acceptodds
Withdrawn from ICLR 2027

PrefixFlow: Projective Consistency for Probabilistic Time-Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Probabilistic forecasts at different horizons share a common future prefix, yet direct generative forecasters need not yield consistent predictive distributions over it. We formalize this requirement as horizon-wise projective consistency. Using Energy Distance, we show that three recent generative forecasters exhibit systematic cross-horizon discrepancies beyond finite-sample Monte Carlo variation. To address this, we propose PrefixFlow, a variable-horizon Flow Matching model trained jointly across horizons and supporting any patch-aligned horizon up to a prescribed maximum. Unlike autoregressive forecasters that require sequential rollout, PrefixFlow generates the entire requested future in parallel. Its prefix-invariant conditioning and prefix-causal velocity field preserve the generative dynamics of an existing prefix under horizon extension; together with a projectively consistent Gaussian source, this yields projectively consistent predictive distributions. Across six benchmarks, PrefixFlow achieves the lowest average CRPS on five, with no detectable cross-horizon discrepancy beyond finite Monte Carlo variation.

Primary area:
learning on time series and dynamical systems
Keywords:
probabilistic time series forecasting, projective consistency, variable-horizon forecasting, flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LU2rOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper motivates itself by arguing that probabilistic forecasts of the same history at different horizons should agree on their shared future prefix (horizon-wise projective consistency). It shows that (sota) generative models, trained separately per horizon, are not consistent. They propose PrefixFlow, with Prefix-masked conditioning and a left-causal velocity network, to make the prefix computation independent of the suffix, which together with a factorized Gaussian source guarantees consistency. PrefixFlow is trained jointly across horizons.

Critical strengths
  • They define the problem of consistency clearly and point out that current metric do not capture it.
  • The consistency evaluation is well done.
  • The single jointly trained model is better in metric than the horizon-specific baselines.
Critical weaknesses
  • It is unclear if the inconsistency is a horizon effect. The analysis compares separately trained models per horizon. However, two independently trained models are expected to disagree, even at the same horizon. The paper does not controll for training variance (seeds etc). The authors’ own Separate Training variant (Excess ED 1.59 with an identical architecture) is consistent with this concern.
  • The practical relevance of consistency is not demonstrated. For the true conditional distribution the marginals are consistent by definition, but it does not follow that disagreement between imperfect models is harmful. Consistency is not correctness: a model can be perfectly consistent and poorly calibrated, and two inconsistent forecasts can be equally well calibrated. MOSES (Yalavarthi et al., 2026) shows how this can be tested: it measures the cost of inconsistency by comparing marginal predictive accuracy. An analogous experiment here, scoring the baselines’ Direct-h and Crop-from-H forecasts against ground truth on the shared prefix, would show whether their inconsistency actually hurts. Also I find the paper to not adequately compare qualitatively and quantitatively to the MOSES paper.
  • Joint training across horizons appears to do most of the work. In table 3 (ETTh1 96 720) wo both almost matches the causal coinstrained one, i.e., the causal constraints only remove this small residual and actually worsen the CRPS. Further, a baseline is missing: a single model trained on all horizons with no causal structure and an explicit horizon encoding.
  • Limited theoretical novelty given MOSES.
Questions for authors

See waeknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer LU2r (1/2)

Thank you for the detailed feedback. We added several controlled experiments and clarified the relation to MOSES.

1. Separately trained models vs. horizon sensitivity [W1]

We agree that the experiment in Sec. 3.2 changes both the forecast horizon and the trained model, so it does not isolate horizon sensitivity. The experiment was intended to reflect the usual horizon-specific setup of these baselines, but independent training can clearly contribute to the disagreement.

We repeated the test with a single jointly trained checkpoint, varying only the requested horizon.

Excess ED from the same checkpoint evaluated at different horizons; 95% bootstrap confidence intervals are shown.

DatasetPairPrefixFloww/o Both
ETTh196–1920.0010 [-0.0011, 0.0022]0.0561 [0.0486, 0.0613]
ETTh196–7200.0013 [-0.0003, 0.0033]0.1697 [0.1374, 0.2169]
Weather96–192-0.0002 [-0.0012, 0.0010]0.0361 [0.0334, 0.0389]
Weather96–720-0.0001 [-0.0010, 0.0009]0.1088 [0.0974, 0.1194]

With the same checkpoint, w/o Both still shows clear cross-horizon disagreement, while PrefixFlow remains near zero.

2. Does inconsistency have a predictive cost? [W2, first part]

We agree with the reviewer that consistency is not correctness: exact consistency alone does not imply better calibration or lower CRPS. Following the suggested MOSES-style test, we score the same H96 target using either a direct H96 forecast or the H96 prefix cropped from an H720 forecast.

CRPS for the same H96 target obtained through Direct-H96 or Crop-from-H720, using 256 predictive samples.

DatasetModelDirectCrop
ETTh1PrefixFlow0.29050.2904
D3U0.32150.3343
TSFlow0.34280.3149
NsDiff0.37830.3608
WeatherPrefixFlow0.14070.1407
D3U0.17640.1676
TSFlow0.19250.1837
NsDiff0.19800.2284

The baseline results can change substantially between Direct and Crop, but the direction varies across models and datasets. We therefore do not claim that inconsistency always lowers predictive accuracy. For these baselines, Direct and Crop come from separately trained horizon-specific models, so the difference is not purely a horizon effect.

We also test whether the structural constraint helps when shorter horizons receive no direct supervision. PrefixFlow and w/o Both are trained only at H720, using the same coverage-aware positional weighting; H96/H192/H336 are never used for training. Direct and Crop use the same checkpoint and the same H720-valid test origins.

H720-only training on ETTh1. Each entry is MSE / CRPS.

ModelRouteH96H192H336
PrefixFlowDirect0.3837 / 0.30410.4073 / 0.31620.4200 / 0.3245
Crop-from-H7200.3834 / 0.30400.4072 / 0.31620.4198 / 0.3244
w/o BothDirect0.4105 / 0.31070.4347 / 0.32090.4551 / 0.3316
Crop-from-H7200.3940 / 0.30560.4257 / 0.31940.4485 / 0.3308

Under H720-only supervision, PrefixFlow gives better direct forecasts than w/o Both at all shorter horizons. In PrefixFlow, the distribution queried at a shorter horizon is exactly the corresponding marginal of the supervised H720 forecast, so H720 training constrains those shorter-horizon predictions as well. This link is absent in w/o Both.

AuthorsOpenReview ↗

Response to Reviewer LU2r (2/2)

3. Joint training and explicit horizon conditioning [W3]

We agree that joint multi-horizon training already produces strong cross-horizon alignment. We view this as a soft consistency bias: sharing one model and training it across horizons encourages the corresponding predictive distributions to agree, even without the structural constraints. This is also why w/o Both performs well in Table 3 and slightly outperforms Full in CRPS.

The reviewer also suggested a shared noncausal model with an explicit horizon encoding. We added this baseline by augmenting w/o Both with an MLP embedding of the requested horizon, while keeping the model noncausal and without prefix masking.

ETTh1 results for jointly trained shared models. Forecasting columns report CRPS.

ModelH96H192H336H720Excess ED 96–192Excess ED 96–720
PrefixFlow0.29720.31450.32350.33750.00100.0013
w/o Both0.29530.31120.31970.33420.05600.1697
+ Horizon Encoding0.29620.31200.31990.33230.05630.1903

Explicit horizon conditioning leaves forecasting accuracy essentially unchanged, but does not reduce the cross-horizon discrepancy. Thus, joint training can learn substantial empirical alignment, but neither parameter sharing nor explicit knowledge of the requested horizon is sufficient to make the resulting distributions projectively consistent.

4. Relation to MOSES and theoretical novelty [W2, second part + W4]

We agree that MOSES is closely related and should have been discussed more clearly in the original submission.

MOSES focuses on probabilistic forecasting for irregular time series, where future queries may contain arbitrary time points and channels. To remain consistent under arbitrary subqueries, MOSES uses separable flow transformations, with dependence across queried variables mainly handled by the Gaussian source and mixture components. The paper also discusses the expressivity trade-off introduced by this separability.

Our problem is different: we consider regular-grid probabilistic forecasting with variable horizons. The query sets are not arbitrary; they are nested prefixes. PrefixFlow only requires the velocity field on an existing prefix to remain unchanged when a suffix is appended. This keeps temporal interactions within the prefix while preserving consistency across horizons.

So the novelty is not the idea of consistency itself, but how to enforce it for nested forecasting horizons without making the flow fully separable. We additionally show that this prefix property is preserved by the Euler discretization used at inference.

We also compared MOSES and PrefixFlow at H=96H=96 on three regular-grid benchmarks.

Forecasting results at H=96H=96. Each entry is MSE / CRPS.

DatasetPrefixFlowMOSES
ETTh10.3760 / 0.29721.3877 / 0.6428
ETTm10.2789 / 0.24681.0918 / 0.6274
Weather0.1458 / 0.14300.5076 / 0.4282

PrefixFlow performs substantially better on all three datasets in this regular-grid forecasting setting.

Thanks a lot for the detailed rebuttal and providing most of the experiments that I asked for. I especially appreciate providing results that did not promote the method. However, these results do mostly confirm my concerns that joint training already captures most of the inconsistency and especially when considering the seed variance, most remaining differences are within run variances. At the same time forcing a causal structure removes most of the benefits that joint modelling provides.

Overall, I see this more as a investigative report of cross-horizon consistency showing very limited impact and keep my score.

One last remark: the moses results seem implausibly poor to me, maybe the setup was misconfigured.

Reviewer_1GrBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues that the non-autoregressive class of probabilistic forecasting models lacks projective consistency; querying for different future horizons leads to different distributions for the common horizons, making it difficult for practitioners to judge which one is correct. While autoregressive models satisfy projective consistency by design, they are expensive during inference as they need to be rolled out for longer horizons.

The authors propose a new probabilistic model called prefixflow, which is a flow-matching-based architecture with projective consistency by design. Prefixflow can be trained once and evaluated for different horizons, and common horizons have the same distribution.

Experimental results state that despite the constraints of projective consistency, prefixflow performs better than diffusion/flow based baselines in terms of CRPS and MSE on most datasets. Projective inconsistency is measured using excess Energy Distance, and PrefixFlow has it within the sampling error.

Critical strengths
  1. PrefixFlow is a useful probabilistic model for practical reasons. It can be trained once and can be evaluated at different horizons. This avoids storing and running different models for different future horizons. Alternatively, one could train a model that does not have projective consistency for long horizons and use it to predict for the short horizons as well. However, to predict the distribution of the next future step, one still needs to query the entire sequence for these models.

  2. Experimental results show that:

  • PrefixFlow trained once is better than the compared baselines that are trained separately for different horizons in most of the datasets in terms of CRPS and MSE
  • PrefixFlow has almost 0 excess ED, while the baselines that do not have projective consistency have larger values
  • Experiments show that PrefixFlow evaluated for different horizons gives good accuracy for the (intermediate) horizons it was not trained for.
Critical weaknesses

Most of the weaknesses are empirical

  1. The evaluation is only in terms of CRPS to measure the distribution accuracy. The consistency comes from the joint law; hence, one should use the energy score and Variogram score for evaluation. This experiment will answer the question of whether the projective consistency does not deteriorate joint distributions.
  2. Following (1), a simple and cheap baseline is missing. The PrefixFlow conditioner trained with quantile loss or a Gaussian distribution with independent marginals. This is guaranteed to be consistent and shows how much flow matching helps.
  3. Table 2 shows baselines trained for different horizons. However, it would be interesting to see the performance of baselines trained for long horizons and evaluate them for multiple horizons.
Questions for authors
  1. Since PrefixFlow is consistent, why do the authors need to train for different horizons in an epoch? One could choose large horizons and just weigh each time point differently. This way, training can be done at every possible intermediate horizon.
  2. On the other hand, the authors weighed earlier horizons more heavily than the farther ones. What happens if they train only for long horizons with uniform weighting and predict for intermediate horizons?
  3. Can the authors report the variance of the results from different seeds in the tables?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 1GrB (1/2)

Thank you for the constructive feedback. We address the main experimental concerns below.

1. Energy Score and Variogram Score [W1]

Following the reviewer’s suggestion, we additionally evaluate multivariate Energy Score (ES) and Variogram Score (VS) at H=96H=96 using the concatenated H×CH\times C forecast samples.

ES/VS at H=96H=96; lower is better.

DatasetModelESVS
ETTh1PrefixFlow11.12230.1109
w/o Both11.10220.1101
D3U11.51170.1311
TSFlow12.80500.1256
NsDiff13.38510.1790
WeatherPrefixFlow11.57490.0559
w/o Both11.67110.0555
D3U12.53340.0703
TSFlow13.83730.0770
NsDiff12.97200.1141

PrefixFlow remains comparable to w/o Both and outperforms the three external generative baselines on these joint scores. We will also clarify that PrefixFlow is channel-independent and therefore does not explicitly model cross-variable dependence.

2. Simple consistent baseline [W2]

We added an independent-Gaussian baseline using the same prefix-invariant conditioner. This tests whether Flow Matching still helps when consistency is already guaranteed by a simple model.

Comparison with the independent-Gaussian consistent baseline. Each entry is MSE / CRPS / temporal VS (T-VS); lower is better.

DatasetModelH96H192H336H720
ETTh1PrefixFlow0.3760 / 0.2972 / 0.11570.4128 / 0.3145 / 0.12290.4336 / 0.3235 / 0.12900.4556 / 0.3375 / 0.1337
Indep. Gaussian0.3790 / 0.2887 / 0.12360.4149 / 0.3068 / 0.13120.4354 / 0.3185 / 0.13880.4594 / 0.3386 / 0.1512
WeatherPrefixFlow0.1458 / 0.1430 / 0.05340.1895 / 0.1760 / 0.06350.2418 / 0.2076 / 0.07230.3184 / 0.2488 / 0.0858
Indep. Gaussian0.1640 / 0.1603 / 0.13340.2132 / 0.1944 / 0.14780.2665 / 0.2265 / 0.16200.3446 / 0.2704 / 0.1892

On ETTh1, the Gaussian baseline has lower marginal CRPS at H96–H336, while PrefixFlow consistently achieves lower temporal VS. On Weather, PrefixFlow also improves MSE and CRPS. This shows that exact consistency can already be achieved by a simple Gaussian baseline, while PrefixFlow better captures temporal dependence, as reflected by its consistently lower temporal VS.

3. Long-horizon baselines evaluated at shorter horizons [W3]

Following the reviewer’s suggestion, we evaluate each H=720H=720 baseline on ETTh1 at h∈{96,192,336,720}h\in\{96,192,336,720\} by cropping its predictive samples. Both the horizon-specific and H=720H=720 models are evaluated on the same H720-valid test origins.

Horizon-specific training versus reusing an H=720H=720 model by cropping on ETTh1. Each entry is MSE / CRPS, using 100 predictive samples.

ModelTraining / evaluationH96H192H336H720
D3Uhorizon-specific0.3792 / 0.31180.4413 / 0.34070.5221 / 0.43400.4985 / 0.3727
H720 →\rightarrow h0.4120 / 0.32890.4327 / 0.33980.4580 / 0.35370.4985 / 0.3727
TSFlowhorizon-specific0.4023 / 0.33300.4739 / 0.37400.5260 / 0.40370.6342 / 0.4534
H720 →\rightarrow h0.4893 / 0.35940.5493 / 0.41020.5701 / 0.41850.6342 / 0.4534
NsDiffhorizon-specific0.5111 / 0.37300.5048 / 0.37330.5532 / 0.39740.7592 / 0.4565
H720 →\rightarrow h0.5447 / 0.38150.5725 / 0.39360.5934 / 0.40590.7592 / 0.4565

Reusing the H720 model is not uniformly competitive with horizon-specific training. For D3U, cropping improves both MSE and CRPS at H192 and H336, but degrades performance at H96. For TSFlow and NsDiff, the H720 model performs worse at all three shorter horizons.

AuthorsOpenReview ↗

Response to Reviewer 1GrB (2/2)

4. Why multi-horizon training? What happens with Hmax⁡H_{\max}-only training? [Q1 + Q2]

Multi-horizon training is not required for the projective-consistency guarantee, which follows from the architecture. We use it to improve predictive quality across requested horizons. We therefore added the following ETTh1 ablation.

ETTh1 forecasting results under different training schemes. Each entry is MSE / CRPS.

TrainingH96H192H336H720
Joint + coverage-aware0.376 / 0.2970.413 / 0.3140.434 / 0.3230.456 / 0.338
Joint + uniform0.374 / 0.2960.412 / 0.3130.434 / 0.3220.459 / 0.337
H720-only + positional weighting0.416 / 0.3190.448 / 0.3340.462 / 0.3410.454 / 0.345
H720-only + uniform0.447 / 0.3290.477 / 0.3430.496 / 0.3520.476 / 0.353

H=720H=720-only models can still be queried at shorter horizons, but their predictive quality is substantially worse at H=96H=96–336336. Positional weighting improves H=720H=720-only training, while joint training gives the best overall performance. Under joint training, coverage-aware and uniform weighting are essentially tied.

5. Variance across random seeds [Q3]

Following the reviewer’s suggestion, we reran PrefixFlow with seeds 2025, 2026, and 2027. We report mean ±\pm standard deviation across seeds; each entry is MSE / CRPS. For Avg., we first average the four horizons within each seed and then compute the cross-seed statistics.

Multi-seed results for the five datasets with horizons {96,192,336,720}\{96,192,336,720\}.

DatasetH96H192H336H720Avg.
ETTh10.3852±0.0129 / 0.3013±0.00820.4228±0.0115 / 0.3182±0.00730.4432±0.0084 / 0.3265±0.00510.4601±0.0331 / 0.3387±0.00630.4278±0.0080 / 0.3212±0.0040
ETTm10.2872±0.0096 / 0.2521±0.00510.3237±0.0122 / 0.2699±0.00490.3583±0.0150 / 0.2859±0.00520.4218±0.0196 / 0.3129±0.00630.3477±0.0141 / 0.2802±0.0054
Weather0.1453±0.0016 / 0.1425±0.00060.1896±0.0020 / 0.1761±0.00060.2418±0.0014 / 0.2077±0.00030.3191±0.0010 / 0.2493±0.00050.2240±0.0015 / 0.1939±0.0003
Exchange0.0939±0.0062 / 0.1565±0.00610.2061±0.0191 / 0.2361±0.01240.3991±0.0428 / 0.3399±0.02241.1751±0.1351 / 0.6405±0.05110.4686±0.0506 / 0.3432±0.0226
ECL0.1341±0.0014 / 0.1767±0.00080.1502±0.0014 / 0.1883±0.00070.1669±0.0014 / 0.2013±0.00060.2053±0.0020 / 0.2261±0.00090.1642±0.0015 / 0.1981±0.0008

Multi-seed results for ILI.

DatasetH24H36H48H60Avg.
ILI2.3651±0.2236 / 0.6330±0.02262.2712±0.1341 / 0.6453±0.01361.9921±0.0958 / 0.6257±0.00981.7954±0.0298 / 0.6277±0.00472.1059±0.1059 / 0.6329±0.0114

Thanks for the additional experiments. My remaining concerns:

  1. ES/VS results are only at H=96 on two datasets without confidence intervals, and w/o Both is slightly better on 3/4 numbers. Please report these values for all the datasets and all the horizons with CIs. You may not need to run the models again just compute the metrics for the samples you already can draw from the trained models.

  2. Gaussian is also on two datasets without energy score, and of them, Gaussian beats PreFixFlow in one dataset. Again, can you report these for all the datasets with confidence intervals? It is hard to judge whether PreFixFlow has a better advantage than Gaussian.

  3. Joint vs. H720-only training:

  • If the joint training improves the MSE of shorter horizons, it should automatically improve it for longer horizons as we take the average at the end. But here, Joint + coverage-aware has a bad MSE at 720 and better ones for smaller horizons. So, is the resampling causing adverse effects at the long horizons? Please explain this.
  • For a prefix-causal model, joint training (resampling horizons) and H720 training with matched per-position weights are equivalent in expectation; only gradient stochasticity differs, which should not cause gaps as large as 0.376 vs 0.416 MSE at H96.

A careful study is required.

Considering these points, I maintain my score.

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

This paper identifies systematic cross-horizon inconsistencies in separately trained generative forecasters. It has shown identical histories can yield conflicting distributions over the same future prefix. It formalizes projective consistency, derives sufficient conditions for its preservation, and proposes PrefixFlow, a variable-horizon Flow Matching architecture that guarantees consistency across supported horizons while generating only the requested length. Experiments show competitive forecasting performance and efficiency gains over maximum-horizon generation followed by cropping. Ablations support the contribution as a structural coherence guarantee with limited observed forecasting cost.

Critical strengths

Strengths:

  1. The work systematically identifies cross-horizon incoherence as a limitation distinct from forecasting accuracy, controlling for finite-sampling variation.
  2. This work has introduced PrefixFlow with guarantees on consistency by construction. They also provide theoretical support extending to shared-grid Euler inference.
  3. Experiments demonstrate competitive forecasting under consistency constraints and computational savings over maximum-horizon generation followed by cropping.
Critical weaknesses

Weakness:

  1. The practical value of exact consistency remains unclear. The experiments show less disagreement across horizons, but do not show that enforcing consistency improves the decision outcomes used to motivate the work.
  2. The main diagnosis changes both the forecast horizon and the trained model. It therefore does not isolate horizon dependence, while the shared-model comparison without the proposed constraints is limited to one dataset and one horizon pair.
  3. The forecasting metrics mainly assess individual time points and variables, rather than their joint behaviour. They therefore provide limited evidence about whether the predicted dependencies are accurate, particularly given that PrefixFlow generates variables independently.
Questions for authors

Questions for Authors

  1. In Table 3, removing both constraints increases cross-horizon discrepancy but leaves MSE and CRPS largely unchanged. What practical problem does the remaining discrepancy in w/o Both cause, and what is gained by eliminating it?
  2. Table 3 omits MSE and CRPS for Separate Training, and the coverage-aware loss weighting is not ablated. Matched-horizon comparisons and a weighting ablation would clarify which training choices contribute to the forecasting gains.
  3. The diagnosis in section 3.2 changes both the forecast horizon and the trained model. Please distinguish disagreement between separately trained models from horizon sensitivity within a shared model.
  4. The efficiency comparison in Table 4 is limited to PrefixFlow’s own maximum-horizon generation followed by cropping. Please compare latency and memory with representative forecasting baselines under comparable settings, alongside predictive performance.
  5. In Figure 3(c) and Table 12, smooth performance trends do not quantify the unseen-horizon performance gap, since forecasting difficulty varies with horizon. Please compare against models trained on the same target horizons to assess this gap.
5 comments
AuthorsOpenReview ↗

Response to Reviewer pXUr (1/2)

Thank you for the constructive review. We address the three main concerns below.

1. Practical value of exact consistency [W1 + Q1]

We agree that projective consistency does not by itself imply better MSE/CRPS or better decisions. Its role is complementary: predictions for the same future prefix should not depend on the arbitrary choice of requesting a longer horizon.

Joint multi-horizon training already provides a strong soft consistency bias, but does not eliminate horizon-dependent disagreement. For the same H96 event queried through H∈{96,192,720}H\in\{96,192,720\}, the any-query decision-flip rate is 7.01% for w/o Both versus 1.48% for PrefixFlow on ETTh1, and 4.29% versus 2.07% on Weather. Thus, the residual distributional discrepancy in w/o Both can still propagate to horizon-dependent decisions, whereas PrefixFlow removes this horizon dependence by construction.

2. Separately trained models vs. shared-model horizon sensitivity [W2 + Q3]

We agree that Sec. 3.2 changes both the forecast horizon and the trained model. The original experiment follows the usual horizon-specific setup used by these baselines. It asks whether forecasts from these separately trained models agree on their shared prefix. To isolate horizon sensitivity itself, we additionally fix the model parameters and vary only the requested horizon.

Cross-horizon consistency of the same jointly trained model. Excess ED is reported with 95% bootstrap confidence intervals.

DatasetPairPrefixFloww/o Both
ETTh196–1920.0010 [-0.0011, 0.0022]0.0561 [0.0486, 0.0613]
ETTh196–7200.0013 [-0.0003, 0.0033]0.1697 [0.1374, 0.2169]
Weather96–192-0.0002 [-0.0012, 0.0010]0.0361 [0.0334, 0.0389]
Weather96–720-0.0001 [-0.0010, 0.0009]0.1088 [0.0974, 0.1194]

With the same checkpoint, w/o Both still shows clear cross-horizon disagreement, while PrefixFlow remains near zero.

3. Joint predictive quality and channel independence [W3]

We agree that CRPS does not assess joint predictive quality. At H=96H=96, we additionally compute multivariate Energy Score (ES) and Variogram Score (VS) on the concatenated H×CH\times C forecast vector.

Joint ES/VS at H=96H=96; lower is better.

DatasetModelESVS
ETTh1PrefixFlow11.12230.1109
w/o Both11.10220.1101
D3U11.51170.1311
TSFlow12.80500.1256
NsDiff13.38510.1790
WeatherPrefixFlow11.57490.0559
w/o Both11.67110.0555
D3U12.53340.0703
TSFlow13.83730.0770
NsDiff12.97200.1141

PrefixFlow remains comparable to w/o Both and outperforms the three external generative baselines on these joint scores. We also clarify that PrefixFlow models temporal dependencies within each variable, but not cross-variable dependencies, since variables are generated independently.

AuthorsOpenReview ↗

Response to Reviewer pXUr (2/2)

We address the remaining questions below.

4. Separate Training and coverage-aware weighting [Q2]

We added matched-horizon results for independently trained PrefixFlow models and ablated the coverage-aware weighting.

ETTh1 forecasting results. Each entry is MSE / CRPS.

TrainingH96H192H336H720
Joint + coverage-aware0.376 / 0.2970.413 / 0.3140.434 / 0.3230.456 / 0.338
Joint + uniform0.374 / 0.2960.412 / 0.3130.434 / 0.3220.459 / 0.337
Separate Training0.377 / 0.2990.430 / 0.3190.447 / 0.3310.477 / 0.353

Joint training is comparable at H=96H=96 and improves the longer-horizon results, while the two weighting schemes are essentially tied. This indicates that the forecasting gains mainly come from shared multi-horizon supervision rather than the particular coverage-aware weighting.

5. Efficiency relative to forecasting baselines [Q4]

We additionally compare forecasting performance on ETTh1 at H=336H=336, and measure inference latency and peak memory for generating 100 predictive trajectories from one test instance.

ETTh1 forecasting quality and inference efficiency at H=336; lower is better.

ModelMSE / CRPSLatency (s)Peak memory (MiB)
PrefixFlow0.4336 / 0.32350.198 ± 0.00217.9
D3U0.4748 / 0.35180.627 ± 0.012307.7
TSFlow0.5668 / 0.42062.949 ± 0.021359.1
NsDiff0.5477 / 0.38486.370 ± 0.06995.8

Under this setting, PrefixFlow has the lowest latency and peak memory while retaining the best predictive performance among the compared methods.

6. Unseen-horizon performance gap [Q5]

To quantify the gap at unseen horizons, we trained target-specific PrefixFlow models directly at H∈{120,240,480,600}H\in\{120,240,480,600\} and compared them with the shared model evaluated at the same horizons.

Shared-model versus target-specific forecasting at unseen horizons. Each entry is MSE / CRPS.

HShared modelTarget-specific
1200.3877 / 0.30260.3839 / 0.3018
2400.4239 / 0.31930.4456 / 0.3311
4800.4361 / 0.32620.4563 / 0.3368
6000.4368 / 0.33000.4384 / 0.3355

The shared model is close to the target-specific model at H=120H=120 and achieves lower MSE/CRPS at H=240,480,600H=240,480,600 in this comparison.

I appreciate the authors’ effort in providing additional experiments and detailed responses which have addressed most of my concerns. However, I still have further questions on two points and they may need stronger justifications.

  1. Regarding W1&Q1, the new analysis provides useful evidence of improved stability on decision making. However it could still be that one method is more stable in making the false decision, which is not desired. I wonder whether it would also be possible to evaluate decision correctness using the same predictions, for example through event classification accuracy. In this way, additional results would help clarify whether the improved consistency also results into more reliable predictions and decision making, beyond reducing inconsistency across future horizons.

  2. Regarding Q2, the improvements from joint training appear relatively modest which supports the main methodology. However coverage-aware weighting performs almost identical to uniform weighting. Without statistical significance tests, it will be difficult to claim the reliability of these gains. There needs stronger empirical support on the chosen approach and the motivation for this design.

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

Related papers

Discussion (0)

Sign in to comment.