acceptodds
Under review as a conference paper at ICLR 2027

PRISM: A Hierarchical Multiscale Approach for Time Series Forecasting

Abstract

Time-series forecasting requires identifying which components of the observed context are informative about the future. Multiscale decompositions expose temporal and spectral structure of time series but the predictive relevance of these components can vary across samples and within the context window. We introduce PRISM, a forecasting framework that learns input-dependent selection over localized time-frequency supports of the context window. PRISM constructs a hierarchy of overlapping temporal regions and uses probabilistic selection to couple frequency weighting within each region with band-specific routing between sibling regions. The resulting allocation is learned end-to-end through the forecasting objective. Across eight standard long-horizon benchmarks, PRISM achieves the lowest mean squared error in 20 of 32 settings, and in 61 of 97 settings on GIFT-Eval. Controlled ablations establish the contribution of coupled hierarchical selection beyond flat aggregation and dense mixing. The learned allocations concentrate on a subset of supports, prioritize components whose deletion substantially impairs forecasting, and shift when predictive structure is relocated in synthetic experiments. These results support structured predictive-support selection as an effective inductive bias for time-series forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
interpretable machine learning, feature attribution, learned selection, time series forecasting

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 3 reviews, above 27% of the others.
Reviewer_chNbOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes PRISM, an MLP-based model that recursively branches input segments and adaptively selects forecast-relevant information from different temporal segments using Haar wavelet transforms. The method is evaluated on the LTSF and GIFT-Eval benchmarks against several prominent forecasting models and achieves strong results.

Critical strengths
  • The use of wavelets across both global and predominantly local segments is interesting.

  • While many recent methods combine spectral transformations with patches and Transformers, PRISM explores a relatively simple MLP-based alternative, which is an interestingly simple and practical direction.

  • The paper provides fair comparisons by using the best results reported in the original papers or through reproduction. This is particularly valuable in LTSF, where many works rely on their own reproduced baselines.

  • The authors evaluate their results on two established benchmarks, thus adding to the robustness of their approach.

Critical weaknesses

1. Novelty

Spectral and wavelet decomposition have been widely explored in deep-learning-based TSF. While PRISM achieves strong results, many of its components build on established techniques, with the main contribution appearing to lie in architectural design.

In particular, WPMixer [1], WaveletMixer [2], and WDformer [3] also employ localized/patch-based wavelet decomposition. The paper should more clearly distinguish PRISM from these closely related approaches and articulate its novel contribution. More importantly, compare their performance.

[1] WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting (AAAI 2025)

[2] WaveletMixer: A Multi-Resolution Wavelets Based MLP-Mixer for Multivariate Long-Term Time Series Forecasting (AAAI 2025)

[3] WDformer: A Wavelet-based Differential Transformer Model for Time Series Forecasting (CIKM 2025)

2. Experiments and Results

  • Why are SCINet and WPMixer omitted from Table 1? They are discussed and included in the significance analysis, but no explanation is provided for their absence from the main comparison.

  • The baselines appear to be relatively outdated, with methods introduced in 2024 or earlier. Given the rapid development of TSF, the evaluation would benefit from including more recent SOTA methods, particularly recent approaches using decomposition or spectral components.

  • The inference analysis is limited and relegated to the appendix. I suggest including additional baselines such as iTransformer, PatchTST, and relevant wavelet-based methods. Since wavelet transforms may introduce computational overhead, a more comprehensive efficiency analysis would strengthen the paper and could be moved to the main text.

3. Clarity and Presentation

  1. In Table 1, the Exchange 336 MAE result for PRISM appears to be missing bold formatting.
  2. Some appendix/main text tables appear incomplete or inconsistent; please check the “Best count” rows in Tables 2,12, 13, etc.
  3. Line 201 states that ties are credited to every tied method, potentially making the win counts exceed 32. Please clarify this.
  4. Section 4.2 feels somewhat unfocused and requires the reader to jump between multiple tables. A more direct discussion of the main findings would make the analysis substantially clearer.
Questions for authors

see above

4 comments

Response to Reviewer chNb

We thank the Reviewer chNb for providing the feedback on our paper. Following your suggestions, we added and compared more recent decomposition-based baselines, extended the efficiency analysis to iTransformer, PatchTST and the wavelet-based methods, and revised the presentation of Section 4.2. We hope what we have done here conveys our effort to address your concerns.

(W1) Novelty relative to WPMixer, WaveletMixer and WDformer

Many components build on established techniques and the contribution appears to be architectural. WPMixer, WaveletMixer and WDformer also use localized or patch-based wavelet decomposition. Please distinguish PRISM from them and, more importantly, compare their performance.

To follow the Reviewer’s suggestion, we compared PRISM’s performance with that of the suggested related works. We observed that, across the standard evaluation settings (combined across the forecasting horizons for brevity), PRISM shows the best MSE performance on 7 out of 8 settings and the best MAE performance on 8 out of 8 settings. These additional comparisons further strengthen the empirical results of our work.

DatasetPRISMWPMixerWDformerWaveletMixer
ETTh10.388/0.4080.411/0.4230.499/0.4810.405/0.423
ETTh20.317/0.3650.343/0.3870.425/0.429--
ETTm10.345/0.3660.352/0.3780.375/0.4020.347/0.377
ETTm20.247/0.2980.258/0.3140.299/0.3420.250/0.313
Weather0.218/0.2470.224/0.2570.243/0.2780.219/0.260
Electricity0.155/0.2440.164/0.2550.163/0.2580.155/0.254
Traffic0.393/0.2520.406/0.2620.374/0.264--
Exchange0.320/0.3740.380/0.4100.420/0.4470.414/0.432

Higher performance of our model comes with architectural differences. WPMixer, WaveletMixer and WDformer all have rigid rules for combining components of decomposed time-series signals, where the mixing is learned once and applied to every input. PRISM recomputes the allocation per sample and lets the same band be treated differently in different parts of the context window. The ablations in Table 6 and Table A.2 show that this design contributes to the increased performance. Replacing tree-structured selection with flat multiscale aggregation worsens the MSE by 5.79%, and replacing explicit selection with implicit mixing worsens it by 4.18%. Both ablations use the same decomposition of the time-series signal, so the contribution of the selection mechanism is separate from that of the decomposition, and separate from what’s done in prior works.

(W2) 2.1 SCINet and WPMixer in Table 1

Why are SCINet and WPMixer omitted from Table 1 although they appear in the significance analysis?

To address the Reviewer’s comment, we provide the requested values below. Across these settings, PRISM shows the best MSE results in 8 out of 8 settings (combined over forecasting horizons here for brevity) and the best MAE in 8 out of 8 settings as well. We did not include these results in Table 1 because, as the counts of relevant baselines have been growing, we aimed to keep the font size of our main results legible.

DatasetPRISMWPMixerSCINet
ETTh10.388/0.4080.411/0.4230.485/0.475
ETTh20.317/0.3650.343/0.3870.406/0.434
ETTm10.345/0.3660.352/0.3780.374/0.395
ETTm20.247/0.2980.258/0.3140.289/0.343
Weather0.218/0.2470.224/0.2570.264/0.295
Electricity0.155/0.2440.164/0.2550.195/0.283
Traffic0.393/0.2520.406/0.2620.589/0.385
Exchange0.320/0.3740.380/0.4100.432/0.454

(W2) 2.2 More recent baselines

The baselines are from 2024 or earlier. Recent methods, in particular decomposition- or spectrum-based ones, should be included.

We also provide the comparisons with more recent baselines. Across these, PRISM shows the best MSE results in 27 of 32 settings and the best MAE results in 32 out of 32 settings. These additional comparisons to newer works further strengthen the empirical contributions of our work.

DatasetHPRISM MSEPRISM MAEFilterNet MSEFilterNet MAEFreqCycle MSEFreqCycle MAEAWEMixer MSEAWEMixer MAEWDformer MSEWDformer MAE
ETTh1960.3550.3740.4240.4260.3710.3970.3710.3970.4150.424
ETTh11920.3900.4050.4530.4490.4020.4180.4060.4160.4630.459
ETTh13360.3860.4070.4350.4350.4260.4360.4330.4330.4690.464
ETTh17200.4210.4450.4500.4610.4540.4750.4660.4730.6500.577
ETTh2960.2610.3220.3440.3880.2850.3470.2870.3470.3380.378
ETTh21920.3110.3590.3590.3970.3460.3880.3470.3870.4250.425
ETTh23360.3080.3610.3860.4230.3740.4140.3800.4140.4690.450
ETTh27200.3870.4200.4110.4410.4250.4500.4180.4490.4680.465
ETTm1960.2820.3260.3140.3560.2880.3420.2910.3420.3160.367
ETTm11920.3230.3520.3400.3780.3280.3660.3310.3670.3550.390
ETTm13360.3540.3770.3840.4030.3600.3850.3680.3890.3850.409
ETTm17200.4230.4090.4350.4320.4200.4180.4230.4200.4450.441
ETTm2960.1580.2340.1790.2620.1640.2530.1740.2570.1940.276
ETTm21920.2140.2670.2430.3060.2220.2920.2350.3020.2720.324
ETTm23360.2640.3190.2890.3360.2750.3300.2790.3320.3240.357
ETTm27200.3520.3720.3730.3940.3620.3850.3650.3870.4050.409
Traffic960.3620.2370.3810.2680.3940.2680.3940.2800.3530.252
Traffic1920.3760.2440.4010.2770.4110.2760.4090.2810.3670.262
Traffic3360.4010.2570.4130.2820.4230.2810.4330.2940.3760.265
Traffic7200.4330.2710.4400.2940.4510.2980.4630.3060.4020.276
Electricity960.1220.2190.1330.2270.1290.2240.1330.2280.1310.226
Electricity1920.1450.2350.1510.2440.1460.2400.1500.2430.1540.248
Electricity3360.1590.2430.1670.2610.1630.2580.1680.2610.1680.263
Electricity7200.1940.2810.2050.2940.2010.2920.2060.2930.1980.293
Exchange960.0810.1980.0980.2230.0930.2160.0980.2210.1130.240
Exchange1920.1730.2970.2220.3400.1940.3140.1900.3110.2100.334
Exchange3360.3140.3960.3650.4400.3590.4330.3770.4480.4150.476
Exchange7200.7120.6060.9280.7240.9420.7290.9750.7460.9420.738
Weather960.1400.1770.1480.2000.1490.2020.1480.1990.1660.216
Weather1920.1870.2270.1890.2390.1920.2450.1890.2390.2100.257
Weather3360.2340.2660.2420.2800.2460.2860.2400.2800.2630.295
Weather7200.3100.3200.3230.3350.3300.3410.3150.3320.3340.342
Best count273200100040

(W2) 2.3 Efficiency analysis

The inference analysis is limited and in the appendix. Please add iTransformer, PatchTST and wavelet-based methods, quantify the overhead of the wavelet transform, and consider moving the analysis to the main text.

We thank the Reviewer for this suggestion and have extended the efficiency analysis to the suggested baselines. In addition to D-PAD, DLinear, TimeMixer and XPatch, the revised table includes iTransformer, PatchTST and the two wavelet-based methods WPMixer and WDformer. All models are built from their released code with their ETTh1 hyperparameters and measured under one setting (context 336, horizon 96, batch 8, fp32, one GPU). We report parameter count, FLOPs per sample, which depends only on the model and not on the hardware, and peak memory at inference and during training. The table lists every model relative to PRISM.

ModelParams (M)MFLOPs / sampleInference peak mem (MB)Train peak mem (MB)
PRISM0.355.113.726.6
DLinear0.10 (0.3×)0.9 (0.2×)9.0 (0.7×)17.9 (0.7×)
iTransformer0.25 (0.7×)5.7 (1.1×)10.7 (0.8×)21.7 (0.8×)
XPatch0.33 (0.9×)9.8 (1.9×)14.5 (1.1×)33.9 (1.3×)
WDformer2.15 (6.1×)41.9 (8.2×)18.8 (1.4×)53.5 (2.0×)
D-PAD3.22 (9.2×)46.8 (9.2×)25.0 (1.8×)78.2 (2.9×)
TimeMixer0.73 (2.1×)182.6 (36×)27.8 (2.0×)83.0 (3.1×)
WPMixer2.26 (6.4×)538.3 (106×)29.0 (2.1×)145.7 (5.5×)
PatchTST5.23 (14.9×)1909.2 (374×)97.8 (7.1×)238.5 (9.0×)

Compute. Comparison shows that PRISM is among the lightest models of the table. It needs 5.1 MFLOPs per sample, the same order as DLinear and iTransformer, and 8 to 374 times fewer than the other decomposition-based models, WDformer (41.9 M), D-PAD (46.8 M), TimeMixer (183 M), WPMixer (538 M) and PatchTST (1909 M). With 0.35 M parameters it is smaller than every model except DLinear and iTransformer, and its training memory of 27 MB is a third of D-PAD and TimeMixer and a fifth of WPMixer.

Wavelet overhead. The Haar decomposition is a fixed linear operator with 0.18 MFLOPs per sample in this setting, 3.5% of PRISM's total of 5.1 MFLOPs, and it runs once per node of the temporal tree. The two wavelet baselines cost 8 and 100 times more per sample than PRISM in total, so the overhead of these methods comes from what is built on top of the transform, not from the transform itself. We move the table to Section 4.2.

(W3) 3.1 Missing bold in Table 1

The Exchange-336 MAE entry of PRISM lacks bold.

Thank you for highlighting this. We have fixed the missing bold and the Best-count rows.

(W3) 3.2 Best-count rows and tie counting

Some Best-count rows (Tables 2, 12, 13) look inconsistent. Line 201 says ties are credited to every tied method, so counts can exceed 32.

We would like to clarify that, as a setting, we consider a (dataset, horizon) pair, and a method is credited whenever its value equals the row minimum at the reported precision. When two methods tie both are credited, so column sums can exceed 32. Throughout the text, we used the same approach as specified in Line 201. To address the reviewer’s comment, we added clarifications regarding best counts throughout the tables.

(W3) 3.3 Section 4.2

Section 4.2 is unfocused and jumps between tables.

We thank the Reviewer for highlighting this, as Section 4.2 is the main results section and has to be clear and easy to follow. To address the Reviewer’s comment, we rewrote Section 4.2 so that each paragraph makes one claim and points to one table. The revised section now reads in this order:

  • Setup: datasets, context length, horizons and baselines.

  • Main result: Table 1 and the best-count rule.

  • Significance: paired t-tests with FDR correction (Tables A.4 and A.5).

  • Robustness: the same comparison at context 96 (Table A.14).

Hyperparameter protocols are referenced once at the end (Tables A.15 to A.17). The full text is in the revised PDF.

I thank the authors for the detailed response and the additional experiments.

The new comparisons with wavelet-based and recent baselines and the clarification of the best-count rule address several of my concerns. However, some concerns remain

  1. Novelty. I still have concerns about the novelty of PRISM compared with WPMixer, WaveletMixer and WDformer.

  2. Efficiency numbers. The new numbers do not match the paper. For ETTh1 (context 336, horizon 96), PRISM is reported at 1.68M FLOPs (Table A.1), 2.52M (Table A.18) and 5.1M (rebuttal). DLinear goes from 679K parameters and 3.63M FLOPs to 0.10M and 0.9M, which reverses the paper’s claim that PRISM needs fewer FLOPs than DLinear. Peak memory goes from 235 MB to 13.7 MB. Latency is also missing from the new table, although Table A.1 shows PRISM to be slower than DLinear and TimeMixer.

  3. Hyperparameter tuning and fairness of the new comparisons. PRISM’s hyperparameters appear to have been tuned on the test set. The search in Tables A.11–A.13 reports test results identical to Table 1, and the selected values are the ones with the best test scores. The released scripts also set hyperparameters per dataset and horizon, and many of these settings are missing from Table A.15.

Therefore, based on the remaining concerns about novelty, the inconsistent efficiency results, and the evaluation protocol, I keep my score.

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

PRISM learns which localized time-frequency components of a historical window are useful for forecasting. It recursively partitions the input into overlapping temporal segments and applies wavelet decomposition to construct frequency-band signals. Routers score statistical descriptors of these signals and combine within-segment frequency weighting with comparisons between sibling segments. The weighted components are passed to forecasting heads and their predictions are combined. The paper reports strong results on standard long-horizon benchmarks and GIFT-Eval, alongside ablations, allocation-concentration statistics, and component-deletion analyses.

Critical strengths
  • I find the idea of decomposing historical information into meaningful time-frequency components and learning their contributions to future prediction appealing. The same frequency band may have different predictive value at different temporal locations, which motivates the joint treatment of frequency and temporal selection.

  • The explicit routing structure offers a useful way to inspect the historical components used by the model. Coupling frequency weighting with sibling comparisons and training through the forecasting objective makes the selection mechanism amenable to targeted ablations.

  • The evaluation considers several datasets and includes flat multiscale aggregation, implicit mixing, routing ablations, and deletion of high- versus low-weight components. These analyses address relevant questions about the proposed mechanism, although some definitions and reported results require verification.

Critical weaknesses
  • Reproduction means and significance statistics are inconsistent. In Table A.3, the four ETTh1 MSE values average to 0.3880 for PRISM and 0.4035 for reproduced D-PAD, giving a difference of -0.0155. Table A.4 instead reports -0.0114. If the tables use the same eight runs and four horizons, reversing the order of averaging over seeds and horizons cannot change the mean difference. The discrepancy of 0.0041 also exceeds the rounding tolerance of approximately 0.00105, accounting for both tables. This is not confined to one entry. I checked the D-PAD MSE comparisons on the other five datasets and found discrepancies beyond rounding tolerance in all five. For example, ETTm1 gives -0.01075 from Table A.3, whereas Table A.4 reports -0.0082. The authors should clarify whether the tables use the same runs, configurations, and aggregation rules, and provide the underlying results and statistical scripts. Until reconciled, the significance claims cannot be reliably checked.

  • Routing weights and allocation statistics are not clearly reconciled. Section 3.3 defines each component weight as a softmax probability between sibling branches. For every frequency band, the left and right weights therefore sum to one. Table 3 analyzes six frequency bands across two temporal regions, giving twelve candidate supports. If its statistics are computed directly from these branch weights, their total is six. After normalization, each component has mass at most 1/6, so the top-3 mass cannot exceed 0.5. However, Table 3 reports 0.88 for ETTh2 and 0.98 for Weather. The analysis may use a different quantity, such as additional frequency weights, path weights, or forecasting-head gates, but the manuscript does not establish this connection. Please define the exact weights and normalization used for entropy, effective support, and top-3 mass, and explain how they relate to the forward computation. This matters because concentrated predictive-support selection is a central claim of the paper.

  • The claimed synthetic experiments lack identifiable supporting details. The abstract and introduction state that controlled synthetic experiments show the allocation shifting when predictive information is relocated. I could not locate the corresponding data-generation process, controlled variables, experimental setup, or result figures and tables in the supplied PDF. Table 4 supports the claim that the model depends on high-weight components, but does not directly test whether allocation changes when predictive information moves. Please identify the relevant material or provide the existing experimental details and results, so this claim can be assessed.

Questions for authors
  1. Do Tables A.3 and A.4 use the same runs and configurations? Please explain the ETTh1 discrepancy (-0.0155 versus -0.0114), check the other datasets, and provide results for each seed and horizon together with the aggregation script. Do the significance conclusions change after correction?

  2. What exact weights underlie Table 3? Please give the full computation from the routing probabilities in Section 3.3 to the analyzed distribution, and explain how the reported top-3 mass can exceed the bound implied by the stated branch-weight definition.

  3. Where are the controlled synthetic experiments reported? Please provide the existing data-generation procedure, explain how predictive information was relocated, and show the resulting allocation changes and forecasting results.

  4. Why does the temporal router operate on normalized frequency probabilities rather than raw component scores? Have alternative normalization schemes been evaluated?

5 comments

Response to Reviewer LyQB

We thank the Reviewer for your time spent on our paper and for finding the idea of learning the contribution of time-frequency components appealing. We hope the responses below could address your concerns.

(W1, Q1) Tables A.3 and A.4

The mean differences implied by Table A.3 (for example ETTh1 MSE, PRISM minus D-PAD, -0.0155) do not match Table A.4 (-0.0114), beyond rounding, on all six datasets. Please clarify whether the tables use the same runs, configurations and aggregation rules, and provide the underlying results and statistical scripts.

As the Reviewer has correctly suggested, the differences in the tables reflect different runs of the model. As we expanded the analyses, we needed additional runs of the model to obtain more detailed statistics. We chose to not change preexisting tables because their internal consistency has already been validated. As a result, we have small discrepancies, as identified by the Reviewer, that exceed the rounding error yet remain negligible (do not change the conclusions of the paper).

(W2, Q2) Table 3 and the top-3 mass bound

Section 3.3 defines each component weight as a softmax between sibling branches, so with six bands and two regions each component has mass at most 1/6 after normalization and the top-3 mass cannot exceed 0.5, yet Table 3 reports 0.88 (ETTh2) and 0.98 (Weather). Please define the exact weights and normalization used for entropy, effective support and top-3 mass.

We would like to clarify that Table 3 analyzes the frequency-selection distribution qθ(Km,c=k∣Xm,c)q_\theta(K_{m,c}=k \mid X_{m,c}), the first softmax of the router, which assigns mass to the K frequency bands within one temporal node and sums to one within that node. The sibling-branch softmax is a second, separate stage that decides between the left and right child of a node and is not part of the analyzed quantity. For every node, branch, channel and test window we take this K-dimensional vector and compute the mass of its three largest entries, and Table 3 reports the averages of these three numbers. Since the only constraint is that each vector sums to one on its own, a single band can carry most of the mass, which is what the Weather row (top-3 mass 0.98, effective support 1.24) shows. The uniform reference row (entropy 1, effective support K, top-3 mass 3/K) is computed for the same vector. We will modify the caption to "This analysis uses K = 12 frequency bands within each temporal node, giving 12 candidate supports."

(W3, Q3) The controlled synthetic experiments

The abstract and introduction claim controlled synthetic experiments showing that the allocation shifts when predictive information is relocated, but no data-generation process, controlled variables, setup or results appear in the PDF. Table 4 shows dependence on high-weight components but does not test whether allocation moves.

We thank the reviewer for this observation. The experiment was run and its results were obtained, but, throughout organizing the paper, we excluded them in favor of other results. We describe the experiment and its results below.

Synthetic dataset. We generated context windows as sums of five predefined time-frequency components. We then generated forecast windows as linear projections of energies of chosen subsets of the context-window time-frequency segments. This way, only limited subsets of known ground-truth components carried information relevant for the forecast, while the remaining components acted as structured distractors of comparable amplitudes. Recovering the ground-truth therefore required the router to place its predictive mass on the projected subset rather than on most energetic or most recent components.

Controlled variable. The only quantity that changes between the two conditions is which component the projection reads from, a 2 Hz component or a 24 Hz component. The generator, the noise level, the model, the training recipe and the evaluation are identical.

Setup. PRISM is trained from forecasting MSE only, with no access to the ground-truth component, and the router weights are averaged over the test windows.

Predictive support allocation result. We’ve put predictive information into lower-frequency (2 Hz) or higher-frequency (24 Hz) signals and compared the router’s learned allocation across the two.

Ground truthComp. 1 (5.8 Hz)Comp. 2 (13.1 Hz)Comp. 3 (21.4 Hz)Comp. 4 (33.4 Hz)Comp. 5 (56.4 Hz)
Low frequency (2 Hz)0.8760.0920.0240.0060.002
High frequency (24 Hz)0.0170.0090.2480.5200.206

We observed that the learned allocation followed the ground truth. 87.6% of the predictive mass goes to the band nearest 2 Hz in the low setting and 52.0% to the band nearest 24 Hz in the high setting, with the adjacent bands (3 and 5) taking most of the remainder. This result supported our statement that PRISM tracks the structure of the signal instead of forming fixed preferences for particular time-frequency components.

(Q4) Normalization before the temporal comparison

Why does the temporal router operate on normalized frequency probabilities rather than raw component scores, and were alternative normalizations evaluated?

Thank you for this question. In PRISM, we use the normalization before temporal comparison because it gives scale invariance across nodes. The raw router scores have no common scale between temporal segments, so a segment with larger signal energy would produce systematically larger outputs, and a branch softmax over raw scores would route bands toward the more energetic sibling for reasons unrelated to predictive value. Normalizing within the node first turns the branch comparison into a comparison of relative within-segment importance, which is the quantity the mechanism is meant to compare.

To follow the Reviewer’s suggestion, we tested the proposed alternative, applying the branch softmax to the raw scores with everything else unchanged. We observed that this change has worsened the results on all four datasets, by 2.6% in MSE and 2.8% in MAE on average and by up to 4.4% / 6.0% on ETTm2. This observation suggests that the normalization that we’ve implemented in the model is useful for increasing the model’s performance.

DatasetRaw scores (seed 13)Two-stage (Table 1)ΔMSEΔMAE
ETTh10.360 / 0.3830.355 / 0.374+0.005 (+1.4%)+0.009 (+2.4%)
ETTh20.265 / 0.3240.261 / 0.322+0.004 (+1.6%)+0.002 (+0.7%)
ETTm10.291 / 0.3320.282 / 0.326+0.009 (+3.2%)+0.006 (+1.9%)
ETTm20.165 / 0.2480.158 / 0.234+0.007 (+4.4%)+0.014 (+6.0%)
Mean+2.6%+2.8%
Reviewer_LyQBOpenReview ↗

Follow up Comments

Thank you for the clarifications and additional experiments. Two main concerns remain:

  1. Statistical verification. Different runs may explain the discrepancy between Tables A.3 and A.4, but the requested per-seed results and statistical scripts are still missing. We therefore cannot verify the significance conclusions.
  2. Allocation evidence. Table 3’s revised interpretation requires confirmation of the actual twelve-band configuration. In the synthetic experiment, the claimed 52% allocation to the band nearest 24 Hz corresponds to 33.4 Hz, although 21.4 Hz is closer. This discrepancy needs clarification, and forecasting errors remain unreported.

These concerns remain unresolved. We maintain our original score of 2.

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

This paper introduces PRISM, a time-series forecasting framework that learns predictions as an input-dependent allocation with a predictive-support selection formulation over localized time-frequency components. The time-series is partitioned into overlapping temporal regions which are further decomposed into frequency bands of Haar Wavelets. A weighted probabilistic selection and routing mechanism is then used for end-to-end training under a forecasting objective.

Critical strengths

The hierarchical framework is easy to follow, well formulated, and principled. Offers some insights into interpretability and robustness. Probabilistic selection over a binary temporal partition tree is an intuitive multi-step forecasting formulation; decouples time-series into time-frequency supports with multiple resolutions. Rigorous empirical evaluation across 8 seeds of standard forecasting benchmarks and GIFT-Eval, with strong performance and ablations (across training times, inference, memory, etc.). Importance scores also provide useful insights. Some empirical evidence of Haar performing better than other alternative decompositions.

Critical weaknesses

There are still several works that use time-frequency or multiscale approaches that should be compared against or discussed more thoroughly. Predictive-support selection is a nice addition, but most of the other components still seem to be repurposed from existing ideas. The paper includes TimeMixer, D-PAD as baselines, but does not clearly articulate the conceptual differences in novelty, nor does it compare against other relevant methods already cited in the paper, such as Pathformer, WaveletMixer, WPMixer, and TFDNet. Here are a few additional recent and uncited works: FreqCycle [1], TimeMixer++ [2], AWEMixer [3], FilterNet [4]. The authors mention “controlled synthetic experiments” for predictive support allocation shifts, but no corresponding experiments or discussion appear to be reported in the paper. Minor typos: line 104 and 182 (‘for’), line 321 (‘entirety of’)

References:

[1] Shiyu Wang et al. TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis, ICLR 2025. [2] Boya Zhang et al. FreqCycle: A Multi-Scale Time-Frequency Analysis Method for Time Series Forecasting, AAAI 2026. [3] Qianyang Li et al. AWEMixer: Adaptive Wavelet-Enhanced Mixer Network for Long-Term Time Series Forecasting, IEEE IoT Journal 2026. [4] Kun Yi et al. FilterNet: Harnessing Frequency Filters for Time Series Forecasting, NeurIPS 2024.

Questions for authors
  1. How does PRISM explicitly compare with and conceptually differ from other time-frequency and multiscale approaches?
  2. Are the synthetic experiments mentioned in the paper actually reported?
3 comments

Response to Reviewer ZDNT

We thank the reviewer for your time. It means a lot to us that you think our predictive-support selection is a worthwhile addition. We have followed your suggestions with care, and we hope you can feel that in the response below.

(W1, Q1) Conceptual difference from time-frequency and multiscale methods

Apart from predictive-support selection, PRISM's components come from existing ideas. How does it differ conceptually from TimeMixer, D-PAD, Pathformer, WaveletMixer, WPMixer, TFDNet, FreqCycle, TimeMixer++, AWEMixer and FilterNet?

We appreciate the Reviewer's remark on the predictive-support selection being a nice design choice. In line with this comment, we would like to highlight that the selection mechanism alone provides a strong contribution to forecasting results. Ablations in Table 6 and Table A.2 illustrate this. Replacing tree-structured selection with flat multiscale aggregation leads to a 5.79% worsening of the MSE and replacing explicit selection with implicit mixing leads to a 4.18% worsening of the MSE (both under the same decomposition).

Second, the selection mechanism has enabled us to use a more granular representation space. Joint time-frequency tiling of the series increases the dimensionality of the input representation, which would normally lead to overfitting. In PRISM, however, the selection mechanism limits the effective predictive support to a subset of these tiles, so the model can admit high-resolution components wherever they are needed without an effective increase in input dimensionality. The same design makes the model robust to hyperparameter choices, since oversampling the search space yields a one-size-fits-all configuration, as the GIFT-eval results across about 100 datasets illustrate.

We further describe specific differences with prior work.

One group of prior work recombines decomposed time-series components in input-agnostic ways. TimeMixer sums its per-scale forecasts with equal weights, WPMixer and WDformer rejoin their wavelet bands with the inverse wavelet transform, FilterNet multiplies the spectrum by one learned filter, TFDNet applies learned time-frequency kernels, and FreqCycle combines its windowed spectra with a trained weight vector. PRISM goes one step further and computes, for every input, how much each time-frequency component of the input should support the forecast. When we remove this step and let the pieces mix densely as these methods do (Table 6, implicit mixing), the MSE rises by 3.59%.

A second group of prior work does not do temporal decomposition. D-PAD gates its modes with sigmoids, AWEMixer computes a softmax over bands from the energy of the window, TimeMixer++ weights its periods by FFT amplitude, and Pathformer picks a few patch sizes. PRISM goes one step further again and makes the decision per time segment as well, by tiling the window jointly in time and frequency and giving each (segment, band) pair its own weight, so a band can carry the forecast in the recent part of the window and be dropped in the early part. When we remove the temporal tree and keep only a whole-window choice over bands, as these methods do (Table 6, flat multiscale), the MSE worsens by 6.26%, and removing temporal selection alone costs 3.92%.

(W1) Experimental comparison with Pathformer, WaveletMixer, WPMixer, TFDNet, FreqCycle, TimeMixer++, AWEMixer and FilterNet

To address this comment, we added the requested comparisons. We found that, among newly added methods, PRISM shows the best MSE on 6 out of 8 settings (the results are averaged across the forecast horizons for brevity) and the best MAE on 8 out of 8 settings. Across all already-reported and newly added results PRISM remains the highest-scoring method, best on 20 MSE and 19 MAE settings of 32, still followed by D-PAD and DLinear. We thank the reviewer for proposing these comparisons, which further strengthen our empirical results.

DatasetPRISMWPMixerFilterNetFreqCycleAWEMixerWDformerTimeMixer++WaveletMixerTFDNet-IKTFDNet-MKPathformer
ETTh10.388/0.4080.411/0.4230.441/0.4430.413/0.4310.419/0.4300.499/0.4810.418/0.4320.405/0.4230.407/0.4210.401/0.4160.412/0.423
ETTh20.317/0.3650.343/0.3870.375/0.4120.357/0.4000.358/0.3990.425/0.4290.339/0.380not reported0.335/0.3770.335/0.3770.337/0.376
ETTm10.345/0.3660.352/0.3780.368/0.3920.349/0.3780.353/0.3800.375/0.4020.368/0.3780.347/0.3770.345/0.3680.346/0.3690.377/0.382
ETTm20.247/0.2980.258/0.3140.271/0.3250.256/0.3150.263/0.3200.299/0.3420.269/0.3200.250/0.3130.245/0.3040.246/0.3050.250/0.309
Weather0.218/0.2470.224/0.2570.226/0.2640.229/0.2690.223/0.2630.243/0.2780.226/0.2620.219/0.2600.220/0.2530.226/0.2590.236/0.263
Electricity0.155/0.2440.164/0.2550.164/0.2570.160/0.2540.164/0.2560.163/0.2580.165/0.2530.155/0.2540.158/0.2490.160/0.2510.175/0.262
Traffic0.393/0.2520.406/0.2620.409/0.2800.420/0.2810.425/0.2900.374/0.2640.416/0.264not reported0.406/0.2680.386/0.2570.501/0.299
Exchange0.320/0.3740.380/0.4100.403/0.4320.397/0.4230.410/0.4310.420/0.4470.357/0.4090.414/0.4320.392/0.4190.391/0.4170.398/0.412

Caption: Dataset averages over H = 96/192/336/720 (MSE/MAE). PRISM, WPMixer, FilterNet, FreqCycle, AWEMixer and WDformer at context 336 (baselines re-run with their released hyperparameters). TimeMixer++ from its paper at input length 96 (no public code). WaveletMixer from its paper at input length 512 (no public code; ETTh2 and Traffic not reported). TFDNet-IK and TFDNet-MK from its paper, which reports for each setting the best of input lengths {96, 192, 336, 512, 720}. Pathformer we take, for each setting, the best of the published input-96 number with re-run of the released code at input lengths 336.

(W2, Q2) The synthetic experiments
The paper mentions "controlled synthetic experiments" on predictive-support allocation shifts, but no corresponding experiment is reported.

We thank the reviewer for this observation. The experiment was run and its results were obtained, but, throughout organizing the paper, we excluded them in favor of other results. We describe the experiment and its results below.

Synthetic dataset. We generated context windows as sums of several predefined time-frequency components. We then generated forecast windows as linear projections of energies of chosen subsets of the context-window time-frequency segments. This way, only limited subsets of known ground-truth components carried information relevant for the forecast, while the remaining components acted as structured distractors of comparable amplitudes. Recovering the ground-truth therefore required the router to place its predictive mass on the projected subset rather than on most energetic or most recent components.

Predictive support allocation result. We’ve put predictive information into lower-frequency (2 Hz) or higher-frequency (24 Hz) signals and compared the router’s learned allocation across the two.

Ground truthComp. 1 (5.8 Hz)Comp. 2 (13.1 Hz)Comp. 3 (21.4 Hz)Comp. 4 (33.4 Hz)Comp. 5 (56.4 Hz)
Low frequency (2 Hz)0.8760.0920.0240.0060.002
High frequency (24 Hz)0.0170.0090.2480.5200.206

We observed that the learned allocation followed the ground truth. 87.6% of the predictive mass goes to the band nearest 2 Hz in the low setting and 52.0% to the band nearest 24 Hz in the high setting, with the adjacent bands (3 and 5) taking most of the remainder. This result supported our statement that PRISM tracks the structure of the signal instead of forming fixed preferences for particular time-frequency components.

(W3). Typos Line 104 and 182 ("for"), line 321 ("entirety of").

Thank you for pointing these out. We have fixed them.

Reviewer_ZDNTOpenReview ↗

acknowledging the rebuttal

thx for the clarifications. My rating is increased to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.