. The learned representation seems to mainly capture geographical information, as the module does not explicitly model temporal patterns. The authors should further explain how this representation helps time series forecasting.

Questions for authors

See weaknessess.

3 comments
AuthorsOpenReview ↗

W1. Benefit beyond direct label conditioning

We thank the reviewer for these comments. The current manuscript already includes the controls and methodological details relevant to all three points, with several located in the Appendix. Below, we summarize this evidence and explain how we will present it more clearly in the main text. Manuscript references use the revised numbering.

W1. Benefit beyond direct label conditioning

Pooling with provided labels.

CLASP is designed to learn how to pool information, whereas learning the assignments themselves is optional. When zone labels are provided, they select shared Gaussian prototype priors, and KL alignment encourages history-dependent posteriors to share statistical strength through these priors. To isolate this mechanism from direct label conditioning, the control experiments in Appendix E.2.1 (Tables 20–21) retain the exact same history encoder, forecasting backbone, and readout protocol while replacing the prototype latent mechanism with direct label vectors.

Table R1. WEATHER-5K MAE / MSE, reproduced from manuscript Table 20.

VariantH=24H{=}24H=72H{=}72H=168H{=}168
One-hot zone conditioning1.355 / 5.7021.931 / 12.1842.345 / 17.514
CLASP, provided zones, K=5K{=}51.291 / 5.1111.902 / 11.8482.290 / 17.193
CLASP, inferred, K=4K{=}41.280 / 5.0491.894 / 11.6912.298 / 17.040

With the same five zone labels, CLASP reduces MAE by 4.7/1.5/2.3%4.7/1.5/2.3\% and MSE by 10.4/2.8/1.8%10.4/2.8/1.8\% at H=24/72/168H{=}24/72/168 (Table R1). These incremental gains are strongest at the shortest horizon and modest at longer horizons. Similar provided/inferred performance does not establish that learned assignments are necessary when useful labels exist.

Overlapping diagnoses in eICU.

The four diagnosis categories overlap within patients (Appendix D.3.2, Figure 18), so a single one-hot diagnosis label would discard comorbidities. Our direct-conditioning control therefore uses all four multi-hot diagnosis indicators; CLASP retains these as metadata and infers responsibilities from history and metadata. Relative to this control, CLASP reduces MAE by 4.6/3.5/1.7%4.6/3.5/1.7\% and MSE by 7.3/5.1/4.9%7.3/5.1/4.9\% at H=6/12/24H{=}6/12/24 (manuscript Table 21).

AuthorsOpenReview ↗

W2. Does the readout explain the gains?

W2. Does the readout explain the gains?

Manuscript Table 3 and Appendix E.2.1 already report readout ablations. Table R2 shows raw CLASP outperforming every WEATHER-5K baseline across metrics and horizons. Manuscript Tables 2 and 21 also show 13.4/9.5/3.3%13.4/9.5/3.3\% lower eICU MAE than NsDiff without readout.

Table R2. WEATHER-5K MAE / MSE, summarized from manuscript Tables 1 and 3.

H=24H{=}24H=72H{=}72H=168H{=}168
Best baseline per metric1.427 / 5.8272.036 / 13.4002.426 / 18.948
Raw CLASP, no readout1.294 / 5.1811.935 / 11.8992.371 / 17.536
Relative reduction9.3% / 11.1%5.0% / 11.2%2.3% / 7.5%

Existing readout controls.

Table 3 shows that readout reduces MAE by 1.1/2.1/3.1%1.1/2.1/3.1\% for the plain backbone and CLASP, using global and prototype-specific profiles, respectively. CLASP’s 12.5–17.7% advantage persists with or without readout. WEATHER-5K uses persistence and diurnal-climatology anchors from history, with profiles fitted on validation entities and fixed for testing. On eICU at H=6H{=}6, readout increases MAE from 5.7815.781 to 5.8015.801.

Existing probabilistic evaluation.

Table 5 in Section 4.4 reports CRPS and coverage from raw predictive draws. CLASP has the lowest CRPS at every horizon: 6.4–6.8% below TSFlow on WEATHER-5K and 7.3–8.5% below NsDiff on eICU.

Presentation revisions.

We will emphasize the existing matched comparison with and without readout in Table 3.

AuthorsOpenReview ↗

W3. Entity-level priors and temporal dependence

W3. Entity-level priors and temporal dependence

As detailed in Eqs. 1–2 and Appendix B.1, CLASP supplies an entity-level prior through shared prototypes, while the backbone models within-entity temporal dependence via historical observations and encoder states MeM_e. Specifically, partial pooling conditions forecasts on information shared across related entities. During training, posteriors condition on history, whereas at deployment, the model samples from shared prototype priors (using window-dependent routing when assignments are inferred). Consequently, strong within-station variation is not required for an entity-level prior to be informative.

Existing evidence for useful priors.

Geographic structure is compatible with this role. On WEATHER-5K at H=24H{=}24, manuscript Table 20 reports MAE of 1.3981.398 without metadata versus 1.5551.555 for the plain backbone. Existing eICU results show 8.4–13.1% lower MAE than NsDiff without geographic inputs.

Presentation revisions.

We will clarify this division of labor at the start of Sec. 3.

Reviewer_oEHVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces CLASP, a probabilistic framework for time-series forecasting that shares information across related entities while preserving entity-specific variation. The method learns shared Gaussian prototypes and aligns history-dependent latent distributions with them through KL regularization. These representations are then used to condition different forecasting architectures. The authors also provide theoretical results that attempt to characterize the relationship between training and deployment latent distributions.

Critical strengths

S1. The paper proposes an interesting probabilistic partial-pooling mechanism that combines shared Gaussian prototypes with history-dependent latent distributions. The KL-based alignment provides a principled way to share information across related entities while allowing entity-specific variation, addressing an important challenge in time-series forecasting.

Critical weaknesses

W1. The methodological novelty is limited relative to existing work on cluster-based and probabilistic latent representations for time-series forecasting. Prior studies have explored shared latent clusters [Ref1], probabilistic latent aggregation [Ref2], cluster-conditioned forecasting [Ref3], and hierarchical prototypes for shared temporal patterns [Ref4]. CLASP introduces Gaussian prototype priors with KL alignment, a mechanism closely related to mixture-prior VAEs [Ref5], but its distinction from these established approaches is not sufficiently clear. Moreover, the experiments do not compare against representative methods of this kind. Notably, the paper cites CCM [Ref3] as closely related but does not include it as a baseline. The evaluated baselines are general-purpose forecasters that do not model population structure, which makes it difficult to assess the added value of the proposed mechanism.

W2. The empirical evaluation is limited to two datasets, WEATHER-5K and eICU, leaving the generalizability of the proposed approach not well established, as widely used time-series forecasting benchmarks, such as ETT, Electricity, and Traffic [Ref6, Ref7], are absent. The WEATHER-5K experiments use only 125 stations, with just 15 held out for testing, which further limits the strength of the empirical conclusions.

W3. The probabilistic evaluation is too limited for a framework whose main mechanism is probabilistic. CRPS, energy score, and P10-P90 coverage are reported only against NsDiff and TSFlow (NsDiff alone on eICU), leaving out the other probabilistic baselines (TSDiff, D3VAE). More importantly, all ablations are evaluated with point metrics only. A more complete evaluation would report CRPS for the ablations, include the remaining probabilistic baselines, and perhaps even add more calibration diagnostics.

[Ref1] Adhikari et al. EpiDeep: Exploiting Embeddings for Epidemic Forecasting. KDD 2019.

[Ref2] Kamarthi et al. CAMul: Calibrated and Accurate Multi-view Time-Series Forecasting. WWW 2022.

[Ref3] Chen et al. From Similarity to Superiority: Channel Clustering for Time Series Forecasting. NeurIPS 2024.

[Ref4] Peng et al. ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting. ICLR 2026.

[Ref5] Jiang et al. Variational Deep Embedding: An Unsupervised and Generative Approach to Clustering. IJCAI 2017.

[Ref6] Wu et al. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting. NeurIPS 2021.

[Ref7] Chen et al. Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting. ICLR 2024.

Questions for authors

Please address the points made in the weakness section.

3 comments
AuthorsOpenReview ↗

W1. Methodological distinction and related baselines

We thank the reviewer for highlighting these connections and missing comparisons. All manuscript references follow the revised paper numbering.

W1. Methodological distinction and related baselines

CLASP’s contribution.

CLASP is an adaptive module for entity-level latent partial pooling that augments existing deterministic and probabilistic forecasters. It aligns history-dependent entity distributions with shared Gaussian prototype priors and conditions the host model on their aggregate latent. Gaussian priors and KL alignment are established; our distinction is independent-draw Gaussian aggregation, rather than categorical mixture selection, and analysis of the training–deployment discrepancy. Unlike CCM’s cluster-specific heads and CAMul’s view-specific graphs, CLASP pools entity-level latent distributions. Adding CLASP improves MAE at every WEATHER-5K horizon for D3VAE, TSDiff, NsDiff, Timer-XL, TSFlow, and CAMul (Table 24, Appendix E.3.2). We added CAMul and CCM. For CCM, we use PatchTST, a representative backbone evaluated in its original paper. EpiDeep’s univariate seasonal setup needs adaptation. VaDE groups similar observations; it would need an additional model to predict future values. ProtoTS reports multivariate results, but its released load pipeline assumes a scalar target, a fixed 23-feature schema including future covariates, and quarter-hour phase alignment. Applying it to our six-channel hourly tasks requires changing feature/target interfaces, alignment, and the treatment of unavailable future covariates; we therefore exclude a custom adaptation.

Table R1 reports MAE / MSE (best in bold), using five provided WEATHER-5K zones and four inferred eICU groups for CLASP.

Table R1

WEATHER-5K

Model24 hours72 hours168 hours
CAMul1.595 / 8.0192.049 / 15.2452.561 / 20.099
CCM1.568 / 5.8272.326 / 16.7662.664 / 23.866
CLASP1.291 / 5.1111.902 / 11.8482.290 / 17.193

eICU

Model6 hours12 hours24 hours
CAMul6.984 / 103.1297.549 / 116.6078.102 / 133.299
CCM8.946 / 117.5019.062 / 133.8479.179 / 153.472
CLASP5.801 / 91.2916.350 / 104.6966.936 / 120.552

Evidence for the mechanism.

CLASP outperforms both added baselines across both metrics and all horizons. Before readout, it lowers WEATHER-5K MAE by 12.5–17.7% over the same plain backbone. Deterministic-prototype, single-prototype, and no-pooling controls support the construction; the last tests alignment and prototype-prior learning jointly.

Presentation revisions.

Clarified modularity and related work (Section 2; Appendix F.1); added CAMul and CCM baselines (Tables 1–2; Appendix C.1) and the CAMul retrofit (Appendices C.2 and E.3.2).

AuthorsOpenReview ↗

W2. Dataset scope and generalization across entities

W2. Dataset scope and generalization across entities

Why entity structure matters.

CLASP pools across entities and tests forecasting on unseen entities. Each station or patient supplies a six-channel time series, with entity identities separated across training, validation, and testing. ETT-small has only two physical transformers; hourly and 15-minute versions do not create additional entities. A collection of many distinct transformers, extending to ETTh3, ETTh4, and beyond, would support our entity-level evaluation. Electricity and Traffic are conventionally evaluated as multivariate forecasting benchmarks. Recasting each customer or sensor as an entity would create a collection of univariate series with entity-disjoint splits, changing the standard baseline comparison protocol.

Dataset scale and evaluation breadth.

WEATHER-5K contains 5,672 stations; our experiments use a balanced subset of 125, with 25 per zone. Our eICU benchmark contains 93,286 patients, including 7,923 held out. The original submission already included imbalanced WEATHER-5K results. Appendix E.3.6 reports cluster-wise forecasting under both balanced and imbalanced training configurations across all five zones and three forecast horizons. Further evaluations, including per-variable accuracy, component ablations, and comparisons across forecasting backbones, appear in Appendices E.1–E.3; eICU’s naturally variable-length records and preprocessing are described in Appendix D.3.1.

A more challenging cross-zone transfer test.

To address the reviewer’s concern about the limited WEATHER-5K holdout, we further test distribution shift to an unseen geographic zone: each of five folds excludes one zone from task-specific training and evaluates on that zone at 24 hours. This goes beyond the headline split, where training and test stations share the same five zones. Table R2 reports MAE / MSE for each held-out zone, with the best values in bold; Timer-XL ZS/FT denotes zero-shot/fine-tuned evaluation. CLASP achieves the lowest MAE in all five zones and the lowest MSE in four; TSFlow has lower Tropical MSE. In the Polar zone, CLASP reduces MAE/MSE by 14.3%/20.4% relative to the strongest baseline, NsDiff.

Table R2

ModelTropicalDryTemperateContinentalPolar
D3VAE2.906 / 20.9322.700 / 19.7842.504 / 17.7042.788 / 21.4953.445 / 33.925
TSDiff1.429 / 4.7331.585 / 6.8411.717 / 7.8941.957 / 11.8632.801 / 27.359
Timer-XL ZS0.917 / 2.5021.441 / 5.2521.552 / 7.3911.694 / 9.3621.857 / 11.157
Timer-XL FT1.134 / 3.7321.863 / 9.3612.109 / 12.2162.065 / 12.2171.873 / 10.663
NsDiff1.071 / 3.0431.555 / 6.5101.821 / 8.9331.772 / 9.1731.633 / 7.842
TSFlow0.890 / 2.3401.436 / 5.5351.520 / 6.3711.571 / 7.7131.713 / 9.403
CAMul1.070 / 3.2291.412 / 5.4641.595 / 6.6801.587 / 7.4051.831 / 11.172
CCM1.149 / 3.6941.809 / 8.5762.066 / 11.2601.897 / 10.4412.027 / 12.422
CLASP0.855 / 2.4311.322 / 5.0641.493 / 6.1731.521 / 7.0001.399 / 6.246

Presentation revisions.

Added leave-one-zone-out WEATHER-5K evaluations (Section 4.1 and Appendices E.3.7), with the corresponding transfer discussion in Appendix F.2.

AuthorsOpenReview ↗

W3. Probabilistic comparisons and ablations

W3. Probabilistic comparisons and ablations

Expanded probabilistic comparisons.

We added TSDiff and CAMul probabilistic comparisons on both datasets. Table R3 reports CRPS / P10–P90 coverage / energy score from raw predictive draws before readout. Lower CRPS and energy score are better (best in bold); coverage is compared with 80%. CLASP has the lowest CRPS and energy score throughout, but coverage is not uniformly closest to 80%. D3VAE and CVAE use one forecast per window (S=1S=1) under our evaluated protocols and are excluded from this ensemble comparison. TSFlow’s evaluated prior construction requires horizons divisible by 24 hours, which excludes eICU’s 6- and 12-hour horizons; we therefore report TSFlow only on WEATHER-5K.

Table R3

WEATHER-5K

Model24 hours72 hours168 hours
NsDiff1.085 / 0.787 / 1.8421.583 / 0.745 / 2.7131.912 / 0.714 / 3.295
TSFlow1.052 / 0.749 / 1.7931.547 / 0.793 / 2.6511.881 / 0.741 / 3.218
TSDiff1.118 / 0.762 / 1.9211.624 / 0.731 / 2.8041.948 / 0.702 / 3.388
CAMul1.097 / 0.758 / 1.8791.591 / 0.736 / 2.7421.902 / 0.709 / 3.326
CLASP0.981 / 0.796 / 1.6741.442 / 0.789 / 2.4931.761 / 0.776 / 3.038

eICU

Model6 hours12 hours24 hours
NsDiff4.823 / 0.833 / 38.4515.412 / 0.796 / 43.2185.891 / 0.780 / 47.583
TSDiff5.387 / 0.812 / 41.2366.024 / 0.785 / 46.0826.541 / 0.769 / 50.417
CAMul4.896 / 0.819 / 39.1245.504 / 0.788 / 43.8725.974 / 0.772 / 48.036
CLASP4.412 / 0.781 / 35.0124.987 / 0.772 / 39.4785.463 / 0.765 / 43.521

CRPS for component ablations.

Table R4 adds CRPS for component ablations on both datasets (lower is better). Readout variants reuse the same raw draws and therefore have identical CRPS.

The eICU provided-cluster and 50%-training-label settings are marked N/A: its four diagnosis indicators overlap and serve as metadata, while cluster assignments are inferred. There are no provided cluster labels to withhold; forcing mutually exclusive diagnosis groups would discard this overlap. Dashes denote unreported results.

Table R4

WEATHER-5K

Variant24 h72 h168 h
No pooling (prior untrained)1.3181.8462.147
Deterministic prototypes1.0491.5411.864
Single prototype1.0721.5181.859
Direct label-vector conditioning1.0981.5561.904
50% training labels1.0261.4891.822
Without latent conditioning1.1581.6622.002
Without encoder states1.0761.5721.881
Without metadata1.1021.6211.949
CLASP (provided groups)0.9811.4421.761
Raw CLASP (no readout)0.9781.4391.763
Full CLASP (inferred groups)0.9781.4391.763

eICU

Variant6 h12 h24 h
No pooling (prior untrained)6.1496.7917.394
Deterministic prototypes4.8495.3725.928
Single prototype4.7425.3346.079
Direct label-vector conditioning5.0215.4895.921
50% training labelsN/AN/AN/A
Without latent conditioning5.2715.8066.243
Without encoder states6.1046.1916.854
Without metadata5.2425.5415.968
CLASP (provided groups)N/AN/AN/A
Raw CLASP (no readout)4.4124.9875.463
Full CLASP (inferred groups)4.4124.9875.463

Presentation revisions.

Added TSDiff/CAMul probabilistic comparisons (Section 4.4, Table 5) and CRPS for component ablations (Section 4.2, Table 3; Appendix E.2, Tables 20–21). Moved energy-score reporting from the appendix into main-text Table 5 and removed the redundant “Predictive Distribution Quality” appendix subsection, table, and discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Summary

We revised the manuscript to clarify CLASP’s contribution, broaden the empirical evaluation, and improve methodological transparency. The main revisions are summarized below; added and revised manuscript text and results are highlighted in red.

  • Method and related work. Clarified CLASP’s entity-level partial pooling, its relationship to clustering and variational methods, and the roles of prototype priors and temporal backbones. Explained independent-draw aggregation and its connection to the pooling objective.

  • Additional baselines. Added CVAE, CAMul, and CCM comparisons on both datasets, using PatchTST for CCM. Aggregate results appear in Tables 1–2 and per-variable results in Tables 16–17.

  • Backbone integration. Added CVAE and CAMul retrofits in Table 24, Appendix E.3.2, extending WEATHER-5K evaluation to seven external forecasters. These comparisons include latent conditioning and the validation-fitted readout.

  • Cross-zone transfer. Added five-fold leave-one-zone-out WEATHER-5K evaluation at 24 hours to test forecasting in unseen geographic zones, reported in Table 30, Appendix E.3.7.

  • Cross-hospital transfer. Added hospital-disjoint eICU evaluation at 6, 12, and 24 hours, excluding patients recorded at multiple hospitals. Tables 32–33 in Appendix E.3.9 report the split and results.

  • Variable-length patient histories. Added split-wise eICU record-length ranges, medians, middle 50% intervals, and 95th/99th percentiles in Table 15, Appendix D.3.2. Clarified in Appendix D.3.1 that these describe patient records, while forecasting uses fixed 12-hour inputs within individual ICU stays.

  • Probabilistic evaluation. Added TSDiff/CAMul comparisons and energy scores alongside CRPS and P10–P90 coverage in Table 5, using raw predictive draws. D3VAE/CVAE remain excluded under their evaluated single-forecast protocols.

  • Ablations and readout. Added raw-draw CRPS to component ablations in Tables 3, 20, and 21 and highlighted existing readout and direct-label controls. Readout variants reuse identical draws for CRPS.

  • Reproducibility and presentation. Expanded baseline configurations, sampling protocols, and retrofit documentation; consolidated probabilistic results and removed the duplicate appendix comparison.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.