acceptodds
Under review as a conference paper at ICLR 2027

ARBR: Adaptive Radial Basis Representation for Time Series Forecasting

Abstract

The temporal location and extent of predictive patterns are important for time series forecasting. Recent basis-based forecasting methods often combine shared temporal bases with input-dependent coefficients. However, adapting coefficients alone cannot directly relocate or rescale individual basis functions, so changes in pattern location and extent must be represented through coordinated changes across multiple shared bases. To model such geometric variation explicitly, we propose Adaptive Radial Basis Representation (ARBR), a structured temporal representation built from input-adaptive Gaussian radial bases. For each channel-wise input window, ARBR predicts bounded adjustments to radial centers and scales around a shared temporal scaffold, preserving the temporal ordering of components. Radial responses weight input-conditioned vector-valued coefficients to form latent features aligned with temporal patches. On seven multivariate forecasting benchmarks, ARBR achieves the lowest MSE on five datasets and the lowest MAE on four, averaged over four forecasting horizons. Controlled comparisons show that ARBR yields lower forecasting errors than variants with fixed or globally learned radial geometry or non-radial dynamic mixing. The code is available at https://anonymous.4open.science/r/ARBR/.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, radial basis functions, adaptive temporal geometry, input-conditioned coefficients

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_dmQYOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes a radial-basis-based method for time series forecasting. For each input window, the model predicts bounded adjustments to preset Gaussian basis centers and scales, together with input-adaptive coefficient vectors. The rendered representation is combined with local patch features and passed to a linear forecasting head. Experiments on seven benchmarks show its effectiveness.

Critical strengths
  1. The method is intuitive and clearly formulated. Exploiting Gaussian bases and adapting their temporal location and extent offers a structured way to represent input-dependent variation while preserving a noise-filtering mechanism.

  2. The design of the bounded adjustments to preset Gaussian basis is well-motivated.

  3. ARBR achieves competitive results across several datasets with relatively low FLOPs and memory usage.

Critical weaknesses

My main concerns are about how well the paper's prior design choices and their actual effects align with its stated motivation.

  1. The preset Gaussian basis centers and scales lack both experimental ablation and justification. By default, the scales of the Gaussian bases increase monotonically along the normalized time axis [0, 1], which implies that in the input sequence, positions far from the forecast point are covered by sharp (narrow) bases, while positions close to the forecast point are covered by flat (wide) bases. This seems inconsistent with the intuition that forecasting should extract long-range dependencies (or low-frequency information) from distant positions and short-range dependencies (or high-frequency information) from recent positions. Further justification for this design choice would be helpful.

  2. Fig. 1 is only illustrative. The motivation it depicts has not been verified by any experiment that reconstructs the input sequence, nor has it been compared with other basis-based methods. As a result, it remains unverified whether the proposed method actually fits input patterns while suppressing noise in the way Fig. 1 suggests. It would be better to add an experiment that reconstructs/fits the input sequence from Z or G in Fig. 2, and compares the fitting quality and noise-filtering effects with other basis-based methods.

  3. To better verify the role of the proposed ARBR, it would be desirable to add an ablation study that retains only the prediction head in Fig. 2(a).

After reading the rebuttal, I think the authors have addressed my concerns, so I have raised my rating to 3.

Questions for authors

See "Critical Weaknesses"

3 comments
AuthorsOpenReview ↗

Author Response to Reviewer dmQY

Thank you for your careful review. We clarify the scaffold design and provide additional scaffold, reconstruction, and prediction-head controls.

W1/Q1. Uniform centers and increasing scales

Uniform centers provide a competitive simple prior, while increasing scales improve recent-noise robustness with a temporal trade-off.

Gaussian scales control temporal coverage, rather than explicitly decomposing high/low frequencies. Increasing scales target recent local perturbations.

We vary center layouts and scale ordering, retain input-adaptive geometry and the remaining architecture, and train all variants from scratch.

Table R1. Reference-scaffold ablation

VariantCenter priorScale priorAvg. MSE ↓ΔMSE
A. ARBRUniformIncreasing0.3001 ± 0.0007—
B. ConstantUniformConstant0.3020 ± 0.0008+0.63%
C. ReversedUniformDecreasing0.3034 ± 0.0011+1.10%
D. Recent-DenseRecent-denseIncreasing0.3018 ± 0.0007+0.57%
E. LearnedLearned orderedIncreasing0.2999 ± 0.0008−0.07%

L=512; Avg. MSE equally averages 7 datasets and 4 horizons; DeltaMSE is relative to ARBR. Increasing scales outperform constant/reversed scales. Recent-dense centers do not help.

We also inject equal-strength local noise into recent or distant history and measure degradation in frozen models.

Table R2. Position-dependent perturbations

Reference scalesRecent ΔMSE ↓Distant ΔMSE ↓Recent − Distant ↓
Increasing (ARBR)+3.1%+2.2%+0.9 pp
Constant+4.2%+1.9%+2.3 pp
Decreasing+5.9%+1.4%+4.5 pp

Increasing versus decreasing scales reduces recent degradation (5.9% to 3.1%) but increases distant degradation (1.4% to 2.2%).

W2/Q2. Signal fitting and noise suppression

In the controlled task, ARBR-G improves clean-signal recovery and noise robustness, while the local residual adds detail and noise sensitivity.

We add controlled synthetic reconstruction with clean ground truth. Fixed Gaussian, Global Learned Gaussian, Cosine, and ARBR retain the same encoder/coefficient-generator architectures, latent dimension, and prediction head; only basis functions and geometry generation differ.

After future-MSE training, predictors are frozen. Separate equal-capacity linear readouts reconstruct histories and are trained only on clean inputs, without noisy-to-clean denoising supervision. ARBR-G and ARBR-Z evaluate the radial representation and its fusion with the local residual. Readouts are patch-shared 128-to-24 linear maps (3,096 parameters), selected by clean-validation reconstruction MSE, without noisy-test selection.

Table R3. Frozen-representation reconstruction

RepresentationClean-fit NMSE ↓Noisy-input NMSE ↓Noise Transfer ↓
Noisy Input—0.1001.000
Fixed Gaussian-G0.055 ± 0.0030.081 ± 0.0040.30 ± 0.02
Global Gaussian-G0.049 ± 0.0020.076 ± 0.0030.26 ± 0.02
Cosine-G0.060 ± 0.0030.093 ± 0.0040.33 ± 0.02
ARBR-G0.038 ± 0.0020.054 ± 0.0020.19 ± 0.01
ARBR-Z0.031 ± 0.0020.068 ± 0.0030.37 ± 0.02

L=512, H=96, K=8, D=128; SNR=10 dB. Values are mean ± SD across three independent training seeds. Clean-fit NMSE measures clean-input reconstruction error; Noisy-input NMSE compares noisy-input reconstructions against clean truth. Noise Transfer measures transmission of independent input-noise changes to reconstructed outputs. Protocols, formulas, and visualizations are in Appendix E.8.

Versus Global Gaussian-G, ARBR-G lowers Clean-fit NMSE from 0.049 to 0.038, Noisy-input NMSE from 0.076 to 0.054 (28.9%), and Noise Transfer from 0.26 to 0.19 (26.9%). Adding the local residual improves clean reconstruction to 0.031, but increases Noise Transfer to 0.37. These results support Fig. 1's motivation within this task.

W3/Q3. Prediction-head-only controls

ARBR outperforms linear and parameter-matched nonlinear direct predictors on all seven datasets.

Linear-Only uses a single linear head; Matched-MLP uses a nonlinear predictor matched to full ARBR's parameter count separately for each horizon. Both remove the sequence encoder, radial geometry, coefficient generator, and local branch, retain RevIN, and train from scratch.

Table R4. Direct-prediction controls

MetricFull ARBR ↓Linear-Only ↓Matched-MLP ↓
Avg. MSE0.3000.3270.319

L=512; MSE equally averages H={96,192,336,720}; Avg. equally weights seven datasets. Average MSE falls by 8.3%/5.9% versus Linear-Only/Matched-MLP. Added nonlinear capacity does not close the gap; together with the original ablations, these results support structured representations.

Reviewer_dmQYOpenReview ↗

Response to the Authors' Rebuttal

I would like to thank the authors for their response, which fully addressed my concerns in Q2 and Q3, and partially addressed my concern in Q1 by experimentally validating the effectiveness of ARBR's Gaussian-prior design.

Regarding Q1, I have one follow-up question. I agree that “increasing scales improve recent-noise robustness” makes sense for positions close to the forecast point, where flat/wide bases are used. However, using sharp/narrow bases at locations farther away from the prediction point still seems counterintuitive to me. I wonder whether there is a more intuitive explanation for this point.

AuthorsOpenReview ↗

Author Follow-up Response to Reviewer dmQY

Thank you for the thoughtful follow-up. We agree that our previous response did not sufficiently explain the role of narrower distant bases.

The key distinction is between a pattern’s temporal duration and its distance from the forecast point. For example, a brief event at a particular phase of an earlier cycle may inform future observations at a corresponding phase. In such a case, distinguishing that historical position from neighboring positions may be useful, even though the predictive dependency spans a long interval. This motivates retaining temporally localized components in the earlier part of the representation, rather than requiring every distant component to have broad temporal coverage.

In ARBR, a narrow basis distributes its input-conditioned coefficient over a more localized set of patch positions; it does not restrict which input positions can inform that coefficient or which future steps it can influence. Both geometry and coefficients are conditioned on the entire input window, and the forecasting head combines all patch representations. Thus, temporal localization in the rendered representation is compatible with long-range forecasting.

Our design intuition is to balance temporal specificity in earlier components with perturbation tolerance near the forecast origin. Narrower distant bases provide the former, while broader recent bases are intended to encourage the latter. The local residual pathway additionally provides access to patch-level details, including recent fine-scale information.

We do not claim that distant positions inherently require narrower bases. The increasing-scale scaffold is an empirically supported reference prior, with bounded input-adaptive adjustments, rather than a universal rule. Tables R1–R2 show that, compared with decreasing scales, increasing scales reduce recent-noise degradation from 5.9% to 3.1%, while increasing distant-noise degradation from 1.4% to 2.2%, and improve average forecasting MSE from 0.3034 to 0.3001. These results support the overall trade-off.

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

This paper proposes Adaptive Radial Basis Representation (ARBR) for long-term time series forecasting, motivated by the fact that predictive patterns can shift in time and vary in duration. For each input window, ARBR generates Gaussian basis functions with adaptive centers and scales, together with input-dependent vector coefficients. Bounded adjustments around a shared temporal scaffold preserve the ordering of the basis components, while a local residual pathway adds patch-level information. Across seven benchmarks, ARBR achieves the lowest average MSE on five datasets and the lowest average MAE on four, with ablations supporting the value of adaptive geometry. The results are promising, although the experiments do not establish that the learned geometry corresponds to genuinely predictive events or that the gains extend to substantial distribution shifts.

Critical strengths
  1. The method has clear structural constraints. The centers move within bounded ranges around shared anchors, allowing input adaptation while preserving the temporal ordering of the components.

  2. The ablation studies are well targeted. The paper separately removes center, scale, and coefficient adaptation, and compares the proposed method with fixed geometry, globally learned geometry, and dynamic mixing.

Critical weaknesses

1.The evidence for the proposed mechanism is limited. The parameters respond to artificial shifts and rescaling, but this does not establish that they identify patterns that are genuinely predictive.

2.The performance gains are inconsistent. ARBR trails the best baseline on Traffic and is not the top performer on ETTh2; its gains on datasets such as Electricity are small.

3.The scope of validation is limited. The model processes variables independently and does not directly capture relationships between them. The experiments also do not assess performance under substantial distribution shifts.

Questions for authors

1.How can the authors verify that the geometric parameters correspond to predictive patterns? On data with known event locations and durations, could they quantify how well the learned centers and scales align with those events?

2.Would the main-table rankings hold under a common input length and comparable parameter counts? This is particularly relevant to the comparisons with PatchTST, PhaseFormer, and BasisFormer.

3.How much does each of the radial and local residual pathways contribute? Beyond ablating the residual pathway, could the authors report the output magnitudes of both pathways or perform intervention experiments?

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer YrZm

Thank you for highlighting predictive relevance. Default seed: 2021.

Q1/W1. Predictive relevance

Controlled tests support predictive-event alignment and adaptive geometry's predictive value.

Known-event forecasting uses future MSE without geometry labels and prespecified event–anchor matching.

Table R1. Known-event alignment

ModelCenter MAE ↓Log-scale MAE ↓Forecast MSE ↓
ARBR0.01770.1410.0251
w/o Center Adapt.0.02510.1620.0273
w/o Scale Adapt.0.01960.2430.0266
Fixed Scaffold0.02510.2430.0294

L=512, H=96, K=8, rho=beta=0.5. Errors use normalized time/natural-log widths. Versus Fixed Scaffold, center/scale/MSE errors fall by 29.5%/42.0%/14.6%. Protocol: Appendix E.7.

Only predictive-event perturbations change the true future. CF Error compares predicted and true future changes. Frozen-model swaps use original (Reverted) or unrelated (Random) geometry, fixing perturbed-input coefficients, local output, head, and normalization; MSE uses counterfactual targets.

Table R2. Counterfactual geometry interventions

EventCF Error (ARBR / Fixed) ↓Full MSE ↓Reverted MSE ↓Random MSE ↓ΔGeo ↑
Predictive0.0156 / 0.02040.02750.03020.0321+9.81%
Nuisance0.0037 / 0.00510.02630.02690.0282+2.28%

DeltaGeo=100×(Reverted MSE/Full MSE-1). Reversion increases MSE by 9.81% for predictive versus 2.28% for nuisance events.

Q2/W2. Unified length and parameter budgets

With L=512 and approximately matched parameters, ARBR has the lowest reported MSE on all seven datasets.

Retraining uses common splits/horizons/metrics, L=512, and approximate parameter matching. For close PatchTST comparisons, ARBR/PatchTST use seeds 42,2021,2026 to assess training variability.

Table R3. Parameter-matched comparison

DatasetARBRPatchTSTPhaseFormerBasisFormer
ETTh10.4038±0.00080.4233±0.00160.4390.441
ETTh20.3344±0.00120.3375±0.00120.3610.368
ETTm10.3453±0.00100.3553±0.00090.3700.374
ETTm20.2488±0.00180.2578±0.00140.2620.275
Weather0.2193±0.00110.2341±0.00150.2250.231
Electricity0.1601±0.00100.1694±0.00130.1810.175
Traffic0.3955±0.00030.3963±0.00060.4040.427

MSE averages H={96,192,336,720}. Parameters: ARBR/PatchTST/PhaseFormer/BasisFormer=2.01/2.09/2.03/1.97M. See Appendix B.3.

Q3. Branch contributions

Both branches contribute, but radial removal increases MSE more on every tested dataset.

One frozen checkpoint measures branch norms at Z=G+lambda×L_local and branch removal, fixing the head and RevIN.

Table R4. Branch diagnostics

Dataset‖G‖₂‖λL_local‖₂ΔMSE_local (%)ΔMSE_radial (%)
ETTh11.7830.857+1.20%+76.08%
ETTh22.2551.037+0.51%+16.99%
ETTm11.8220.763+5.64%+85.77%
ETTm22.4511.092+4.22%+39.32%
Weather1.9850.677+10.38%+28.18%
Electricity1.4310.652+9.36%+47.42%
Traffic1.0960.662+12.13%+22.44%

Mean L2 norms; DeltaMSE averages four horizon-wise relative increases. Local contributions vary.

W3. Cross-variable dependencies and shifts

CI isolates temporal geometry, and ARBR outperforms both transfer controls in all four tested directions.

Cross-variable comparisons [1–3] follow.

Table R5. Cross-variable baselines

DatasetsARBRTimeMixer++VPNetMS-Flow
Etth10.4030.4190.4340.416
Etth20.3330.3390.3560.306
Ettm10.3440.3690.3760.368
Ettm20.2470.2690.2700.264
Weather0.2180.2260.2380.234
Electricity0.1590.1650.1620.157
Traffic0.3950.4160.4210.440

Zero-shot transfer uses source-only training, frozen parameters, and no target tuning/selection. L=512; H={96,192,336,720}. Standardization uses source-training statistics; RevIN uses current windows.

Table R6. Cross-site transfer

Source → TargetMMD² Within-source / Cross-domainARBR Target-domain ID MSE ↓ARBR OOD MSE ↓Fixed Scaffold OOD MSE ↓PatchTST OOD MSE ↓ARBR Transfer Gap ↓
ETTm1 → ETTm20.018 / 0.0790.2470.3460.3630.372+40.08%
ETTm2 → ETTm10.021 / 0.0750.3440.4670.4880.498+35.76%
ETTh1 → ETTh20.026 / 0.0910.3330.4170.4290.439+25.23%
ETTh2 → ETTh10.024 / 0.0860.4030.5010.5230.528+24.32%

MSE averages four horizons; ID/OOD train on target/source. Gap=100×(OOD/ID-1). MMD²: normalized 512-step windows, unbiased RBF estimate, source median bandwidth, 1,024 windows, 3 sampling seeds, 7-channel mean. Transfer degrades by 24.32%–40.08%.

References:

[1] Zhang et al., “What If We Let Forecasting Forget?”, ICML 2026.

[2] Liu et al., “Are Global Dependencies Necessary?”, ICLR 2026.

[3] Wang et al., “TimeMixer++”, ICLR 2025.

Reviewer_K9EHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the ARBR model for time-series forecasting, utilizing Gaussian radial basis functions that adapt to the input data. The centers and scales of these basis functions are adjusted based on a shared temporal framework, while input-dependent coefficients generate patch-aligned representations. A local residual pathway is incorporated to capture additional temporal details. Experiments across seven datasets demonstrate highly competitive forecasting performance, with controlled comparisons highlighting the advantages of an adaptive radial geometric structure.

Critical strengths
  1. The paper is well-written and generally easy to follow.
  2. ARBR achieves the lowest average MSE (across forecasting horizons) on five out of the seven datasets used for primary comparison.
  3. Controlled representation comparisons and multi-seed experiments reinforce the benefits of an input-adaptive radial geometric structure.
  4. The idea of adapting radial basis centers and scales to capture changes in temporal location and extent is interesting.
Critical weaknesses
  1. The advantage of maintaining a specific order for the centers remains unverified, particularly when compared to a variant with an unconstrained adaptive geometric structure.
  2. Specific parameters regarding input sequence lengths and the configurations of baseline models are not clearly stated, making it difficult to discern the impact of varying context lengths.
  3. The paper lacks comparisons against language-model-based forecasting models, such as Chronos [1], SCALER [2], and TimeLLM [3].

[1] Ansari, A. F., et al. Chronos: Learning the Language of Time Series. Transactions on Machine Learning Research, 2024.

[2] Le, X.-M., et al. Efficient Test-Time Scaling for LLM-based Time Series Forecasting (SCALER). KDD, 2026.

[3] Jin, M., et al. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. ICLR, 2024.

Questions for authors

See the weakness section!

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer K9EH

Thank you for your positive assessment of the presentation, multi-seed experiments, and adaptive radial geometry. We address your three concerns below.

W1/Q1. Center ordering

Preserving center order reduces MSE at the same displacement bound.

Equation (10) guarantees ordered centers for ρ≤0.5\rho\leq0.5. We compare original ARBR (A) with two retrained variants at ρ=1\rho=1: B applies a bounded order-preserving projection, retaining center–scale–coefficient associations; C permits crossings.

Table R1. Center-ordering controls

VariantρOrderingAvg. MSE ↓Raw CrossingFinal Crossing
A. ARBR0.5Bounded0.3000%0%
B. Relaxed-Ordered1.0Projection0.303raw 8%0%
C. Relaxed-Unordered1.0None0.30612%12%

L=512; all 7 datasets and 4 horizons. Avg. MSE equally weights datasets. Crossing rates are proportions of test windows with an inverted adjacent pair, before (Raw) and after (Final) constraints. At the same bound, B reduces MSE by 0.98% versus C; A further reduces it by 0.99% versus B. These results support ordering and bounded local anchoring.

W2/Q2. Unified input length

At L=512, ARBR achieves the lowest or tied-lowest horizon-averaged MSE on all seven datasets.

In the original main table, ARBR uses L=512; baselines use official configurations.

Table R2. Unified L=512 comparison

DatasetARBRTimeGSLGAMixLinearPhaseFormerTimeBaseTimeMixerPatchTSTBasisFormer
ETTh10.4030.4150.4160.4060.4590.4190.4320.4420.452
ETTh20.3330.3560.3380.3440.3800.3580.3530.3440.377
ETTm10.3440.3820.3590.3960.3880.3610.3730.3550.384
ETTm20.2470.2610.2640.2570.2690.2590.2710.2570.287
Weather0.2180.2280.2380.2420.2290.2420.2320.2340.235
Electricity0.1590.1690.1680.1650.2040.1670.1920.1690.181
Traffic0.3950.4350.4660.4330.4180.3950.4200.3960.436

All models use seed 2021. MSE equally averages H={96,192,336,720}. Configurations are detailed in Appendix B.3. Representative original settings are summarized below.

Table R3. Original configurations

ModelOriginal LKey settings
ARBR512D=128, K=8, patch/stride=24/12
TimeGS96grid=24×24, hidden=8/16
LGA336 (ETTh), 512 (ETTm/Weather/Electricity)3 layers, d_G=32/64
MixLinear720Periodic/frequency mixing
PhaseFormer720Phase-based representation
TimeBase720Periodic basis representation
TimeMixer963 downsampling layers
PatchTST336patch/stride=16/8, 3 layers
BasisFormer96d_model=100, 10 bases

The unified-length results indicate that input-length differences alone do not explain ARBR's competitiveness. This comparison controls input length while retaining model-specific architectures. It is separate from the approximately parameter-matched comparison in Appendix B.3.

W3/Q3. Chronos, Time-LLM, and SCALER

ARBR remains competitive with far fewer trainable parameters, although SCALER at L=672 achieves lower MSE on all seven datasets.

We use official implementations and paper configurations: pretrained Chronos-T5-Small, and LLaMA-backed Time-LLM and SCALER. We evaluate L=512 throughout and also SCALER's paper setting, L=672.

Table R4. Additional forecasting baselines

DatasetARBR (L=512)Chronos (L=512)Time-LLM (L=512)SCALER (L=512)SCALER (L=672)
ETTh10.4030.4110.4090.4060.391
ETTh20.3330.3380.3540.3430.328
ETTm10.3440.3410.3510.3470.334
ETTm20.2470.2450.2540.2510.244
Weather0.2180.2200.2260.2210.217
Electricity0.1590.1640.1590.1590.154
Traffic0.3950.3910.3880.3920.384

MSE averages four horizons. Trainable parameters: ARBR 2.01M; Chronos 46.15M; Time-LLM 43.34M; SCALER (512/672) 27.99M/35.20M. Time-LLM and SCALER each also contain about 7B frozen parameters.

ARBR outperforms Chronos on 4/7 datasets and Time-LLM and SCALER (L=512) on 5/7. These results support competitiveness under the evaluated settings.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Summary of Rebuttal and Revisions

We thank the reviewers and Area Chair for their feedback. We summarize the main findings from the additional experiments below.

1. Predictive relevance and adaptive geometry (YrZm, K9EH). In controlled synthetic tasks, adaptive basis centers and scales aligned more closely with the locations and durations of known predictive events than the fixed scaffold. With the rest of the model held fixed, reverting the geometry to its pre-perturbation values increased forecast MSE by 9.81% after perturbing predictive events and by 2.28% after perturbing nuisance events (Appendix E.7). A separate control showed that preserving center order reduced MSE at the same displacement bound.

2. Controlled comparisons and zero-shot transfer (YrZm, K9EH). With a common input length of L=512 and approximately matched trainable parameter counts, ARBR achieved the lowest reported horizon-averaged MSE among the matched models on all seven benchmarks (Appendix B.3). At L=512, ARBR used 2.01M trainable parameters and achieved lower MSE than SCALER on five of seven datasets. At its original input length of L=672L=672, SCALER achieved lower MSE, using 35.20M trainable parameters. In all four zero-shot ETT cross-site transfer directions, ARBR achieved lower MSE than both Fixed Scaffold and PatchTST.

3. Representation and scaffold design (dmQY, YrZm). In frozen-representation reconstruction tests, the ARBR radial branch recovered clean signals more accurately and transferred less noise than the globally learned Gaussian baseline. Adding the local residual further improved clean reconstruction (Appendix E.8). In frozen-checkpoint tests, removing either branch increased forecast MSE, with a larger increase when removing the radial branch on all seven datasets. Full ARBR also outperformed both the linear-head-only and parameter-matched MLP controls. Under recent-history noise, increasing scales reduced the MSE increase from 5.9% to 3.1% relative to decreasing scales. Basis width governs temporal coverage in the representation, and we use increasing scales as an empirically supported design prior.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.