acceptodds
Under review as a conference paper at ICLR 2027

Variational Inference for Dynamic Mixtures in Probabilistic Time-Series Forecasting

Abstract

Probabilistic time-series forecasting requires modeling complex, dynamic, and often multimodal conditional distributions while maintaining computational efficiency. We propose a lightweight finite mixture framework that uses a mixture sieve model and an analytically derived optimal posterior for reconstruction, thereby reducing the variational inference problem for conditional distribution estimation to a multi-class classification task. The framework defines the model output in the fractionally differenced domain, which alleviates distribution shift in the target variable and yields a dynamic mixture representation with time-varying component locations on the original scale. The proposed objective provides a quantization-level distributional approximation guarantee and has bounded logit-level gradients. Experiments on real-world benchmarks demonstrate that our method consistently outperforms existing probabilistic forecasting approaches while generating plausible future trajectories and competitive point estimates.

Primary area:
learning on time series and dynamical systems
Keywords:
Probabilistic forecasting, Variational inference, Mixture model, Fractional differencing, Multi-class classification

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_5RuLOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a method, CLS-PF, for probabilistic time-series forecasting. The method builds on mixture sieve model, which reduces the variational inference problem to a standard multi-class classification task, as the decoder is fixed and has no learnable parameters. The method also incorporates a fractional differencing to alleviate distribution shift.

The main contribution is the design of the mixture model for time-series forecasting and corresponding theoretical results demonstrating robustness of the method. The authors also conducted rigorous experiments and ablations which emphasise the importance of fractional differencing and deeply discussed the importance of each component of the model on final results.

Critical strengths
  1. The proposed method is strengthened with the corresponding theoretical claims and proofs
  2. The authors conducted rigorous experiments which demonstrate the improvements on multiple datasets and on various metrics.
  3. In addition, the authors conducted extensive ablation studies and discussed the importance of each component of the model
Critical weaknesses
  1. The proposed method is for one-dimensional outcome only. However, the authors mention it in the conclusions and call it an important direction for future work
  2. The authors motivate this work as a trade-off between expressiveness and computational cost but do not really discuss why mixture-based model was selected instead of some other lightweight probabilistic model
Questions for authors

For me was not very clear if πm\pi_m was implemented using backbone architectures?

2 comments

We thank the reviewer for the positive assessment and encouraging feedback.

W2. Why a mixture-based model?

We chose Gaussian mixtures for their ability to represent multimodal densities and support efficient sampling. By fixing component locations and scale, our formulation reduces mixture density estimation to a simple component classification task. We will clarify this motivation in the revised manuscript.

Q1. Role of backbone architectures

Yes. In πm(⋅;ϕ)\pi_m(\cdot;\phi), ϕ\phi denotes the neural network parameters of the backbone architecture. We will explicitly clarify this in Section 3.1 of the revised manuscript.

Thank you for your answers!

Reviewer_Gn4yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CLS-PF, a probabilistic time-series forecasting framework that combines fractional differencing with classification-based conditional density estimation. The method models the differenced target using a finite Gaussian mixture with fixed component locations, where the mixture weights are learned through a cross-entropy objective derived from a variational lower bound using a reconstruction-optimal hard posterior. The inverse fractional differencing transformation induces history-dependent mixture component locations in the original forecasting space. The paper provides theoretical analyses of cell-level distributional approximation and bounded logit gradients. Experiments on 10 real-world datasets with three neural network backbones demonstrate strong probabilistic forecasting performance, achieving the best CRPS on 8 datasets. Additional ablation studies and sensitivity analyses examine the effects of fractional differencing and key hyperparameters.

Critical strengths
  1. Simple and computationally efficient probabilistic forecasting framework.

The proposed method provides a simple and practical approach to probabilistic time-series forecasting by converting conditional density estimation into a classification task. Its mixture-based formulation enables efficient sampling without requiring iterative denoising procedures. The framework can also be integrated with different neural network backbones, making it potentially useful in practice.

  1. Strong empirical performance across multiple benchmarks.

The paper reports competitive results across 10 real-world datasets and three backbone architectures, achieving the best CRPS on 8 out of 10 datasets. The experiments also include comparisons with several probabilistic forecasting methods, ablation studies, and hyperparameter sensitivity analyses. These results provide encouraging empirical evidence for the effectiveness of the proposed framework.

Critical weaknesses
  1. Limited methodological novelty beyond combining fractional differencing with classification-based density estimation.

The main methodological contribution appears to be defining a classification-based conditional density estimator in the fractionally differenced domain. The variational inference formulation provides a probabilistic interpretation of the classification objective, but it is unclear whether it introduces a substantive methodological advance. In particular, the proposed hard posterior is optimal for the reconstruction term rather than the full ELBO, and the cross-entropy objective follows directly once this posterior is fixed. Therefore, the role of variational inference seems primarily to justify an otherwise straightforward discretization-based classification approach, rather than to introduce a fundamentally new inference method.

  1. Limited evidence for the unique benefits of fractional differencing over simpler history-dependent shifts.

A potentially interesting aspect of the proposed framework is that fractional differencing induces history-dependent mixture component locations in the original forecasting space. However, it remains unclear to what extent the improvements can be attributed specifically to fractional differencing rather than to the introduction of history-dependent shifts more generally.

In particular, setting d=1 yields a simpler residual-based mixture model that predicts first-order increments while retaining time-varying component locations. Although Figures 4(c) and 8(c) compare different differencing orders, the advantages of non-integer orders over d=1 do not appear uniformly substantial across datasets. Moreover, the differencing order is selected through validation-based hyperparameter search, and the preferred order appears to vary across datasets. A more explicit comparison with fixed d=1, as well as simpler alternatives such as moving-average-based or learned residual shifts, would help clarify the unique contribution of fractional differencing to the proposed framework.

Questions for authors
  1. What is the substantive contribution of the variational inference formulation beyond providing a probabilistic interpretation of the classification objective?

In Proposition 2.1, the proposed hard posterior maximizes the reconstruction term rather than the full ELBO. Once this posterior is fixed, the resulting objective reduces directly to standard cross-entropy. Could the authors clarify what methodological or theoretical advantages the VI formulation provides over directly applying a discretization-based classification objective?

  1. Could the authors further clarify the specific benefits of fractional differencing over simpler history-dependent shifts?

I appreciate the ablation and sensitivity analyses already provided, particularly Figures 4(c) and 8(c). However, the additional improvements from non-integer differencing orders over standard first-order differencing d=1 appear somewhat dataset-dependent and are not always substantial. It would be helpful to see a more explicit comparison between the validation-selected differencing orders and fixed d=1 across datasets. Additionally, I wonder whether simpler history-dependent shifts, such as moving-average-based or learned residual shifts, might offer similar benefits within the proposed mixture framework. Any further clarification or supporting evidence would help me better understand the unique role of fractional differencing.

3 comments

We thank the reviewer for the thoughtful feedback and helpful suggestions. We will incorporate the clarifications and additional analyses presented in our responses into the revised manuscript.

Q1. Contribution of the variational inference formulation

We do not claim a new variational inference method, and the reduction to cross-entropy is intended for stable and efficient training. The hard posterior and the full ELBO are discussed in our response to Reviewer DbrJ-W1.

The contribution beyond interpretation is methodological.

(1) The decoder connects cross-entropy training to a continuous predictive distribution. Replacing the Gaussian decoder with a piecewise-uniform one recovers a discretization-based classifier, whose density is piecewise constant; our formulation instead yields a smooth Gaussian-mixture density while keeping the same cross-entropy training.

(2) Proposition 2.1 derives the target assignments from the decoder itself, so that the training targets and the predictive distribution are specified by the same decoder.

(3) This choice translates into empirical performance: using the same trained classifier, Gaussian decoding achieves lower CRPS than the uniform decoding on 9 of 10 datasets.

CRPSCLS-PF(Gaussian)Uniform
BE.287.290
DE.419.422
FR.245.248
NP.220.228
PJM.299.309
ETTh1.077.078
ETTh2.174.181
ETTm1.046.046
ETTm2.125.197
Weather.159.163

Q2-1. Non-integer differencing order

datasetΔ\DeltaCRPSΔ\DeltaRMSE
BE−.006−.027
DE−.016−.022
FR−.004−.004
NP−.015−.010
PJM−.008−.008
ETTh1+.001−.001
ETTh2+.002+.004
ETTm1.000−.001
ETTm2−.024−.018
Weather−.001+.001

We compared tuned dd against fixed d=1d=1. Δ\Delta denotes tuned-minus-fixed performance; negative values indicate improvements over d=1d=1. Although the gains vary across datasets, selecting the differencing order improves CRPS on 7 of 10 datasets, with one tie, and RMSE on 8 of 10 datasets.

ddcount
0.38
0.712
1.02
1.36
1.72
2.00

Additionally, the table summarizes the validation-selected differencing orders across the 30 dataset–backbone settings. Non-integer orders were selected in 28 settings, showing that the results reported in the main text predominantly use fractional rather than first-order differencing.

Q2-2. Ablation study: History-dependent shifts

We evaluate two alternative history-dependent shifts:

  1. moving-average shift (MA): the mean of the most recent WW observations.
  2. cycle-average shift (Cycle): the mean of observations within the lookback window at lags P,2P,…P,2P,\ldots relative to the next prediction step.

Since target discretization blocks gradients to the shift, we replace CycleNet's learnable cycle with a cycle average computed within the lookback window. For each variant, WW or PP is selected from {24,48,72,96,144,168}\{24,48,72,96,144,168\} based on validation performance, matching the six-candidate search budget for the differencing order (FD).

CRPSFDMACycle
BE.287.293.300
DE.419.442.438
FR.245.256.256
NP.220.235.245
PJM.299.446.332
ETTh1.077.079.087
ETTh2.174.180.189
ETTm1.046.047.059
ETTm2.125.154.191
Weather.159.265.491
RMSEFDMACycle
BE.623.617.611
DE.557.627.612
FR.639.617.589
NP.456.466.477
PJM.261.338.280
ETTh1.167.167.188
ETTh2.275.280.293
ETTm1.100.102.128
ETTm2.168.202.251
Weather.019.027.040

Fractional differencing achieves lower CRPS than both moving-average and cycle-average shifts on all 10 datasets, and the best or tied-best RMSE on 8 datasets. These results show that the evaluated alternative history-dependent shifts do not reproduce its benefits, supporting the effectiveness of fractional differencing.

I appreciate the authors' response and additional experiments. The authors' response partially addressed my concerns, but my major concern still remains.

In my view, the core work of the paper lies in incorporating fractional differencing into mixture density estimation to construct a dynamic mixture model. The classification loss, however, is a choice of training objective rather than a necessary component of this model. While the variational formulation connects classification-based training with continuous predictive distributions, I remain unconvinced that this constitutes sufficient methodological novelty to be the central contribution. The authors should further justify the choice of classification loss and demonstrate its advantages over alternative objectives, such as exact maximum likelihood estimation under the same mixture model. The existing comparisons do not isolate the contribution of the classification objective, as the baselines use different predictive distribution parameterizations.

Besides, I believe the current title and narrative place disproportionate emphasis on variational inference, which is not truly in line with the content.

Overall, the novelty of the methodology seems incremental/trivial, and a modification on the Gaussian decoder and training objective does not yield sufficient theoretical contribution in the field.

We thank the reviewer for the continued engagement.

We agree that the dynamic mixture induced by fractional differencing should be the center of the paper. In the revision, we will change the title to "Dynamic Mixture Sieves via Fractional Differencing for Probabilistic Time-Series Forecasting", restructure the narrative around the dynamic mixture sieve, and include the additional experiments from this discussion.

(1) The core of CLS-PF is the dynamic mixture sieve.

CLS-PF integrates a mixture sieve with fractional differencing: its components are fixed on a grid in the differenced space but have time-varying locations in the original space. Moreover, fractional differencing alleviates distribution shift in the target variable. This model is not reproduced by simpler history-dependent shifts: within the same mixture sieve, moving-average and cycle-average shifts yield higher CRPS than fractional differencing on all 10 datasets, and non-integer orders were selected in 28 of the 30 dataset–backbone settings.

(2) The cross-entropy follows from the sieve setting and separates estimation from decoding.

Since the components are fixed on the grid, the reconstruction term of the ELBO is constant with respect to the network, so with the one-hot posterior fixed at the nearest bin, maximizing the ELBO reduces to cross-entropy classification with fixed targets. This reduction is specific to the sieve: with a learned decoder, the reconstruction term would change during training.

The nearest bin is reconstruction-optimal for every decoder satisfying Proposition 2.1, regardless of β\beta, so a single training serves all such decoders and values of β\beta. The exact likelihood depends on both, so tuning β\beta requires one training per candidate, whereas the cross-entropy requires only one.

Under the same dynamic mixture sieve and the same β\beta, the cross-entropy attains a CRPS comparable to the exact likelihood (standard deviation over seeds in parentheses; Δ\Delta = cross-entropy minus exact likelihood).

CRPScross-entropyexact likelihoodΔ\Delta
BE.287 (.003).286 (.003)+.001
DE.419 (.003).416 (.004)+.003
FR.245 (.004).245 (.003).000
NP.220 (.002).228 (.003)−.008
PJM.299 (.004).299 (.003).000
ETTh1.077 (.001).075 (.001)+.002
ETTh2.174 (.003).172 (.002)+.002
ETTm1.046 (.000).046 (.001).000
ETTm2.125 (.003).141 (.003)−.016
Weather.159 (.002).154 (.004)+.005

Thus, the cross-entropy shows no systematic loss in performance relative to the exact likelihood, while a single training serves all decoders and values of β\beta.

Reviewer_DbrJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a probabilistic forecasting model called CLS-PF for predicting distribution at next time step. Authors propose the following:

  1. compute the fractionally differenced target to reduce distributional shift
  2. quantize the differenced target into M many bins
  3. assume fixed parameter gaussian distribution for bins
  4. entire distribution is considered as a mixture of Gaussians
  5. Train for the cross entropy on bin probabilities

Authors demonstrated that training for ELBO of the GMM with a specific posterior is equivalent to training for the cross-entropy of the component probabilities.

Experiments are performed on 10 standard time series forecasting and evaluated with CRPS and QICE. Proposed CLS-PF performs better in 8/10 datasets for CRPS, 7/10 for QICE when averaged over three neural architectures.

Critical strengths
  1. CLS-PF is a simple and cheap baseline for probabilsitic forecasting. It can be applied on top of any point forecasting model by just changing the output heads and input processing.
  2. Experimental results are better compared to many complex baselines. Ablation studies demonstrate that the fractional binning is useful.
Critical weaknesses
  1. The claimed contribution of Variational Inference framework does not hold because the cross-entropy loss is achieved by choosing posterior that minimizes only the reconstruction loss. ELBO gets tight when posterior is optimized for full ELBO. With uniform distribution on the bin, optimizing cross entropy would be same as optimizing NLL and Variational Inference is unnecessary. Hence, the model is effectively trained for uniform distributed bins rather than Gaussians, yet the samples are drawn from Gaussian Mixture.

  2. It is not clear which components (cross entropy loss or fractional differencing) drive the gains. In Table 8, CLS-PF wins in 5 datasets by thin margin and loses in 4 datasets when all are trained after fractional differencing. Furthermore, the differencing hyperparameter d is not tuned for baselines.

Questions for authors
  1. The training of proposed model using cross entropy is similar to many existing approaches such as tabpfn. The new thing is mainly Gaussian distribution on bins. Can authors provide an experiment comparing Gaussian and Uniform distributions on bins?

  2. Fractional differencing is similar to removing trend as in DLinear and cycles of cyclenet in the spirit. There is no clear argument that the fractional differencing would perform better than the existing approaches. An ablation would help to clarify.

4 comments

We thank the reviewer for the careful evaluation and constructive criticism. We will incorporate the clarifications and additional analyses presented in our responses into the revised manuscript.

W1. Validity of the variational formulation and Q1. Gaussian vs. Uniform bin distributions

  1. We clarify that selecting the posterior by the reconstruction term is consistent with the variational formulation.

Let yy denote the target and xx the lookback window, and define m∗=arg⁡min⁡m∣y−θm∣=arg⁡max⁡mN(y∣θm,β2)m^* = \arg\min_m |y-\theta_m| = \arg\max_m \mathcal{N}(y\mid\theta_m,\beta^2) and q(θm∣y)=I(m=m∗)q(\theta_m\mid y) = \mathbb{I}(m=m^*). Jensen's inequality gives a valid bound (Eq. (3)):

\log p(y\mid x;\phi) \ge \mathrm{ELBO}(m^* ) = \log\mathcal{N}(y \mid \theta_{m^* },\beta^2) + \log\pi_{m^* }(x;\phi). \end{aligned}$$ **Choosing the posterior by the reconstruction term is itself a variational step.** Our procedure can be interpreted as one sweep of variational EM over the one-hot family, initialized with a uniform prior. In the E-step, the KL term under the uniform prior is constant for every one-hot $q$, as in VQ-VAE [1], so maximizing the ELBO reduces to maximizing the reconstruction term, whose solution is $m^{* }$. In the M-step, the reconstruction term is constant in $\phi$, so maximizing the ELBO reduces to minimizing the cross-entropy $- \log\pi_{m^* }(x;\phi)$. **Optimizing the posterior for the full ELBO improves it by no more than the cross-entropy.** Maximizing the full ELBO over the one-hot family, including the prior term $\log\pi_m(x;\phi)$, gives $\widetilde{m} = \arg\max_m\{\log\mathcal{N}(y\mid\theta_m,\beta^2)+\log\pi_m(x;\phi)\}$. Since $\mathcal{N}(y\mid\theta_{\widetilde{m}},\beta^2)\le\mathcal{N}(y\mid\theta_{m^* },\beta^2)$, we have $$\begin{aligned} 0 \le \mathrm{ELBO}(\widetilde{m})-\mathrm{ELBO}(m^* ) \le \log\frac{\pi_{\widetilde{m}}(x;\phi)}{\pi_{m^* }(x;\phi)} \le -\log\pi_{m^* }(x;\phi), \end{aligned}$$ so the improvement is zero whenever the nearest component $m^*$ is also the most probable component under the prior, and is otherwise bounded by the cross-entropy being minimized. Tightness is not required for the bound to be valid, and the one-hot family is a modeling choice of the variational family for which the bound remains valid. Therefore, our method falls within the variational inference framework, with the posterior inferred over this family and the prior learned by maximizing the same ELBO, and the claimed variational derivation of the cross-entropy loss holds. 2. **Cross-entropy training is consistent with the Gaussian decoder, not only with uniform bins.** We agree that with uniform bins the cross-entropy coincides with the negative log-likelihood. However, this does not mean the model is trained only for uniform bins. With the target assignments derived from the Gaussian decoder (Proposition 2.1), the same cross-entropy maximizes a lower bound on the Gaussian-mixture log-likelihood, so the objective is shared by the Gaussian and uniform decoders rather than specific to uniform bins. The variational formulation is unnecessary for uniform decoding, but it is what makes Gaussian sampling consistent with the training objective rather than mismatched with it. | **CRPS** | CLS-PF(`Gaussian`) | `Uniform` | |---|---:|---:| | BE | **.287** | .290 | | DE | **.419** | .422 | | FR | **.245** | .248 | | NP | **.220** | .228 | | PJM | **.299** | .309 | | ETTh1 | **.077** | .078 | | ETTh2 | **.174** | .181 | | ETTm1 | **.046** | **.046** | | ETTm2 | **.125** | .197 | | Weather | **.159** | .163 | To evaluate the decoder choice empirically, we compared Gaussian and uniform decoding using the same trained classifier. Gaussian decoding achieves lower CRPS on 9 of 10 datasets and ties on ETTm1. - Reference [1] Van Den Oord, Aaron, and Oriol Vinyals. "Neural discrete representation learning." Advances in neural information processing systems 30 (2017).

Q2. Ablation study: History-dependent shifts

We evaluate two alternative history-dependent shifts:

  1. moving-average shift (MA): the mean of the most recent WW observations.
  2. cycle-average shift (Cycle): the mean of observations within the lookback window at lags P,2P,…P,2P,\ldots relative to the next prediction step.

Since target discretization blocks gradients to the shift, we replace CycleNet's learnable cycle with a cycle average computed within the lookback window. For each variant, WW or PP is selected from {24,48,72,96,144,168}\{24,48,72,96,144,168\} based on validation performance, matching the six-candidate search budget for the differencing order (FD).

CRPSFDMACycle
BE.287.293.300
DE.419.442.438
FR.245.256.256
NP.220.235.245
PJM.299.446.332
ETTh1.077.079.087
ETTh2.174.180.189
ETTm1.046.047.059
ETTm2.125.154.191
Weather.159.265.491
RMSEFDMACycle
BE.623.617.611
DE.557.627.612
FR.639.617.589
NP.456.466.477
PJM.261.338.280
ETTh1.167.167.188
ETTh2.275.280.293
ETTm1.100.102.128
ETTm2.168.202.251
Weather.019.027.040

Fractional differencing achieves lower CRPS than both moving-average and cycle-average shifts on all 10 datasets, and the best or tied-best RMSE on 8 datasets. These results show that the evaluated alternative history-dependent shifts do not reproduce its benefits, supporting the effectiveness of fractional differencing.

W2. Gains from the cross-entropy loss

datasetCLS-PFPPMWTADDIMCRPSMLE
BE.287.527.375.450.291.321
DE.419.620.456.850.428.431
FR.245.359.266.451.252.256
NP.220.461.286.411.232.226
PJM.299.391.311.536.295.298
ETTh1.077.098.077.083.082.079
ETTh2.174.252.178.225.185.169
ETTm1.046.054.047.057.047.046
ETTm2.125.129.129.202.101.109
Weather.159.360.285.187.150.256

We additionally tuned the differencing order dd for each baseline over the same search space as CLS-PF, selecting it based on validation performance. CLS-PF still achieves the best or tied-best performance on 6 of 10 datasets.

Together with Table 8, these results show that CLS-PF remains competitive both in the same forecasting space and with individually tuned differencing orders. This supports the effectiveness of the cross-entropy loss.

I appreciate the detailed response to my comments and the additional experiments. My main concerns remain:

  1. The variational formulation theory does not justify a Gaussian decoder. Any unimodal symmetric distribution will do the job, and the rationale for using a Gaussian is unclear. When comparing Uniform and Gaussian, the differences are very small in many datasets.

  2. The rebuttal argues that the current nearest-bin posterior is ELBO-optimal, but this is true only if the prior is uniform. However, the learned prior may not be uniform, and hence the bin that maximizes the full ELBO can differ from the nearest bin. Since the assignments are never updated, training does not optimize the ELBO.

  3. When the baselines use the same tuned differencing, the gains of CLS-PF are very small compared to the CRPS baseline. Without std, it is not clear if the differences are statistically significant.

Therefore, I keep my score.

We thank the reviewer for the follow-up.

The core of CLS-PF is the integration of a mixture sieve with fractional differencing, which yields a dynamic mixture sieve and alleviates distribution shift in the target variable: its components are fixed on a grid in the differenced space but have time-varying locations in the original space.

This sieve setting addresses the first two concerns: the Gaussian is the leading case of a mixture sieve for targets with unbounded support, and since the components are fixed on the grid, the reconstruction term of the ELBO is constant with respect to the network.

1–2. Decoder choice and ELBO optimization in the sieve setting

(1) The sieve setting justifies the Gaussian decoder and reduces the ELBO to cross-entropy.

Since the fractionally differenced targets have unbounded support, we choose normal densities, which are the leading case of continuous mixture estimators (Fox et al., 2016).

Because the decoder is fixed and known in the sieve, with the one-hot posterior fixed at the nearest bin, maximizing the ELBO reduces to estimating the mixing weights by cross-entropy. This reduction is specific to the sieve: with a learned decoder, the reconstruction term would change during training.

We agree that the nearest bin maximizes the full ELBO only under a uniform prior; in practice, fixing it costs little, as maximizing the exact likelihood at the same β\beta yields comparable CRPS (see our response to Reviewer Gn4y). In return, since the nearest bin does not depend on the decoder or β\beta, a single training serves all of them, whereas the exact likelihood must be retrained for each.

The variational formulation does not select the decoder, and it is not intended to: the decoder belongs to the mixture model, and the variational formulation concerns its estimation. Other decoders also satisfy Proposition 2.1 (lines 167-168), and as shown below, the Gaussian performs consistently well.

(2) The advantage is consistent across backbones and arises from multimodality.

backbonedecoderBEDEFRNPPJMETTh1ETTh2ETTm1ETTm2Weather
CrossLinearGaussian.299.429.252.224.312.078.180.045.135.162
Uniform.304.432.255.233.319.078.177.046.210.162
TSMixerGaussian.288.423.245.216.295.077.164.047.116.156
Uniform.290.424.246.223.306.077.180.049.184.163
TimeXerGaussian.274.405.237.221.289.077.179.045.124.159
Uniform.275.409.243.229.300.078.187.045.198.165

Although the differences are small on several datasets, Gaussian decoding is better in 25 of the 30 dataset–backbone pairs and tied in 4.

In a density estimation example with 0.6N(0,0.62)+0.4N(μ,0.62)0.6 \mathcal{N} (0, 0.6^2) + 0.4 \mathcal{N} (\mu, 0.6^2), we decode the same estimated weights with each decoder (β\beta selected on validation data):

μ\mu0.00.61.21.82.43.0
num peaks111222
Uniformexcess CRPS0.280.270.220.230.340.54
bias$^2$0.180.170.120.130.240.43
Gaussianexcess CRPS0.490.470.320.160.190.33
bias$^2$0.190.170.120.070.080.14

(averaged over 300 replications, CRPS ×103\times 10^{3})

Uniform decoding is better for unimodal distributions, whereas Gaussian decoding is better once the distribution becomes bimodal, the setting our sieve targets, mainly through a lower bias.

3. Baselines with the tuned differencing order

CRPSCLS-PFCRPS-baseline
BE.287 (.003).291 (.005)
DE.419 (.003).428 (.011)
FR.245 (.004).252 (.005)
NP.220 (.002).232 (.007)
PJM.299 (.004).295 (.005)
ETTh1.077 (.001).082 (.001)
ETTh2.174 (.003).185 (.009)
ETTm1.046 (.000).047 (.001)
ETTm2.125 (.003).101 (.004)
Weather.159 (.002).150 (.005)

Although the CRPS-baseline directly optimizes the evaluation metric, CLS-PF attains a lower mean CRPS on 7 of 10 datasets with an equal or smaller standard deviation over seeds on every dataset, indicating that the cross-entropy objective is comparable to direct CRPS training and stable across seeds.

Comments on the paper

1 comment

Dear Reviewers,

The authors have now submitted their responses. Please review them and indicate whether your concerns have been addressed. Please also update your scores if the responses have changed your assessment.

Thank you.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.