acceptodds
Under review as a conference paper at ICLR 2027

Conditional Projective Scenario Systems for Long-Horizon Joint Forecasting

Abstract

Modern time-series foundation models (TSFMs) can estimate value distributions at individual future time steps, but these marginals do not specify which values are likely to occur together. This matters for quantities such as the future maximum, maximum one-step change, or time spent above a threshold, which depend on the complete forecast trajectory. In this study, we introduce the Conditional Projective Scenario System (CPSS), which transforms marginal forecasts from a frozen TSFM (TimesFM-2.5) into 64 weighted complete trajectories. CPSS first represents broad future alternatives and then introduces finer variations while preserving consistency across forecast horizons, so identical truncated prefixes are merged and their probabilities summed. Across eight datasets, CPSS achieves increasing gains over a closely matched 64-trajectory baseline as the forecast horizon grows. On a separate 12-domain, 128-step benchmark, CPSS reduces complete-trajectory error by 6.80% and error on path-dependent quantities by 16.80%. Under present evidence, these results demonstrate that CPSS can extend strong marginal forecasters into coherent scenario generators, showing great promise for more accurate probabilistic reasoning over complete future trajectories.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, probabilistic forecasting, joint forecasting, trajectory forecasting, scenario generation, time-series foundation models, long-horizon forecasting

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

This paper proposes CPSS, which learns a distribution over 64 weighted trajectories around a frozen TimesFM-2.5 model. Hierarchical, suffix-supported refinements allow related scenarios to share early prefixes, which merge with their probability masses when truncated. Training combines the energy score, marginal CRPS, and entropy regularization. The results show improvements at longer horizons and on path-dependent quantities, but the attribution of these gains and the reliability of the evaluation require further clarification

Critical strengths

**Practically relevant motivation. **The paper clearly distinguishes marginal forecasting from joint trajectory forecasting and evaluates meaningful path-dependent quantities, including maxima, ramps, and exceedance duration.

Clear structural design. Suffix-supported refinements and hierarchical probability allocation provide an interpretable branching representation with exact prefix merging under truncation.

Useful controlled evaluation. CPSS and GB64 share the frozen carrier, scenario budget, and similar parameter counts. The paper also examines path-dependent scores, calibration, and scenario diversity.

Critical weaknesses

The main ablation conflates hierarchy with basis design. CPSS uses local, suffix-supported basis functions, whereas GB64 uses only eight global cosine components per trajectory. These representations differ in their ability to capture local variations, particularly at long horizons. Consequently, the observed gains cannot be attributed solely to hierarchical branching. A flat model using comparable local bases and an ablation removing suffix constraints would strengthen the evidence.

The cross-horizon consistency guarantee has a limited scope. The construction guarantees prefix merging when truncating a single terminal forecast distribution. However, the heads for H=192,336,720H=192,336,720 are trained separately, so consistency between these independently trained forecasts is not established. The paper should distinguish shared-prefix structure from consistency across horizon-specific models and demonstrate the practical value of the former.

Small evaluation panels and unclear data separation limit confidence. The horizon sweep uses only eight forecast origins per dataset, while the 12-domain joint evaluation contains 96 cases. Case identifiers are disjoint, but windows may overlap, and both training and part of the evaluation draw from GiftEvalPretrain. The authors should clarify separation by series and time interval. Case-level bootstrapping may also underestimate uncertainty when windows or channels are correlated; grouped or temporal-block resampling would be more appropriate.

Evidence for improved dependence modeling is mixed. Although CPSS improves energy and path-functional CRPS relative to TimeMCL, its variogram score worsens by approximately 5.78%, with an interval excluding zero. This result deserves discussion. Moreover, the long-horizon conclusions rely primarily on GB64 rather than a broader set of competitive joint forecasting methods.

Sensitivity analyses and implementation details are incomplete. The scenario budget, branching locations, basis size, and entropy threshold are fixed without corresponding ablations. More detail on training-data sampling and exclusion rules, the TimeMCL adaptation, and computational costs is needed to assess reproducibility and generalizability.

Questions for authors

see weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer gyNs (1/3): Q1, W1–W2

Q1

see weakness

Matched tests identify predictive gains from nested sharing, and exact prefix aggregation reduces measured scoring costs. The five responses below give the supporting experiments and implementation details.

The reviewer refers to the five weaknesses. We quote and address each below.

W1

The main ablation conflates hierarchy with basis design. CPSS uses local, suffix-supported basis functions, whereas GB64 uses only eight global cosine components per trajectory. These representations differ in their ability to capture local variations, particularly at long horizons. Consequently, the observed gains cannot be attributed solely to hierarchical branching. A flat model using comparable local bases and an ablation removing suffix constraints would strengthen the evidence.

Matched ancestry isolates gains in Energy, variogram and maximum CRPS with local bases and capacity fixed. CPSS also represents H336 forecasts with at most 16 distinct prefixes, versus 64 for the flat local-basis control.

  1. The sharing contribution is isolated with local geometry held fixed. The matched ancestry intervention preserves the dictionary, capacity, initialisation, terminal-logit probability map and training batches, with zero entropy penalty in both arms.

    Nested sharing versus matched rewiringSkill, points95% interval
    Energy+0.411[0.222, 0.568]
    Variogram+0.106[0.013, 0.257]
    Maximum CRPS+2.168[0.863, 3.496]

    The Energy estimates favour nesting in every dataset and seed. This answers whether ancestry contributes beyond choosing local basis functions.

  2. Both requested additional controls were evaluated. The 638,464-parameter flat local-basis system uses the same suffix dictionary but a different coefficient factorisation and masses. It gains 2.069 [1.136, 2.892] Energy points over CPSS. Relaxed suffix support gains +2.793 [1.520, 3.921] Energy points, while exact-support CPSS gains +12.233 [4.569, 20.393] maximum-ramp CRPS points over that relaxation.

  3. The interventions answer distinct questions. Matched rewiring isolates the sharing effect. The flat-system and suffix comparisons measure predictive flexibility alongside exact prefix structure. At H336 the flat control retains 64 distinct prefixes, while CPSS has at most 16 and supports lossless aggregation. Section 5.2, Table D.2 in Appendix D.1 and Appendix D.4 give these results and their protocols.

W2

The cross-horizon consistency guarantee has a limited scope. The construction guarantees prefix merging when truncating a single terminal forecast distribution. However, the heads for H = 192, 336, 720 are trained separately, so consistency between these independently trained forecasts is not established. The paper should distinguish shared-prefix structure from consistency across horizon-specific models and demonstrate the practical value of the former.

Exact prefix aggregation preserves one forecast's restricted distribution and prefix-dependent objectives. At H336, it reduces 64 records to 16 and lowers measured scoring memory under both evaluators, with generation costs reported separately.

  1. The guarantee is exact shared-prefix consistency within one terminal forecast. Related scenarios are identical before their prescribed divergence, and descendant masses sum under restriction. Separately trained heads are separate forecast laws. Section 3.3 states this scope explicitly.

  2. Exact grouping preserves the predictive object and its prefix-dependent uses. At H336, grouping the 64 terminal records into 16 prescribed prefix records leaves the restricted forecast distribution and every prefix-dependent expected-cost objective unchanged.

  3. The compact law reduces measured scoring cost. These are measurements of the same issued CPSS forecasts at batch size 64.

    H336 scoring operation64 unmerged records16 prefix records
    Original evaluator time1.733 ms0.974 ms
    Original evaluator incremental peak353.371 MB22.086 MB
    Lower-memory evaluator incremental peak44.46 MB11.43 MB

    The original scoring operation is also 2.76 times as fast after H540 aggregation. Table D.7 includes H720 without prefix reduction. The score-only measurements exclude generation and merge preparation. Appendix D.4 gives the regrouping identity, equality checks and both evaluators. W5 below reports generation and combined-operation costs separately.

AuthorsOpenReview ↗

Response to Reviewer gyNs (2/3): W3–W4

W3

Small evaluation panels and unclear data separation limit confidence. The horizon sweep uses only eight forecast origins per dataset, while the 12-domain joint evaluation contains 96 cases. Case identifiers are disjoint, but windows may overlap, and both training and part of the evaluation draw from GiftEvalPretrain. The authors should clarify separation by series and time interval. Case-level bootstrapping may also underestimate uncertainty when windows or channels are correlated; grouped or temporal-block resampling would be more appropriate.

H128 scenario-head training excludes all evaluation domains, and the 569 new cases exclude recorded window overlap. Grouped resampling keeps related observations together, with the H720 origin-spacing study reported separately.

  1. Training separation is verified for the H128 scenario heads. Their 2,368-record, 37-domain bank excludes all 12 H128 evaluation domains. H720 uses a separate 36-domain training bank with frozen schedules, and its run receipts record no benchmark test-target access. These checks concern the scenario heads, with carrier-pretraining overlap and cross-domain aliases remaining unverified.

  2. The new H128 cases remove recorded full-window overlap. The 569-case evaluation excludes complete windows intersecting either earlier evaluation and enforces nonoverlap within source items, regardless of channel. Cases were selected before forecasting using fixed source hashes. The original evaluations retain their reported overlapping windows, with these populations distinguished in Appendix D.2.

  3. Uncertainty analyses preserve relevant groups. Dataset resampling keeps each dataset's paired comparison together, the five-family sensitivity groups the four ETT datasets, temporal blocks retain all channels at an origin, and H128 item resampling keeps a source series' windows together.

    Grouped resultCPSS advantage95% interval
    Original H720 Energy against GB64, dataset resampling+8.518 points[4.631, 11.744]
    Matched sharing Energy, ETT treated as one family+0.321 points[0.093, 0.548]
    New H128 functional composite against GB64, domain-and-item resampling+3.921%[0.975, 21.106]

    The supplementary H720 origin-spacing study is identified separately because 55 of its 67 origins overlap earlier target intervals. Nonoverlapping windows are not automatically independent observations. Appendix D.2 gives the separation checks and grouped protocols.

W4

Evidence for improved dependence modeling is mixed. Although CPSS improves energy and path-functional CRPS relative to TimeMCL, its variogram score worsens by approximately 5.78%, with an interval excluding zero. This result deserves discussion. Moreover, the long-horizon conclusions rely primarily on GB64 rather than a broader set of competitive joint forecasting methods.

CPSS improves Energy and the functional CRPS composite over TimeMCL in the 96-case evaluation, with the composite gain also present across 569 cases. Four additional H720 interfaces broaden the comparison.

  1. CPSS improves complete-path and functional scores in the original H128 comparison. Against TimeMCL, relative loss reductions are +6.69% [1.96, 9.67] for Energy and +11.18% [2.32, 19.14] for four-functional CRPS. The variogram reduction is -5.78% [-14.49, -0.14], favouring TimeMCL. Energy assesses complete trajectories, while variogram examines moments of lagged increments, so the measured trade-off concerns different properties of the forecast law.

  2. The expanded evaluation also favours CPSS on the functional composite. Across 569 cases, the reduction over TimeMCL is +2.926% [0.365, 19.352] under domain-and-item resampling. The gain is mainly associated with threshold duration. The constant-history Wind Power case remains included. Appendix D.3 reports both aggregation conventions, with the ratio-of-risks Energy interval against TimeMCL including zero and variogram favouring TimeMCL under mean within-domain skill.

  3. Long-horizon comparisons now include four additional joint forecasting interfaces. The following H720 comparisons use 64 paths on the same 1,480 context-512 cases, with positive Energy skill favouring CPSS.

    Evaluated H720 interfaceCPSS Energy skill, points95% interval
    TimeMCL carrier adapter-0.116[-0.903, 0.527]
    TACTiS-2 carrier adapter+0.137[-1.007, 1.101]
    TimePrism-64 adaptation+38.043[19.608, 58.671]
    Toto 1.0 native sampler+93.005[48.546, 136.061]

    The evaluated TimePrism and Toto interfaces favour CPSS, while the TimeMCL and TACTiS-2 intervals include zero. These system comparisons use differing training and forecasting interfaces, including mixed 1,000/6,000-update TimePrism checkpoints. They broaden the evaluation alongside the matched architectural tests. Section 5.3 and Appendix D.5 give the interfaces and results.

AuthorsOpenReview ↗

Response to Reviewer gyNs (3/3): W5

W5

Sensitivity analyses and implementation details are incomplete. The scenario budget, branching locations, basis size, and entropy threshold are fixed without corresponding ablations. More detail on training-data sampling and exclusion rules, the TimeMCL adaptation, and computational costs is needed to assess reproducibility and generalizability.

The 48-fit study covers the requested sensitivities, with training sampling and TimeMCL details specified below. Matched implementation tests reduce CPSS head time to 3.18 ms and show memory savings in combined generation and scoring.

  1. The requested sensitivities are evaluated in the 48-fit component study. At H720, the following estimates are relative to the 640,059-parameter reference, with positive Energy skill favouring the modified configuration.

    Modified configurationActive parametersEnergy skill, points95% interval
    32 scenarios492,315-1.976[-2.702, -1.065]
    128 scenarios935,547+1.743[0.971, 2.463]
    Two basis functions per node509,757-2.293[-3.225, -1.222]
    Eight basis functions per node900,663+2.111[1.099, 3.041]

    At fixed capacity, the original schedule gains +0.269 [0.105, 0.417] and +0.606 [0.315, 0.898] points over uniform and later refinement. Removing the entropy penalty or raising its floor from 0.35 to 0.70 leaves reference Energy performance stable, and disabling it preserves the relaxed-suffix comparison. Appendix D.1 provides all configurations.

  2. Training selection and exclusion rules are explicit. The H128 scenario-head bank has 2,368 records from 37 domains, excluding all 12 evaluation domains. Each 16-window training batch used eight domains and two sampled windows per domain, with a fixed hash-based domain schedule and seeded sampling of eligible records and forecast origins. Contexts cycled through 128, 256 and 512 for H128, and 96, 128, 256 and 512 for the horizon sweep. Fixed source hashes select evaluation cases. The new 569-case evaluation excludes recorded full-window overlap within source items and against both earlier evaluations. Appendix D.2 states the source, channel and temporal rules and the scope of the verified exclusions.

  3. The TimeMCL adaptation is specified. It uses the official conditional class at commit 666840a, 160 pooled history-and-carrier features, 64 uniformly weighted hypotheses and relaxed winner-takes-all training. The H128 adapter has 1,759,192 parameters, uses the original 37-domain batches, seeds 11 and 22, and 1,000 updates. AdamW has a 50-update warm-up and cosine decay from 0.0003 to 0.00003. Appendix D.5 gives the implementation.

  4. Computational measurements include both generation and prefix evaluation. Basis caching and vectorisation reduce CPSS head time from 67.32 ms to 3.18 ms in a new matched H720 test with frozen checkpoints, float32, an RTX A6000 and batch size 128. Cached GB64 and local-basis heads take 1.97 ms and 3.26 ms. All 1,480 cases and three seeds passed numerical checks.

    Streamed H720 head and H336 scoring, batch 64CPSSGB64Local control
    Median time4.02 ms3.14 ms4.00 ms
    Incremental peak allocation45.43 MB49.65 MB49.65 MB

    The combined operation includes the conditioner, prefix preparation, Energy and sorted marginal CRPS. It excludes frozen-carrier generation and transfers. Ten warm-ups and 50 synchronised measurements per seed produce the reported medians. These measurements show the memory benefit of prefix aggregation with streamed decoders and the lower-memory evaluator. Appendix D.4 and Table D.8 give every implementation stage and both evaluators.

Thank you for the additional experiments and clarifications. The response addresses most of my concerns. I will maintain my original score.

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

This paper introduces the Conditional Projective Scenario System (CPSS), which converts marginal forecasts from a frozen TimesFM-2.5 model into 64 weighted trajectories through hierarchical suffix-supported refinements. The model is optimized using energy score, marginal CRPS, and entropy regularization. CPSS shows increasing gains over GlobalBasis-64 (GB64) as the forecast horizon extends, reaching +8.52 energy skill points at H = 720. On a 12-domain benchmark, it reduces complete-path energy score and path-functional CRPS by 6.80% and 16.80%, respectively.

Critical strengths

S1. Well-motivated problem with appropriate evaluation. The paper addresses an important limitation of marginal forecasts in capturing temporal dependencies. The evaluation metrics are well aligned with the goal of joint trajectory forecasting.

S2. Fairly controlled comparisons and encouraging long-horizon performance. The GB64 comparison uses the same frozen carrier, scenario budget, and comparable parameter counts. Dataset-level results and confidence intervals further support the reported gains.

Critical weaknesses

W1. Limited ablations and insufficiently established novelty. The comparison with GB64 changes multiple components simultaneously, including the hierarchical structure, suffix-supported refinements, and probability allocation. Without targeted ablations, the sources of improvement cannot be clearly identified. The novelty of exact shared-prefix consistency beyond standard marginalization and existing scenario-tree methods also requires clarification.

W2. Insufficient robustness evaluation. The H = 128 joint evaluation includes only 96 cases across 12 domains, using selected forecast origins and channels. Since training and evaluation data both originate from GIFT-Eval Pretrain, the data separation protocol also needs clarification.

Questions for authors

Q1. Could the authors provide ablations isolating the contributions of the hierarchical shared-prefix structure, suffix-supported refinements, and tree-based probability allocation?

Q2. How were evaluation channels and forecast origins selected, and how was potential training–evaluation overlap ruled out?

Q3. How sensitive is CPSS to the scenario budget and predefined refinement locations?

Q4. Is exact shared-prefix consistency guaranteed across independently trained horizon-specific heads, or only when truncating predictions from a single head?

6 comments
AuthorsOpenReview ↗

Response to Reviewer fFTi (1/2): W1 / Q1

We address W1 with Q1 and W2 with Q2, quoting each in full before the shared response.

W1

Limited ablations and insufficiently established novelty. The comparison with GB64 changes multiple components simultaneously, including the hierarchical structure, suffix-supported refinements, and probability allocation. Without targeted ablations, the sources of improvement cannot be clearly identified. The novelty of exact shared-prefix consistency beyond standard marginalization and existing scenario-tree methods also requires clarification.

CPSS combines carrier-conditioned trajectory refinement with exact shared-prefix aggregation. Matched nested sharing improves prediction, while the compact restricted forecasts deliver measured scoring-memory savings without changing the predictive distribution.

Q1

Could the authors provide ablations isolating the contributions of the hierarchical shared-prefix structure, suffix-supported refinements, and tree-based probability allocation?

Matched sharing improves Energy by +0.411 points with local bases and capacity fixed. Separate tests identify the +12.233-point maximum-ramp benefit of exact suffix support and measure the effect of probability allocation.

  1. Nested sharing has a controlled predictive contribution beyond local basis geometry. The nested and rewired models share the local dictionary, coefficient count, conditioner, 640,059 active parameters, initial tensors, terminal-logit map and training batches, with zero entropy penalty. Rewiring changes ancestry assignments while preserving node usage counts.

    InterventionMetricCPSS skill, points95% interval
    Matched ancestryEnergy+0.411[0.222, 0.568]
    Matched ancestryVariogram+0.106[0.013, 0.257]
    Matched ancestryMaximum CRPS+2.168[0.863, 3.496]
    Exact suffix supportMaximum-ramp CRPS+12.233[4.569, 20.393]
    Exact suffix supportEnergy-2.793[-3.921, -1.520]

    Positive values favour nested ancestry or exact support in their respective interventions. The matched Energy effect is positive in every dataset and seed, and remains +0.321 [0.093, 0.548] when ETT is grouped as one family.

  2. Suffix support and probability allocation are tested separately. Exact support preserves prescribed prefix identities and benefits maximum-ramp forecasting in the tested comparison, while relaxation improves Energy. Replacing branch probabilities with terminal logits gains +0.070 [0.004, 0.136] Energy points. The separate flat local-basis system gains 2.069 [1.136, 2.892] Energy points and retains 64 distinct H336 prefixes versus at most 16 for CPSS.

  3. The contribution combines learned refinement with an exact, usable forecast structure. CPSS is a carrier-conditioned finite scenario process with shared-prefix identities and mass-preserving horizon restriction. Our contribution is not ordinary marginalisation or the scenario-tree abstraction itself, but the combination of carrier-conditioned learned refinement, a measured predictive contribution from nested sharing, and exact prefix aggregation that preserves shorter-horizon forecast laws. Controlled sharing improves prediction, while exact aggregation preserves prefix-dependent objectives and reduces evaluation memory. Sections 2.2, 3.3 and 5.2, Table D.2 in Appendix D.1, and Appendix D.4 state and evaluate these contributions.

AuthorsOpenReview ↗

Response to Reviewer fFTi (2/2): W2 / Q2–Q4

W2

Insufficient robustness evaluation. The H = 128 joint evaluation includes only 96 cases across 12 domains, using selected forecast origins and channels. Since training and evaluation data both originate from GIFT-Eval Pretrain, the data separation protocol also needs clarification.

The additional 569-case H128 evaluation supports the functional CRPS composite gain under grouped resampling. All 12 evaluation domains are excluded from scenario-head training, and the new cases exclude recorded full-window overlap.

Q2

How were evaluation channels and forecast origins selected, and how was potential training–evaluation overlap ruled out?

Scenario-head training excludes all 12 H128 evaluation domains. Fixed source hashes select cases before forecasting, and the 569 new cases exclude recorded full-window overlap within source items and against both earlier evaluations.

  1. The scenario heads were fitted with all H128 evaluation domains excluded. The original bank contains 2,368 records from 37 domains. Its inventory hash matches the original CPSS and GB64 run records, and it contains no matching evaluation domain-and-item identifiers.

  2. Case selection used fixed rules before scoring. The original joint evaluation used eligible channels from the first 128 source records per domain, contexts 128/256/512 and 33 regularly spaced origins. A fixed source-identifier hash order selected eight cases per domain after excluding exact tuples from the marginal evaluation. Eligibility required at least 90% finite history and a fully finite target. This removed duplicate cases while permitting overlapping windows.

  3. The expanded cases strengthen temporal separation. Candidate origins 128, 384, 640 and so on define disjoint 256-observation history-plus-target windows. We excluded complete windows intersecting either earlier evaluation regardless of channel, kept at most one channel per item and origin, and selected by source hashes before forecasting.

    H128 separation checkVerified result
    Evaluation domains excluded from head-training bankAll 12
    Additional cases569, comprising 11 times 48 plus 41 from NN5
    Recorded complete-window overlaps within source items or against earlier evaluations0

    Appendix D.2 gives the rules and checks. These verify scenario-head separation. Frozen-carrier pretraining overlap and cross-domain aliases remain unverified.

Q3

How sensitive is CPSS to the scenario budget and predefined refinement locations?

The budget study quantifies the accuracy–capacity choice across 32, 64 and 128 scenarios. The original refinement schedule outperforms both tested alternatives at fixed capacity, with parameter counts reported for budget changes.

  1. The budget study quantifies the quality-capacity relationship. The 32-, 64- and 128-scenario models use 492,315, 640,059 and 935,547 active parameters, respectively.

    Configuration relative to the 64-scenario referenceEnergy skill, points95% interval
    32 scenarios-1.976[-2.702, -1.065]
    128 scenarios+1.743[0.971, 2.463]
  2. The original refinement schedule gives the best result among the tested schedules. Its advantage over uniform refinement is +0.269 [0.105, 0.417] points, and over later refinement is +0.606 [0.315, 0.898] points. These schedule comparisons keep capacity fixed.

  3. The scenario budget offers a measurable accuracy–capacity choice. Larger budgets improve Energy with more parameters. Appendix D.1 reports the scenario, basis-size, schedule and entropy studies with their parameter counts and protocols.

Q4

Is exact shared-prefix consistency guaranteed across independently trained horizon-specific heads, or only when truncating predictions from a single head?

Exact consistency holds across restrictions of one terminal forecast. At H336, lossless aggregation reduces prefix records from 64 to 16, and the horizon advantage over GB64 persists with the trained head fixed.

  1. Exact consistency is guaranteed across restrictions of a single terminal forecast. Independently trained horizon-specific heads are not constrained to agree. Section 3.3 now states the guarantee explicitly.

  2. That guarantee provides compact evaluation across horizons. One H720 law has at most 16 distinct H336 prefixes from 64 terminal scenarios. Aggregating equal prefixes preserves the restricted distribution and prefix-dependent expected-cost objectives. With the lower-memory evaluator, H336 scoring allocation decreases from 44.46 MB to 11.43 MB.

  3. The predictive horizon trend also holds with the head fixed. Restricting one H720 forecast law gives +3.258 [1.920, 4.499] Energy skill points per horizon doubling over GB64. Appendix D.2 reports this single-law analysis, and Appendix D.4 reports exact aggregation and its scoring measurements.

Thank you to the authors for their response. However, my concerns remain unresolved:

  1. With the prediction head held fixed, CPSS’ advantage over GB64 continues to increase with the forecast horizon. However, Table D.3 shows that CPSS has worse Energy scores than the flat local basis model at every evaluated horizon. Could CPSS’ advantage over GB64 therefore be primarily attributable to the limitations of GB64’s global basis representation?

  2. Table D.8(c) shows that CPSS saves only approximately 4 MB of memory relative to the flat local basis model, while runtime remains essentially the same. Meanwhile, CPSS' Energy skill is 2.556 points lower than that of the flat model. In my assessment, this limited computational benefit does not justify the worse Energy performance.

Reviewer_LdH3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Conditional Projective Scenario Systems (CPSS) for converting marginal forecasts from a frozen time-series foundation model into a weighted distribution over complete future trajectories. CPSS organizes 64 scenarios through hierarchical, suffix-supported refinements, allowing related trajectories to share identical prefixes while introducing additional uncertainty at later forecasting steps. The resulting distribution supports exact prefix merging under horizon restriction. The method is trained using a combination of energy score, marginal CRPS, and entropy regularization. Experiments demonstrate increasing advantages over a capacity-matched global-basis baseline as the forecasting horizon grows, together with improvements in complete-path and path-dependent probabilistic forecasting.

Critical strengths
  1. Meaningful problem formulation. The paper addresses an important limitation of marginal time-series forecasting: accurate timestamp-wise distributions do not necessarily represent temporal dependence or path-dependent uncertainty. Modeling complete future trajectories is relevant to practical forecasting and decision-making.

  2. Structured scenario representation. Hierarchical suffix-supported refinements provide an interpretable coarse-to-fine construction. Exact shared prefixes and probability-mass aggregation establish a clear structural relationship between scenarios at different forecasting horizons.

  3. Promising long-horizon performance. Figure 2 and Appendix B demonstrate an increasing performance advantage over GB64, reaching +8.52+8.52 energy skill points at H=720H=720. The matched scenario budget, backbone, and parameter count strengthen this comparison.

  4. Multi-level evaluation. The paper evaluates complete-path energy, CRPS of path-dependent functionals, and marginal forecasting quality. Comparisons with TimeMCL, modern foundation models, and the supplementary decision-regret analysis provide useful supporting evidence.

Critical weaknesses
  1. The scope of projective consistency requires clarification. Equations (8)--(11) establish consistency under truncation of a single complete-trajectory distribution. However, any joint distribution admits consistent marginals under restriction. The distinctive property of CPSS is its exact shared-prefix support structure. Moreover, Appendix A.1 indicates that H=192H=192, 336336, and 720720 use separately trained heads, so consistency between independently generated forecasts at different horizons is not established. The distinction between support-level prefix consistency and cross-model projective consistency should be made explicit.

  2. Insufficient isolation of the hierarchical design. GB64 provides a useful capacity-matched baseline, but the comparison changes several factors simultaneously: hierarchical parameter sharing, suffix-supported bases, and tree-structured probability assignment. Consequently, the results do not establish which component drives the improvement. Additional ablations separating these mechanisms, together with different scenario budgets and refinement schedules, are necessary.

  3. Mixed evidence regarding temporal dependence quality. Table 1 shows that CPSS improves energy score over TimeMCL by 6.69%6.69\%, but worsens variogram score by 5.78%5.78\%, with a paired confidence interval of [−14.49,−0.14][-14.49,-0.14]. This suggests that improvements in overall trajectory scoring do not uniformly translate to better temporal dependence modeling. Furthermore, the 10.74%10.74\% improvement in threshold-duration CRPS has a confidence interval crossing zero. These qualifications deserve more discussion.

  4. Limited evaluation of cross-horizon generalization and statistical robustness. The horizon sweep combines restrictions of an H=96H=96 forecast with separately trained longer-horizon heads, complicating attribution of the observed trend exclusively to horizon length. The H=128H=128 joint evaluation contains only 96 cases across 12 domains and two training seeds. Appendix A.1 also notes that evaluation windows may overlap temporally. Additional independent forecast origins and larger evaluation panels would strengthen the statistical conclusions.

  5. Restricted modeling scope. The experiments primarily consider univariate forecasts with a fixed budget of 64 trajectories and predetermined refinement locations. The robustness of the approach to higher-dimensional joint forecasting, alternative scenario budgets, and adaptive branching remains untested. This limits conclusions about broader scalability and applicability.

Questions for authors
  1. The scope of projective consistency requires clarification. Equations (8)--(11) establish consistency under truncation of a single complete-trajectory distribution. However, any joint distribution admits consistent marginals under restriction. The distinctive property of CPSS is its exact shared-prefix support structure. Moreover, Appendix A.1 indicates that H=192H=192, 336336, and 720720 use separately trained heads, so consistency between independently generated forecasts at different horizons is not established. The distinction between support-level prefix consistency and cross-model projective consistency should be made explicit.

  2. Insufficient isolation of the hierarchical design. GB64 provides a useful capacity-matched baseline, but the comparison changes several factors simultaneously: hierarchical parameter sharing, suffix-supported bases, and tree-structured probability assignment. Consequently, the results do not establish which component drives the improvement. Additional ablations separating these mechanisms, together with different scenario budgets and refinement schedules, are necessary.

  3. Mixed evidence regarding temporal dependence quality. Table 1 shows that CPSS improves energy score over TimeMCL by 6.69%6.69\%, but worsens variogram score by 5.78%5.78\%, with a paired confidence interval of [−14.49,−0.14][-14.49,-0.14]. This suggests that improvements in overall trajectory scoring do not uniformly translate to better temporal dependence modeling. Furthermore, the 10.74%10.74\% improvement in threshold-duration CRPS has a confidence interval crossing zero. These qualifications deserve more discussion.

  4. Limited evaluation of cross-horizon generalization and statistical robustness. The horizon sweep combines restrictions of an H=96H=96 forecast with separately trained longer-horizon heads, complicating attribution of the observed trend exclusively to horizon length. The H=128H=128 joint evaluation contains only 96 cases across 12 domains and two training seeds. Appendix A.1 also notes that evaluation windows may overlap temporally. Additional independent forecast origins and larger evaluation panels would strengthen the statistical conclusions.

  5. Restricted modeling scope. The experiments primarily consider univariate forecasts with a fixed budget of 64 trajectories and predetermined refinement locations. The robustness of the approach to higher-dimensional joint forecasting, alternative scenario budgets, and adaptive branching remains untested. This limits conclusions about broader scalability and applicability.

4 comments
AuthorsOpenReview ↗

Response to Reviewer LdH3 (1/3): W1–W2 / Q1–Q2

The five weaknesses are repeated verbatim in the questions. We quote each once under both labels and answer it below.

W1 / Q1

The scope of projective consistency requires clarification. Equations (8)--(11) establish consistency under truncation of a single complete-trajectory distribution. However, any joint distribution admits consistent marginals under restriction. The distinctive property of CPSS is its exact shared-prefix support structure. Moreover, Appendix A.1 indicates that H = 192, 336, and 720 use separately trained heads, so consistency between independently generated forecasts at different horizons is not established. The distinction between support-level prefix consistency and cross-model projective consistency should be made explicit.

CPSS guarantees exact shared prefixes within one terminal forecast. At H336, lossless aggregation reduces 64 records to 16 and lowers measured scoring memory from 44.46 MB to 11.43 MB with the lower-memory evaluator.

  1. CPSS enforces exact shared-prefix identities. Related scenarios remain identical before their prescribed divergence. Truncation therefore merges their shared prefixes by summing descendant probabilities, giving a compact representation of the same restricted forecast law.

  2. The guarantee covers restrictions of one terminal forecast. The independently trained H192, H336 and H720 heads are separate forecast laws and are not constrained to agree. Section 3.3 explicitly states the within-forecast guarantee.

  3. The structure has a measured practical benefit. At H336, the same CPSS forecast can be evaluated using 16 prescribed prefix records instead of 64 unmerged records.

    H336 evaluation of the same issued CPSS forecastUnmergedAggregated
    Prefix records6416
    Incremental scoring memory, lower-memory evaluator44.46 MB11.43 MB

    Aggregation preserves the restricted distribution and every prefix-dependent expected-cost objective, including its optimum when attained. The memory row measures Energy plus sorted marginal CRPS at batch size 64 and excludes generation. Section 3.3 and Appendix D.4 give the identity and numerical checks.

W2 / Q2

Insufficient isolation of the hierarchical design. GB64 provides a useful capacity-matched baseline, but the comparison changes several factors simultaneously: hierarchical parameter sharing, suffix-supported bases, and tree-structured probability assignment. Consequently, the results do not establish which component drives the improvement. Additional ablations separating these mechanisms, together with different scenario budgets and refinement schedules, are necessary.

Matched ancestry improves Energy by +0.411 points, alongside variogram and maximum CRPS gains, with local bases and capacity fixed. The 48-fit study also tests suffix support, probability allocation, budgets and schedules separately.

  1. The matched intervention identifies a predictive contribution from nested sharing. Both configurations use the same local dictionary, 640,059 active parameters, conditioner, initial tensors, terminal-logit map and training batches, with zero entropy penalty. Only the intermediate ancestry assignments change, while node usage counts are preserved.

    Matched ancestry, positive skill favours nestingSkill points95% interval
    Energy+0.411[0.222, 0.568]
    Variogram+0.106[0.013, 0.257]
    Maximum CRPS+2.168[0.863, 3.496]

    The Energy effect is positive in all eight datasets and three seeds, and remains +0.321 [0.093, 0.548] when the ETT datasets form one family.

  2. Exact suffix support adds a distinct path-functional and structural benefit. CPSS gains +12.233 [4.569, 20.393] maximum-ramp CRPS points over relaxed support and preserves exact prescribed prefixes. Relaxation gains +2.793 [1.520, 3.921] Energy points. The separate flat local-basis system gains 2.069 [1.136, 2.892] Energy points, with a different coefficient factorisation and masses. These outcomes separate the sharing contribution from the predictive consequences of prefix constraints.

  3. The remaining controls test probability allocation and sensitivity. Terminal logits gain +0.070 [0.004, 0.136] Energy points over branch probabilities. The 48-fit study also tests scenario budgets, basis sizes and refinement schedules, with their results detailed in W5/Q5. Section 5.2 and Tables D.1-D.2 in Appendix D.1 provide the complete comparisons.

AuthorsOpenReview ↗

Response to Reviewer LdH3 (2/3): W3–W4 / Q3–Q4

W3 / Q3

Mixed evidence regarding temporal dependence quality. Table 1 shows that CPSS improves energy score over TimeMCL by 6.69, but worsens variogram score by 5.78, with a paired confidence interval of [-14.49, -0.14]. This suggests that improvements in overall trajectory scoring do not uniformly translate to better temporal dependence modeling. Furthermore, the 10.74 improvement in threshold-duration CRPS has a confidence interval crossing zero. These qualifications deserve more discussion.

CPSS improves Energy and the functional CRPS composite over TimeMCL in the 96-case evaluation. The 569-case evaluation also supports the composite gain under grouped resampling, with metric-specific comparisons and uncertainty detailed below.

  1. CPSS improves complete-path and functional forecasting in the original TimeMCL comparison. The selected results are ordinary relative loss reductions, with positive values favouring CPSS.

    Original 96-case H128 comparison against TimeMCLReduction, %95% interval
    Energy+6.69[1.96, 9.67]
    Four-functional CRPS+11.18[2.32, 19.14]
    Variogram-5.78[-14.49, -0.14]
  2. The metrics examine different properties of the forecast law. Energy scores complete trajectories, the functional scores assess specified path quantities, and variogram assesses moments of lagged increments. The results establish a metric-specific trade-off, with TimeMCL giving the better variogram result in this comparison.

  3. The expanded evaluation also supports the functional composite. On 569 additional cases, CPSS reduces the composite by 2.926% [0.365, 19.352] over TimeMCL under domain-and-item resampling. Separately, the original high-duration comparison against GB64 is +10.74% [-17.10, 22.61], whose interval includes zero. The composite improvement and that individual uncertainty are reported separately in Sections 5.3-5.4 and Appendix D.3.

W4 / Q4

Limited evaluation of cross-horizon generalization and statistical robustness. The horizon sweep combines restrictions of an H = 96 forecast with separately trained longer-horizon heads, complicating attribution of the observed trend exclusively to horizon length. The H = 128 joint evaluation contains only 96 cases across 12 domains and two training seeds. Appendix A.1 also notes that evaluation windows may overlap temporally. Additional independent forecast origins and larger evaluation panels would strengthen the statistical conclusions.

The horizon advantage over GB64 grows by +3.258 points per doubling with one H720 head fixed. The 569 additional H128 cases also support the functional CRPS composite gain under grouped resampling.

  1. The horizon trend persists within one terminal forecast law. We restricted the same frozen H720 forecasts to H96, H192, H336 and H720, keeping the trained head fixed.

    Single-H720-law comparison against GB64Estimate95% interval
    Energy skill slope, points per horizon doubling+3.258[1.920, 4.499]
    H720 Energy skill, points+8.518[4.631, 11.744]
  2. The additional H128 evaluation expands the evidence beyond the original 96 cases. The 569 new cases use fixed context 128, exclude full-window overlap within each source item and against both earlier evaluations, and retain the two frozen seeds. The functional-composite reductions are +3.921% [0.975, 21.106] over GB64 and +2.926% [0.365, 19.352] over TimeMCL under domain-and-item resampling.

  3. The supplementary H720 study tests origin-spacing sensitivity. Its 1,294 cases cover 67 dataset-level origins with internally disjoint target blocks and give +6.476 [2.753, 9.963] Energy points over GB64. Because 55 origins overlap earlier target intervals, we identify this as a sensitivity study rather than an independent holdout. Appendix D.2 distinguishes the populations, and the grouped uncertainty analyses retain associated datasets, channels or source-item windows.

AuthorsOpenReview ↗

Response to Reviewer LdH3 (3/3): W5 / Q5

W5 / Q5

Restricted modeling scope. The experiments primarily consider univariate forecasts with a fixed budget of 64 trajectories and predetermined refinement locations. The robustness of the approach to higher-dimensional joint forecasting, alternative scenario budgets, and adaptive branching remains untested. This limits conclusions about broader scalability and applicability.

The 48-fit study quantifies performance across scenario budgets, basis sizes and refinement schedules. The original schedule leads both tested alternatives at fixed capacity, and larger scenario budgets improve Energy while increasing model capacity.

  1. Scenario-budget sensitivity is now measured. Energy skill is relative to the 64-scenario reference, with positive values favouring the alternative.

    ScenariosActive parametersEnergy skill, points95% interval
    32492,315-1.976[-2.702, -1.065]
    64640,0590Reference
    128935,547+1.743[0.971, 2.463]
  2. The original refinement schedule outperforms both tested alternatives. At fixed capacity, its Energy advantages are +0.269 [0.105, 0.417] points over uniform refinement and +0.606 [0.315, 0.898] over later refinement. Two and eight basis functions were also tested against four, giving -2.293 [-3.225, -1.222] and +2.111 [1.099, 3.041] points, respectively. Budget and basis changes also change capacity, so parameter counts accompany the results.

  3. The sensitivity study covers multiple scenario budgets, basis sizes and refinement schedules. Multivariate forecasts and adaptive branching remain extensions for further study. Section 6 states this scope, and Appendix D.1 gives all 48 fits.

Reviewer_LdH3OpenReview ↗

Response to Authors' Rebuttal

I appreciate the authors' detailed and thorough rebuttal. The additional ablation studies, particularly the matched ancestry comparison, and the fixed-head horizon evaluation have addressed most of my major concerns. I also appreciate the authors' clarification of the scope of projective consistency and their transparent discussion of the trade-offs in temporal dependence evaluation.

While some limitations remain, I believe the new evidence substantially strengthens the paper. I therefore maintain my previous positive rating and am more inclined to support its acceptance.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response (1/2): Component, horizon and H128 results

Conditional Projective Scenario Systems for Long-Horizon Joint Forecasting

CPSS combines learned joint forecasting with exact progressive scenario refinement. Our experiments establish a predictive contribution from nested sharing and a computational benefit from its shared-prefix structure. We thank the three anonymous reviewers for their comments. The evaluation includes 48 component-study fits, H128 panels of 96 and 569 cases, and checks of training exclusions and window overlap.

1. Controlled component tests identify predictive benefits.

The matched ancestry intervention holds the local dictionary, 640,059 active parameters, conditioner, initialisation, terminal-logit probability map and training batches fixed, with zero entropy penalty in both arms. The following selected H720 results distinguish the predictive contribution of nested sharing from the consequences of exact suffix support. Positive skill favours nested ancestry or exact support, respectively.

InterventionMetricSkill (points)95% interval
Matched ancestryEnergy+0.411[0.222, 0.568]
Matched ancestryVariogram+0.106[0.013, 0.257]
Matched ancestryMaximum CRPS+2.168[0.863, 3.496]
Exact suffix supportMaximum-ramp CRPS+12.233[4.569, 20.393]
Exact suffix supportEnergy−2.793[−3.921, −1.520]

The sharing Energy gain is positive in all eight datasets and all three seeds. Grouping the ETT datasets into one family gives +0.321 [0.093, 0.548]. Table D.2 in Appendix D.1 gives all nine metrics per intervention. The same subsection reports mass, budget, basis-size and schedule sensitivities.

2. The long-horizon gain persists without changing heads.

Restricting the same H720 forecasts to H96, H192, H336 and H720 gives an energy advantage over GB64 that increases by +3.258 [1.920, 4.499] points per horizon doubling. The H720 gain is +8.518 [4.631, 11.744] points under dataset resampling.

3. The H128 evaluation supports path-functional forecasting.

Across all 569 H128 cases with fixed context 128, CPSS improves the four-functional CRPS composite by 3.921% [0.975, 21.106] over GB64 and 2.926% [0.365, 19.352] over TimeMCL, using domain-and-source-item resampling.

Both fitted seeds give positive reductions against both methods. The 96-case mixed-context panel gave 16.80% over GB64. The smaller positive effect in the 569-case panel was mainly associated with threshold duration. Appendix D.3 explains these panel differences.

4. Exact refinement makes shorter-horizon evaluation more compact.

At H336, CPSS requires at most 16 distinct prefixes, whereas the flat local-basis forecasts have 64 in every checked case and seed. Aggregation preserves the restricted forecast distribution and prefix-dependent objectives while reducing incremental scoring memory from 353 MB to 22 MB and making the tested operation 1.78 times as fast, or 2.76 times as fast at H540.

Table D.7 gives these score-only measurements, including H720 with no prefix reduction. A matched implementation test reduced CPSS generation from 67.32 ms to 3.18 ms with the same frozen checkpoints.

Including generation and H336 evaluation, incremental peak memory was 60.3% lower than the cached controls with the evaluator that materialises pairwise trajectory differences. The memory advantage persisted with streamed decoders and a lower-memory evaluator (Table D.8). Carrier generation is excluded.

AuthorsOpenReview ↗

General response (2/2): Path-dependent forecasts, efficiency and conclusion

5. CPSS improves specified path-dependent forecasts over flat local-basis models.

We evaluated frozen H720 forecasting heads across 1,480 cases, eight datasets and three fitted seeds.

Both four-basis models produce 64 scenarios with closely matched parameter counts: 640,059 for CPSS and 638,464 for flat local basis.

The following two findings retained positive simultaneous confidence bands after accounting for the broader exploratory screen, together with positive grouped sensitivity results in both evaluated populations.

Table 1. Selected path-dependent forecasting results

Forecast quantityHorizonBases, CPSS/flatLoss reduction95% interval
Largest accumulated fall, CRPSH964/4+5.57%[3.35, 7.38]
Future-average q90 decisionH1924/4+6.63%[3.63, 9.50]

The percentage reductions give all eight datasets equal weight. The q90 decision penalises underestimation nine times more heavily than an equal overestimation.

These findings demonstrate measurable improvements on specific path-dependent objectives. They complement the complete-path Energy results by evaluating quantities directly relevant to forecasting decisions.

Appendix D.6 provides the full exploratory-screening protocol, comparisons across horizons, statistical adjustments and sensitivity analyses, including the overlapping origin-spacing evaluation.

6. CPSS improves decision quality and computational efficiency on the same workload.

For the H192 future-average decision, CPSS achieves both 6.63% lower loss and 1.97× faster execution under the same 64 MB memory limit.

For a queue of 256 requests, four-basis CPSS completes the decision workload in 2.475 ms, compared with 4.884 ms for the closely matched flat local-basis model.

Direct restriction also provides substantial computational benefits.

Table 2. Direct H336 restriction and scoring

MeasurementCPSS, four basesFlat local basis
Prefix records1664
Execution time2.941 ms3.806 ms
Incremental peak memory12.585 MB49.648 MB

This corresponds to 74.7% lower incremental peak memory and 1.29× faster execution.

Both decoders were allowed to omit inactive suffix computations. CPSS additionally exploits its prescribed identical prefixes and aggregates the corresponding probability masses. Appendix D.7 distinguishes direct restriction from full H720 generation.

The computational benefits extend to conditional-probability queries. With matched eight-basis heads and 64 scenarios, generating H96 conditional-category tables for 256 requests required 3.648 ms for CPSS versus 29.351 ms for flat, an 8.05× speed-up under a 64 MB live-allocation limit.

This is a computational comparison of the same requested output, distinct from conditional-probability accuracy. The gain varies with the available memory budget, as documented in Appendix D.8.

All reported timings use an NVIDIA RTX A6000 with cached TimesFM forecasts, excluding carrier generation and transfers. Numerical checks confirm agreement with the corresponding original forecast restrictions within the reported tolerances.

Conclusion

The additional experiments strengthen the evidence for CPSS in three important respects.

First, the matched ancestry intervention establishes a positive predictive contribution from nested sharing beyond local basis geometry.

Second, the fixed-head and expanded evaluations support the long-horizon and path-functional forecasting results.

Third, exact prefix structure provides measurable advantages in forecast representation, decision quality and computational efficiency.

For the H192 future-average decision, CPSS achieves 6.63% lower loss while completing the same workload 1.97 times faster under a 64 MB memory limit. Direct H336 restriction also reduces incremental peak memory by 74.7% and provides a 1.29× execution speed-up.

These findings demonstrate why path-dependent forecasting quality, decision performance and computational efficiency should be evaluated alongside aggregate Energy when assessing structured joint forecasting methods.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.