acceptodds
Under review as a conference paper at ICLR 2027

TASCT: Temporal–Spectral Adaptation and Shape Calibration for Time Series Forecasting

Abstract

We introduce Temporal–Spectral Adaptation and Shape Calibration for Time Series Forecasting (TASCT), which combines temporal–spectral input adaptation with level-preserving forecast calibration. Its Temporal–Spectral Adapter (TSA) combines centered local-temporal and frequency-based corrections, preserving the historical mean. Its Level-Preserving Calibrator (LPC) combines centered contextual and consistency corrections, preserving the pre-calibration forecast mean. Across five backbones, seven datasets, and four forecasting horizons, TASCT improves three-seed mean MSE in 130 of 140 settings, with an average relative reduction of 4.17%. Experiments with LIFT and CCM further show that TASCT can provide additional gains on top of existing forecasting plugins. These results demonstrate individual and combined gains from input adaptation and output calibration.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, temporal-spectral adaptation, forecast calibration, mean-preserving correction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_brVCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TASCT enhances forecasting backbones with history input adaptation and forecast output calibration. The combination of multiscale temporal corrections and corrections from frequency-grouped input views, preserving historical mean in TSA is motivated and relevant. This and the backbone are jointly trained, followed by static and dynamic gate fitting . LPC then learns contextual residual corrections and corrections using neighboring predicted leads and previous-origin forecasts, while preserving each forecast's temporal mean. Across five backbones, seven datasets, four horizons, and three seeds, TASCT improves MSE in 130/140 settings, with a 4.17% average reduction. Ablations and composition/comparison with LIFT and CCM are also evaluated.

Critical strengths
  1. The paper provides reasonable evidence of improvement with matched Native/TSA configurations to support the main empirical claim. We see gains across several architecture families with 130/140 MSE wins and 123/140 MAE wins, and average reductions of 4.17% and 2.34%. It is interesting to note that TASCT can add value to LIFT and CCM enhanced forecasters as well. These experiments establish a useful practical contribution.

  2. The ablations reported in the paper help address alternative explanations, like the comparison of TSA with a parameter-matched residual MLP, individual branches, and the fixed 0.5/0.5 combination before gate fitting. Appendix F evaluates the contextual and consistency components, removal of centering, and simple temporal-averaging controls which are meaningful controls. TASCT's main contribution is the specific mean-preserving combination (of input/output adapters, spectral adaptation, and contextual revision) and the evidence supporting it.

Critical weaknesses
  1. Some implementation details in the paper are missing. I understand the TSA correction-energy penalty and the separate gate penalty. However, the it is not clear how the final LPC correction is fitted. How many pooled history/forecast features are used, whether the regression is shared across channels, and exactly how the consistency coefficients and final mixing weights are estimated. The chronological calibration blocks also need explanation of how overlapping forecast targets are handled.

  2. The benefit of the extra fitting stages is less clear than the benefit of TASCT overall. TSA already achieves a 3.49% average MSE reduction; adding LPC raises the total to 4.17%. This is useful, but Table 9 shows that the extra consistency component contributes little in several settings. The paper reports LPC's 0.38 ms average batch latency and 3.2–25.0K deployment coefficients, but it would also be beneficial to add the fitting time and a breakdown showing what the consistency component adds in accuracy and cost. Similarly, Table 8 compares fixed equal weights with fully fitted gates, but does not separate the benefit of learned static weights from the subsequent input-dependent gate. Adding these may improve the practical case for using every stage.

Questions for authors
  1. Could you provide one section or algorithm or table listing each stage's objective, fitted parameters, and regularization? For LPC, please specify the pooled features, ridge-penalty selection, coefficient-fitting procedure, including how the overlapping target intervals are handled.

  2. Since Table 8 already evaluates fixed equal weights, how much of the remaining improvement comes from learned static channel-wise weights, and how much comes from the dynamic gate?.

  3. Can you summarize the incremental improvement and fitting/inference cost of full LPC over contextual correction alone? To justify its added fitting and rolling-inference steps.

  4. In Table 12 δ-Adapter worsens all eight ETT backbone–dataset averages. Please explain this behavior, as there is clearly a difference in how the frozen-backbone is handled in δ-Adapter (frozen) v/s TASCT (joint training).

(nice to have) An inference-cost-matched weighted ensemble using an independently trained forecaster would provide another practical reference point, as LPC is a residual correction model and can be viewed as additive model combination.

Reviewer_vA2uOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TASCT, a modular time-series forecasting enhancement framework that combines a jointly trained, mean-preserving Temporal–Spectral Adapter (TSA) for historical input correction with a post-hoc, zero-mean Level-Preserving Calibrator (LPC) for output trajectory refinement across diverse forecasting backbones.

Critical strengths
  1. The authors evaluate TASCT across five backbones, seven benchmarks, four horizons, and three seeds, which convincingly establishes the reliability and broad applicability of the reported gains.
  2. Constraining both input adaptation and post-hoc output calibration to zero-mean temporal updates prevents level drift and conflicts with internal backbone normalizations like RevIN, which the ablations show is critical for avoiding accuracy degradation across diverse forecasting settings.
Critical weaknesses
  1. TASCT requires a four-stage training and post-hoc fitting procedure with numerous ad-hoc design choices, while improvements on stronger non-linear architectures are often marginal (<1% MSE).
  2. The local consistency correction in LPC requires caching consecutive stride-one rolling-origin forecasts, while Table 9 shows it contributes negligible accuracy gains compared to the contextual residual branch.
  3. The main paper lacks comparisons with other adapter methods.
Questions for authors
  1. How much performance is lost if the dynamic gate in TSA is trained jointly with the correction branches and backbone in a single stage rather than using two separate eight-epoch frozen gate-fitting stages?
  2. Could the local consistency correction be generalized to arbitrary inference strides or reformulated as a training-time regularization term to eliminate the need for stateful rolling-origin caching during deployment?
Reviewer_8ygoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TASCT, a modular framework for time-series forecasting that combines Temporal-Spectral Adaptation (TSA) and Level-Preserving Calibration (LPC). TSA adapts the input representation through multi-scale temporal and frequency-based corrections, while LPC adjusts the forecast trajectory while preserving its mean. Experiments across five forecasting backbones, seven datasets, and four prediction horizons show consistent improvements over baseline models. Ablation studies and comparisons with existing methods further investigate the contributions of the two components.

Critical strengths
  1. Broad experimental evaluation. The method is evaluated across five backbones, seven datasets, and four prediction horizons, providing evidence of its applicability across different forecasting settings.

  2. Modular design. The separation between input adaptation and forecast calibration makes the framework flexible and allows the contributions of the two components to be studied independently.

  3. Empirical improvements. The reported results show improvements across many experimental settings, and the ablation studies help demonstrate the contributions of the proposed components.

Critical weaknesses
  1. Limited clarity on novelty. The paper could more clearly distinguish TASCT from existing input-adaptation and post-hoc forecast-revision methods, including the closely related approaches evaluated in the paper.

  2. Insufficient robustness analysis. Although the results are generally promising, a more detailed analysis of variability across random seeds and experimental settings would help establish the reliability of the improvements.

  3. Unclear necessity of the full framework. Further comparisons with simpler correction mechanisms would help determine whether both components are necessary to achieve the reported gains and whether the additional complexity is justified.

Questions for authors
  1. What are the key technical differences between TASCT and existing input-adaptation and forecast-revision methods? Could the authors clarify which aspects constitute the main methodological contribution?

  2. How sensitive are the reported improvements to random seeds and individual datasets? Can the authors provide additional variability statistics or significance analyses?

  3. How does TASCT compare with simpler temporal or level-correction baselines? Are both TSA and LPC consistently necessary to obtain the reported improvements?

Reviewer_Jb8uOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes TASCT, which integrates Temporal–Spectral Adapter (TSA) for calibrating historical input and LEVEL-PRESERVING CALIBRATOR (LPC) for calibrating output when in the test stage. Across five backbones, seven datasets, and four prediction horizons, the paper reports lower three-seed mean MSE in 130 of 140 settings, with an average relative reduction of 4.17%. Ablations examine the individual modules, temporal and spectral branches, gating, and output centering.

Critical strengths
  1. The proposition of calibrating input and output signal separately is reasoning. The implement of TSA and LPC effectively capture the deviation from original signal. The framework leaves the backbone architecture unchanged, and LPC can be applied to either native or TSA-enhanced forecasts.
  2. Experimental results show that TASCT enhances different backbone models consistently. The ablation study attributes TASCT's effectiveness to a specific module detail.
  3. Figure 2 (overview of TASCT) is clear and makes it easy for the reader to follow.
Critical weaknesses
  1. The title named “TEMPORAL–SPECTRAL ADAPTATION” but I don’t really sure that the author knowing about the concept of “spectral”. Generally, we said the coefficients or strengths associated with a set of basis functions (e.g., the amplitude spectrum and the Fourier basis). So to my mind, TASCT primarily adapts the data distribution using frequency-grouped representations, rather than explicitly adapting its spectral characteristics. The authors should clarify what is being adapted in the spectral domain and why this constitutes spectral adaptation rather than distribution adaptation.
  2. In Figure 1, the authors aim to show that TSA improves the distribution of prediction errors, but I do not find the chosen metric convincing. The comparison based on a Gaussian distribution is not straightforward, and its relevance to forecasting accuracy needs clarification. If the authors intend to evaluate probabilistic predictions, a proper scoring rule such as the Continuous Ranked Probability Score (CRPS) would be more appropriate.
  3. The notation and definitions in the equations are not sufficiently clear. (1) In Equation 3, the different superscripts attached to Y are not clearly explained, making the purpose of the equation obscure. (2) The definitions in Equations 7 and 8 should appear before Equation 6 so that readers can follow the formulation more easily. (3) In Equation 11, what does interior leads h mean? Does h refer to a network component or a position in the forecast sequence? If it refers to the sequence position, Equation 11 appears to implement a moving-average operation, and the authors should clarify this interpretation
  4. Mean preservation is mathematically clear, but preserving a backbone’s forecast mean can also retain an incorrect level prediction. The manuscript does not establish when avoiding level correction is preferable to learning a suitably regularized level adjustment.
  5. The presentation of this manuscript needs further polishing. The concepts are introduced as separate additions rather than developed into a coherent framework. This makes it difficult for readers to follow the discussion, understand how the components are connected, and grasp the central idea of the proposed method.
Questions for authors

See weaknesses

1 comment

I am grateful to the authors for clarifying the notation and definitions. However, there are still some problems concerned with me:

  1. SKOLR encodes original time series to latent space, which can be viewed as a manifold, thus "spectra" in SKOLR are the eigenvalues of the manifold. In contrast, TASCT uses “spectra” to refer to amplitudes at fixed Fourier bins. These are conceptually distinct objects, so citing SKOLR as a precedent for the terminology risks conflating manifold-based spectra with the Fourier amplitude spectrum.
  2. "The branch therefore performs frequency-grouped input adaptation and does not directly learn per-frequency amplitude gains or phase shifts" is confused to me. Although the masks are fixed and the MLP operates on time-domain reconstructions rather than parameterizing per-frequency gains or phase shifts, its learned adaption can still change the signal’s Fourier amplitudes and phases.
  3. With CRPS metric, the adaptation results can be shown directly. To establish that the observed CRPS improvements generalize, the authors should evaluate them across a broader range of datasets and backbone models.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.