acceptodds
Under review as a conference paper at ICLR 2027

Bayes Averaging or Over-Smoothing? Excessive Contraction in Time Series Forecasting

Abstract

Time series forecasting models often produce predictions that vary much less than observed futures, a behavior commonly interpreted as over-smoothing. Yet under squared loss, even the Bayes-optimal predictor averages over plausible futures, reducing variation under target uncertainty. Reduced variation alone therefore does not establish a modeling deficiency. We develop a Bayes-referenced framework for variation *across* predictions for similar inputs and *within* individual forecast trajectories. For both perspectives, we derive an observable criterion that relates predictive amplitude to target alignment and identifies excessive contraction along the model's current prediction direction, without directly estimating the Bayes-optimal predictor. Using the proposed contraction measures, we examine eight forecasting models across six benchmark datasets and confirm that predictive contraction is widespread and that stronger contraction is consistently associated with larger forecasting error. The Bayes-referenced analysis, however, reveals two reversals in this interpretation. First, observed contraction does not imply excessive contraction: sample-wise contrasts generally favor further attenuation, while temporal contrasts span both excessive- and insufficient-contraction regimes. Second, these aggregate tendencies themselves conceal heterogeneous window-level regimes. Within the same model–dataset pair, some windows favor attenuation while others favor amplification, and, in both regimes, larger departures from unit scaling tend to accompany higher error. A contrast-error decomposition further shows that optimal uniform rescaling captures only part of the pooled contrast error, motivating a finer analysis of alignment and window-level heterogeneity. Overall, our results establish a Bayes-referenced view of predictive contraction that distinguishes over-smoothing from uncertainty-induced attenuation and reveals heterogeneous scaling behavior across individual forecasts.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Predictive Contraction, Over-Smoothing, Bayes-Optimal Prediction, Regression to the Mean, Forecasting Behavior, Model Diagnostics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SrmrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies predictive contraction in time-series forecasting, in which predictions show less variation than the observed targets. Its central question is whether this reduction reflects appropriate averaging over uncertain futures or excessive contraction introduced by the forecasting model, considering variation across samples with similar observed histories and variation within individual forecast trajectories. The authors define contrast operators and contraction ratios for these two perspectives. They then derive an observable criterion that combines predictive amplitude with target alignment to identify the optimal scalar rescaling of predicted contrasts. Under the stated assumptions, this coefficient is equivalent to the coefficient obtained using a Bayes-optimal conditional-mean reference, without requiring that reference to be estimated.

Experiments across eight models and six benchmark datasets show that contraction is widespread (Figure 3). However, accounting for alignment changes its interpretation: sample-wise contrasts generally favor further attenuation, whereas temporal contrasts exhibit both amplification and attenuation regimes (Figure 4). The paper also examines retrospective scaling preferences across individual windows (Section 4.3) and shows that global temporal rescaling produces limited forecasting improvements, even with test-oracle fitting (Table 4). Its conclusion is that reduced variation relative to observed targets alone is insufficient evidence of excessive smoothing.

Critical strengths

The paper distinguishes contraction caused by optimal averaging under uncertainty from excessive contraction. This is valuable because reduced predictive variation shouldn't simply be interpreted as a model deficiency without considering whether that variation is predictable.

The paper defines an observable criterion that combines amplitude and target alignment to identify optimal scalar rescaling without estimating the Bayes reference. Analysis across eight models and six datasets demonstrates that amplification is not universally beneficial (Figure 4), providing useful evidence for interpreting forecasting behavior.

Critical weaknesses

The contribution beyond prior work is insufficiently discussed and established. A dedicated discussion of existing theoretical and empirical studies on forecast rescaling, error decomposition, and uncertainty-induced contraction would help clarify which aspects are established and what new insights this work provides. This distinction is particularly important for assessing the novelty and significance of an analysis paper. Please compare the paper’s main findings directly with prior theoretical or even empirical work on rescaling, error decomposition, and uncertainty-induced contraction. Explain how your findings either confirm existing knowledge, challenge it, or provide new insights.

Model error is not really or clearly distinguished from irreducible uncertainty. A controlled validation of the diagnostic could answer this! In fact, the paper would benefit from synthetic experiments where the true conditional mean is known. Comparing this optimal predictor with predictions whose amplitude is intentionally reduced or whose shape is distorted would test whether the diagnostic correctly identifies excessive contraction. These experiments would also clarify which window-level patterns reflect model errors and which can arise from unpredictable noise, even for an optimal predictor.

Questions for authors

Please address the concerns raised in the weaknesses section.

4 comments

We appreciate the reviewer’s recognition of the importance of distinguishing excessive model contraction, as well as the value of an observable criterion that avoids explicitly estimating the Bayes-optimal predictor. The constructive suggestions have helped us further clarify the relationship between our contributions and prior work, and strengthen the validation and interpretation of our findings. We address the two concerns below.

W1: On the novelty and contributions relative to prior work

Regarding the theoretical contribution, our analysis is grounded in established statistical results, including least-squares forecast recalibration (Mincer and Zarnowitz, 1969), forecast error decomposition (Theil, 1966), and the orthogonality property of conditional expectation. We have made these works explicit in the revised manuscript. The central theoretical result of our work is an observable characterization of predictive contraction under conditional uncertainty. This characterization enables a purely post hoc assessment of Bayes-referenced contraction, requiring only model predictions and ground-truth targets, without estimating the Bayes-optimal predictor. We establish this characterization separately for sample-wise and temporal contrasts, providing a unified analytical framework for examining predictive contraction along these two dimensions.

Regarding the empirical contribution, prior studies on over-smoothing have identified the attenuation of predictive dynamics as a potential modeling limitation and proposed approaches to preserve meaningful future variation (Le Guen and Thome, 2019; Sun et al., 2025; Yu et al., 2026; Zhang et al., 2026). In contrast, Green et al. (2026) demonstrate that reduced predictive variation can also arise from Bayes-optimal averaging under conditional uncertainty.

Our empirical findings help reconcile these perspectives. We confirm that predictive contraction is widespread and positively associated with forecasting error, consistent with the concerns raised by over-smoothing studies. We find that, after accounting for target alignment, sample-wise predictions predominantly favor further attenuation, whereas temporal predictions exhibit both excessive- and insufficient-contraction regimes, challenging the widely adopted assumption that predictive contraction reflects a modeling deficiency and needs to be rectified.

Furthermore, while existing studies primarily characterize over-smoothing through aggregate benchmark performance, we further examine retrospective scaling preferences at the individual-window level. This finer-grained analysis reveals substantial heterogeneity within the same model–dataset pair: some windows favor attenuation, while others favor amplification, and larger deviations from unit scaling in either direction tend to accompany higher forecasting error. This heterogeneity also helps explain why a single global temporal rescaling yields only modest gains, even under test-oracle fitting.

W2: On distinguishing model error from irreducible uncertainty

We have added controlled synthetic experiments in the Appendix A.5, using an AR(2) process with latent frequency uncertainty, for which the Bayes-optimal conditional mean can be computed exactly. This enables direct validation of the proposed diagnostic and an examination of the respective contributions of model deviations and irreducible uncertainty.

First, we validate the diagnostic through controlled interventions. Scaling the Bayes-optimal predictions by a factor c yields the expected Bayes-referenced coefficient βˉ⋆=1/c\bar{\beta}^{\star}=1/c, correctly distinguishing contraction from amplification. We further apply an energy-preserving shape transformation, which alters prediction alignment without changing amplitude. The resulting scaling coefficients deviate from one, confirming that the diagnostic captures both amplitude and alignment effects (Figure 12).

Second, we examine the sources of window-level heterogeneity. As raised by the reviewer, irreducible uncertainty can produce window-level heterogeneity even under the Bayes-optimal predictor. Nevertheless, in our controlled setting, learned predictors exhibit additional heterogeneity in their Bayes-referenced scaling preferences, accounting for 33%–62% of the variance. This demonstrates that model deviations can contribute substantially to the observed heterogeneity alongside irreducible uncertainty.

Thank you again for your thoughtful review and constructive suggestions. We would like to follow up and wonder if our responses have addressed your concerns. For ease of reference, we summarize below the corresponding revisions and newly added experiments. The revised passages are highlighted in blue in the updated manuscript.

W1: On the novelty and contributions

We have expanded the Introduction (Pages 2–3) to clarify the relationship between prior studies on over-smoothing and Bayes-optimal averaging under conditional uncertainty, and to better position our theoretical and empirical contributions.

The theoretical foundation and the proposed observable criterion are further clarified in Section 4.1 (Page 6), while Section 4.4 (Page 8) clarifies the contrast-error decomposition based on Theil (1966), explaining the limitations of uniform amplitude correction.

W2: On distinguishing model error from irreducible uncertainty

We have added controlled synthetic experiments in Appendix A.5 (Pages 16–19).

  • Section A.5.1 (Page 16): Constructs an AR(2) process with latent frequency uncertainty, for which the Bayes-optimal conditional mean is known.

  • Section A.5.2 (Pages 17–18): Validates the proposed diagnostic through controlled amplitude and energy-preserving shape interventions. Figure 12 summarizes their effects on predictive amplitude and target alignment.

  • Section A.5.3 (Pages 18–19): Decomposes window-level scaling coefficients into conditional-mean scaling preferences and residual fluctuations due to irreducible uncertainty. Table 6 quantifies their respective contributions, showing that the former accounts for 32.5%–61.6% of the combined variance in window-level scaling coefficients across the learned predictors.

We would appreciate any further feedback or guidance you may have, and would be willing to address any remaining questions or concerns.

Dear authors,

Thank you for your detailed responses to my questions and also to the points raised by the other reviewers. The discussion has clarified the paper’s contributions and addressed my main concerns. I now have a more positive assessment of the work and have raised my score.

We sincerely appreciate your positive reassessment and your decision to raise your rating. Thank you for the time and thoughtful consideration devoted to reviewing our work and engaging in this discussion. Your insightful feedback prompted us to conduct deeper analyses and clarify key aspects of our methodology, which helped us further strengthen our manuscript. We are pleased that our revisions have addressed your main concerns, and remain grateful for the valuable feedback for improving this work.

Reviewer_Zs1HOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors focus on the observation that predictions often show reduced variance compared to the true future values, a problem identified as oversmoothing, and try to connect that to examine if this behavior connects to model deficiency. They leverage the fact that under squared loss, the Bayes-optimal predictor averages possible future trajectories, which vary less than true future values. A Bayes-referenced framework is developed for measuring variation across predictions for similar inputs (sample-wise) and individual forecast trajectories (temporal). This is achieved via an observable criterion that relates predictive amplitude to target alignment and measures contraction with the direction of the model’s prediction. The authors test 8 forecasting models on 6 benchmark datasets to show that stronger contraction is related to larger forecasting error. The overall analysis introduces a Bayesian view of prediction contraction that separates over-smoothing from attenuation due to uncertainty and shows how scaling behavior differs between forecasts.

Critical strengths
  1. Significance of contribution: The paper sheds light on a very interesting concept in time series forecasting. Specifically, contraction is studied from different angles when it comes to true targets and Bayesian-optimal forecasts, and a new simple criterion to measure this behavior is introduced. This helps us question the common assumption behind methods that target less smooth forecasts. Figure 1 provides a strong motivation for the study.
  2. Clear experimental setup: The experimental evaluation covers common benchmark datasets in the forecasting field and several commonly employed baselines. Window-level analyses, error decomposition, and sensitivity studies, accompanied by nice visualizations, complete the presentation of important results.
Critical weaknesses
  1. Performance generalization and significance: Lookback and horizon lengths are fixed to 96, while contraction is supposed to change with horizon lengths. Additionally, no standard deviation across runs is depicted, which makes contraction stability hard to assess. Also, results are limited to point forecasts.
  2. Connections of the provided theoretical analysis: The authors fail to clearly make connections between the main formulations and the relevant literature for points such as the scaling coefficient, the error decomposition, and the properties of the conditional mean. The novelty of the theoretical analysis should be justified more. It is also unclear whether the employed train and test subsets for the real-world datasets follow the same distributions, as assumed.
  3. Noise seems to affect results: For similar inputs, target differences are mostly noise, thus any prediction variation seems large. Additionally, for targets with high variance, contraction and error increase, thus are correlated even for a good forecasting method. This means findings reflect the chosen measures.
Questions for authors
  • Authors should provide justifications for the novelty of the provided theoretical analysis, as explained above.
  • Can the authors explain how train-test distributions affect results and how performance varies across seeds and different lookback/history lengths?
  • Are the findings dependent on noise and the selected measures?
  • Since temporal contraction is related to the attenuation of high-frequency components, can the authors discuss frequency-based training objectives?
2 comments

We thank the reviewer for recognizing the significance of studying predictive contraction from both realized-target and Bayes-optimal perspectives, as well as the value of our proposed criterion and empirical analysis. We have addressed the remaining concerns, as detailed below.

W1 / Q2. Performance Generalization and Significance

Forecast horizon. Our original manuscript included a sensitivity analysis of input and forecast lengths in Appendix A.1 (Figure 10). Specifically, we adopted a joint-scaling setting, where P=F∈96,192,336,720P=F\in{96,192,336,720} following prior work (e.g., Yi et al., 2023).

To further isolate the effect of forecast horizon, we additionally considered another widely adopted setting in long-term time series forecasting, which fixes the input length at 96 while varying the forecast horizon over 96,192,336,720{96,192,336,720} (e.g., Wu et al., 2021; Liu et al., 2024). Under this setting, we also expanded the experimental coverage by including two models (DLinear and iTransformer) and two datasets (ETTm1 and Electricity). The results show that predictive contraction remain broadly consistent.

Stability across runs/seeds. Our reported results are already averaged over multiple runs, following the evaluation protocol adopted in prior work. To further examine the stability of our findings across random seeds, we additionally conducted experiments with three seeds (2021, 2022, and 2023) for all 64 model–dataset pairs. As reported in Appendix A.1 (Table 1), the median standard deviations are 0.0043 and 0.0054 for RsampR_{\mathrm{samp}} and RtempR_{\mathrm{temp}}, and 0.0153 and 0.0108 for βsamp\beta_{\mathrm{samp}} and βtemp\beta_{\mathrm{temp}}, respectively. The results demonstrate that our main findings remain consistent across different random seeds.

Point forecasting. Our focus on point forecasting is intentional. Under squared-error loss, the Bayes-optimal point forecast is the conditional expectation, whose predictive variation can be attenuated by averaging over uncertain future outcomes. This Bayes-averaging effect is central to the phenomenon investigated in our work and motivates our point-forecast formulation. Extending the analysis to probabilistic forecasting would require a different formulation at the predictive-distribution level, which is beyond the scope of the current study.

W2 / Q1. Connections to Existing Theory and Theoretical Novelty

We have revised the manuscript to clarify the theoretical foundations of our analysis and its connections to existing literature, including classical least-squares forecast recalibration (Mincer and Zarnowitz, 1969), error decomposition (Theil, 1966), and the orthogonality property of conditional expectation.

Our contribution lies in connecting these established results to develop an observable, Bayes-referenced criterion for assessing excessive contraction. Specifically, we establish the equivalence, in expectation, between optimal rescaling toward observed contrasts and rescaling toward the corresponding Bayes-optimal contrasts. Combined with the amplitude–alignment relationship β=A/R\beta=A/R, this allows us to assess excessive contraction in both sample-wise and temporal contrast spaces without explicitly estimating the Bayes-optimal predictor.

W2 / Q2. Train–Test Distributional Assumptions

Our theoretical analysis does not require identical training and test distributions. In Section 4.1, the trained predictor is held fixed, while the Bayes-optimal reference and all relevant expectations are defined under the evaluation distribution. The orthogonality property of conditional expectation, and consequently the equivalence between optimal rescaling toward observed and Bayes-optimal contrasts, holds under this evaluation distribution regardless of the distribution from which the predictor was trained.

W3 / Q3. Influence of Noise and Conditional Uncertainty

Different targets associated with similar inputs may arise from realization noise or unobserved exogenous factors, both contributing to conditional uncertainty and can affect the observed contraction ratio RR, even for a Bayes-optimal predictor. This is precisely the motivation for introducing our Bayes-referenced criterion in Section 4.

The Bayes-referenced scaling coefficient β\beta indicates whether predictive variation is excessive. By the orthogonality property of conditional expectation, a Bayes-optimal contrast predictor satisfies E⟨P,T⟩=E∣P∣2\mathbb{E}\langle P,T\rangle=\mathbb{E}|P|^2, yielding β=1\beta=1 whenever its contrast energy is nonzero. Consistently, the empirical β\beta values across our real-world experiments remain within a reasonable range.

Q4:Frequency-based training objectives

Temporal contraction is related to spectral energy, but not limited to high-frequency attenuation. By Parseval's identity, temporal contraction ratio can be expressed in the frequency domain as

\frac{\sum_{f\ne 0} E_f^P} {\sum_{f\ne 0} E_f^T} =1- \frac{\sum_{f\ne 0}(E_f^T-E_f^P)} {\sum_{f\ne 0}E_f^T} ,$$ where $E_f^P$ and $E_f^T$ denote the pooled prediction and target energies at frequency $f$, respectively. Prior studies often associate temporal contraction with the attenuation of high-frequency components. However, as indicated by the above formulation, observed temporal contraction reflects energy deficits across all non-DC frequencies. For further examination, we decomposed the aggregate spectral energy deficit into low- and high-frequency components under the commonly used forecasting setting. Across six datasets and two forecasting models (DLinear and iTransformer), the lowest 25% of non-DC frequencies account for 53.6%–94.3% of the total energy deficit across all 12 model–dataset pairs. These results reveal that low-frequency components also contribute substantially to observed temporal contraction, accounting for the majority of the total energy deficit in all evaluated settings. Regarding frequency-based training objectives, such methods may improve the learning or preservation of frequency-specific structures, but recovering spectral energy does not necessarily imply reduced excessive contraction. Under squared-error loss, appropriate predictive variation depends on its alignment with the predictable target structure. Our Bayes-referenced criterion therefore provides a complementary perspective for evaluating whether frequency-based objectives alleviate excessive contraction, beyond their effects on observed spectral energy.
Reviewer_XsBiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines whether reduced variation (i.e., over-smoothing) in time series predictions reflects a modeling flaw or legitimate Bayes-optimal averaging under squared error loss. Using contrast operators, the authors evaluate predictive variation both across nearby input windows (sample-wise contrasts) and within individual forecast trajectories (temporal contrasts).

The contributions are threefold. First, the authors derive an observable Bayes-referenced criterion that identifies excessive contraction without estimating the unobserved Bayes-optimal predictor. Second, this paper shows that while reduced variation relative to realized targets is widespread and correlates with higher error, sample-wise predictions generally require further attenuation, while temporal predictions span both excessive and insufficient contraction regimes. Finally, the authors analyze window-level heterogeneity where attenuation and amplification needs coexist within the same model-dataset pair, and decompose the contrast error to show that directional alignment error dominates global amplitude mismatch.

The authors evaluate existing forecasting architectures across benchmark datasets under the TFB protocol. A test-oracle global temporal rescaling experiment is conducted. Sensitivity analyses are also provided across neighborhood quantiles, distance metrics, sampling strides, and sequence lengths.

Critical strengths
  • By leveraging the properties of conditional expectation under squared error loss, the authors evaluate forecast amplitude against the unobserved Bayes-optimal predictor directly from realized target data. This is important because it yields a hyperparameter-free diagnostic that compares directional target alignment with observed amplitude contraction. In this way, it overturns the widespread assumption that reduced forecast variation automatically implies over-smoothing, showing instead that predictions across similar inputs generally require further attenuation due to low directional accuracy.

  • The authors decompose normalized contrast error into the component removable by uniform amplitude rescaling and the residual error governed by directional alignment. This clean separation is beneficial because it explains both why global temporal rescaling gives negligible accuracy gains even under a test-set oracle, and why retaining greater sample-wise variation hurts multivariate models like iTransformer on noisy series, while helping on structured series.

Critical weaknesses
  • The core theoretical bridge, rescaling toward observed targets matches rescaling toward the unobserved Bayes-optimal predictor, relies on taking expectations so that unpredictable target noise cancels out (i.e., E[T−T∗∣G]=0\mathbb{E}[T - T^* \mid \mathcal{G}] = 0). Note that, The identity E⟨P,T∗⟩=E⟨P,T⟩\mathbb{E}\langle P, T^*\rangle = \mathbb{E}\langle P, T\rangle holds only in expectation. However, Section 4.3 drops the expectation and computes window-level coefficients βk,i:=⟨pik,tik⟩/∥pik∥22\beta_{k,i} := \langle p_i^k, t_i^k \rangle / \Vert{}p_i^k\Vert{}_2^2 directly against single noisy target realizations, while still framing the findings as "Bayes-referenced" instance-level heterogeneity. Without expectation pooling, realization noise does not vanish, this conflates single-realization noise with genuine Bayes-referenced heterogeneity.

  • Although the reviewer appreciates the diversity of experiments and analyses presented in the manuscript, concerns remain regarding the selection and coverage of datasets. Several referenced studies evaluate methods on more benchmark datasets for time series forecasting. For instance, the Solar dataset has been used in both DUET (Qiu et al., 2025) and iTransformer (Liu et al., 2024); the Weather dataset has also been considered in these prior studies and also PDF (Dai et al., 2024). The omission of these commonly adopted benchmarks raises critical concerns over potential cherry-picking bias in dataset selection.

Questions for authors

The authors are encouraged to address the issues mentioned in the Weaknesses section in their response.

An additional question: to ensure reproducibility, will the authors make the source code for the experiments publicly available in the camera-ready version?

1 comment

We thank the reviewer for recognizing the value of our Bayes-referenced diagnostic and the decomposition of contrast error into amplitude and directional components. We address the remaining concerns below.

W1: On the expectation-level equivalence and window-level analysis

The analysis in Section 4.3 serves to characterize the heterogeneity of scaling preferences across individual windows, where the window-level coefficients are intended to describe scaling behavior at the realization level, rather than to serve as Bayes-referenced measures of instance-level model errors or estimates of window-specific Bayes coefficients.

Nevertheless, the original presentation did not make this distinction sufficiently explicit. To clarify the scope and interpretation of this analysis, we have revised the Abstract, Introduction, Section 4.3, and Conclusion. We have also added an exact decomposition in Appendix A.5.3:

βk,i=βk,i⋆+ηk,i,\beta_{k,i}=\beta_{k,i}^{\star}+\eta_{k,i},

where βk,i⋆\beta_{k,i}^{\star} represents the scaling preference relative to the conditional mean, and βk,i\beta_{k,i} is a residual projection with zero conditional mean. This decomposition separates the Bayes-referenced component from the variability introduced by individual target realizations, clarifying the interpretation of our window-level results.

W2: On dataset coverage and potential selection bias

While Solar and Weather were already included in the original manuscript's broader evaluation in Section 4.4, alongside Exchange, PEMS04, and PEMS08, they were not part of the main contraction analysis due to presentation concerns. We have now incorporated Solar and Weather into the main analysis, extending the original evaluation.

The expanded results preserve the overall directional pattern: sample-wise predictions predominantly exhibit insufficient contraction, whereas temporal predictions show substantial deviations in both directions. Accordingly, we have updated the corresponding figures, empirical discussion, and conclusions to reflect the expanded results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.