acceptodds
Under review as a conference paper at ICLR 2027

NEIN-TSF: Not Every Irregularity Is Noise in Time-Series Forecasting

Abstract

Modern time-series forecasting models capture increasingly fine-grained historical dynamics, yet not every observed pattern should be extrapolated. Irregular changes may be transient disturbances or genuine regime shifts, while seemingly stable patterns may lose predictive relevance beyond the input window. We propose NEIN-TSF, a unified framework that regulates historical dynamics through forecasting consequence, historical support, and correction utility. Forecast-Damage Ranking Pretraining (FDRP) learns relative teacher-derived forecast damage from matched local counterfactuals, encouraging the backbone to capture predictive consequence rather than input-space deviation. Supported Predictive Evidence (SPE) combines historical retrieval, periodicity, local trends, and persistence to produce bounded, support-conditioned residual corrections. Horizon-Adaptive Evidence Utility (HAEU) calibrates correction strength across prediction positions, while Chronological Forward Utility Selection (CFUS) selects model states on chronologically later validation data. Experiments on fifteen real-world benchmarks across multiple horizons show consistent MSE and MAE improvements over strong recent baselines. Mechanism analyses further show stronger alignment between FDRP-predicted and teacher-derived damage than raw input deviation, positive support-benefit associations, and horizon- and time-dependent correction utility.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Forecasting, Non-Stationarity, Counterfactual Learning, Historical Evidence, Horizon-Adaptive Forecasting

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_8SoaOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper discusses the challenge in multivariate time-series forecasting, where the authors are distinguishing observed historical irregularities (such as transient surges or abrupt shifts) that carry predictive value for future horizons from those that act as transient noise. They propose NEIN-TSF (Not Every Irregularity is Noise in Time-Series Forecasting:

  • the first component is FDRP (Forecast-Damage Ranking Pretraining), the idea is to train a forecasting backbone fθf_\theta to capture predictive consequences rather than raw input-space deviations;

  • the second component is Supported Predictive Evidence (SPE) With the pretrained backbone fθ0f_{\theta_0} frozen to produce base forecast Y^0=fθ0(X)\hat{Y}^0 = f_{\theta_0}(X), SPE constructs a bank of historical candidates paired with support scores SkS^k.

  • The third component is Utility-Controlled Evidence Extrapolation (HAEU + CFUS), Horizon-Adaptive Evidence Utility (HAEU) fits position-wise scalar multipliers αh∈[0,αmax⁡]\alpha_h \in [0, \alpha_{\max}] via ridge-regularized least squares on calibration split Dcal\mathcal{D}_{\text{cal}}.

Critical strengths

I found the Consequence-Aware Formulation is novel, in fact rather than relying on heuristic input-space anomaly thresholds or unconditional extrapolation, NEIN-TSF introduces forecast-consequence learning through teacher-derived damage ranking (FDRP) and couples it with support-bounded residual refinement (SPE). This provides explicit structural bounds on maximum correction magnitude (∣ΔYb,h,c∣≤γσb,cSb,h,c|\Delta Y_{b,h,c}| \le \gamma \sigma_{b,c} S_{b,h,c}) and prevents catastrophic forecast drift. Also they have provided empirical validation thorough, they used 15 diverse benchmarks, multi-seed variance bounds.

Critical weaknesses

I am bit concerned about the multi-stage offline training, basically they have proposed pipeline involves multiple sequential training, calibration, and selection steps: (i) pretraining a teacher forecaster fθˉf_{\bar{\theta}}, (ii) counterfactual generation and FDRP backbone pretraining via pairwise damage ranking, (iii) freezing fθ0f_{\theta_0} and optimizing SPE evidence fusion networks, (iv) fitting HAEU ridge coefficients and (v) CFUS checkpoint selection on Dfwd\mathcal{D}_{\text{fwd}}.

However what I noticed inference remains lightweight and single-pass in equation 2, the paper lacks quantitative reporting on total offline training time and FLOPs relative to end-to-end baseline models. Maybe reporting wall-clock training times and memory usage for each stage across representative datasets would clarify offline overhead.

Questions for authors

Could you provide a breakdown of wall-clock training times and memory footprints for each of the offline stages (Teacher training, FDRP pretraining, SPE optimization, HAEU calibration, CFUS selection) compared to training standard end-to-end models like iTransformer or PatchTST on large datasets (such as ECL and Traffic)?

For HAEU, one scalar αh\alpha_h is fit per horizon position hh, shared across all variables c∈{1,…,C}c \in \{1, \dots, C\}. Have you considered variable-specific or channel-grouped utility coefficients αh,c\alpha_{h,c}, especially for highly heterogeneous multivariate datasets like Weather or Traffic where variables exhibit vastly different noise levels and persistence lifetimes?

5 comments
AuthorsOpenReview ↗

Author Response to 8Soa (1/3): Weakness and Q1

We thank the reviewer for recognizing the consequence-aware formulation, the explicit bounds on residual correction, and the breadth of evaluation. We address the requested computational transparency and HAEU coefficient granularity with stage-level profiling (E3) and controlled coefficient-granularity comparisons (E6).

Weakness and Q1. What does the multi-stage offline procedure cost?

We agree that the practical value of improved forecasting accuracy must be considered alongside computational cost. The distinct stages of NEIN-TSF serve complementary purposes: teacher-based consequence supervision, FDRP representation learning, support-conditioned SPE correction, horizon-wise utility calibration, and chronological checkpoint selection. We therefore profiled the wall-clock time and peak GPU allocation of the five core operations, as requested.

Table R6a — E3-A: stage-wise offline profiling of NEIN-TSF. GPU peak is the maximum observed within each named stage (and the maximum across stages in the Total row), not the sum.

Offline operationWall-clock (s)Wall-clock (min)GPU peak (GiB)
Teacher training108.2761.8050.092
FDRP pretraining1,665.49527.7581.851
SPE: evidence + optimization3,822.06463.7010.173
HAEU coefficient fitting0.0170.0000.069
CFUS selection0.0170.0000.069
Sum of five stages5,595.86993.2641.851

Additional Comparison with End-to-End Forecasting Baselines

To further address the reviewer's concern regarding computational overhead, we additionally reproduce iTransformer and PatchTST on Electricity (H=96H=96) using an NVIDIA GeForce RTX 5080. We report their measured offline wall-clock time, peak GPU memory, model parameters, and inference latency.

Table R6b. Computational cost of reproduced end-to-end forecasting baselines on Electricity (H=96H=96).

ModelOffline Time (s)Peak GPU Memory (GiB)Inference Latency (ms/batch)
iTransformer987.601 ± 198.8410.1411.293 ± 0.230
PatchTST20,229.311 ± 5,759.0211.08624.595 ± 0.079

All measurements were obtained on an NVIDIA GeForce RTX 5080 using two recorded seed runs (2022 and 2023). The current baseline records contain two completed seed measurements (2022 and 2023); the reported offline times reflect the recorded experimental units. These results complement the stage-wise computational analysis of NEIN-TSF in Table R6.

The largest measured costs among these operations are FDRP pretraining and SPE evidence construction/optimization, whereas coefficient fitting and the final checkpoint-selection calculation themselves are lightweight. This decomposition clarifies where the reported framework's computational work is concentrated.

AuthorsOpenReview ↗

Author Response to 8Soa (2/3): Weakness and Q1

We also provide a reference-protocol resource comparison on Electricity and Traffic following the efficiency setup of OLinear . To maximize transparency, we reproduce all 12 baseline rows from OLinear's published Table 41 (corresponding to Table 37 in an earlier version) and add our newly measured NEIN-TSF inference footprint. The baseline rows are published measurements from OLinear, not new re-runs on our GPU, so cross-study latency/memory numbers should not be interpreted as controlled same-hardware speedups.

Table R7a. E3-B Computational efficiency on Electricity (ECL).

ModelParams (M)FLOPs (G)T.T. (ms/iter)T.M. (GB)I.T. (ms/iter)I.M. (GB)
OLinear4.7901.6507.7500.4502.1100.170
OLinear-C4.5801.6506.9400.4202.0900.170
Leddam8.5605.32065.1901.44014.7800.490
CARD1.3905.07078.2505.08028.2100.630
Fredformer12.1205.55026.5901.2407.0800.260
iTransformer4.8301.87010.6600.7003.4600.230
TimeMixer++1.19064.020106.0601.02034.1000.360
TimeMixer0.1201.74057.0904.24022.7900.820
DLinear0.0200.0101.3700.0400.2000.020
TimesNet300.580293.980509.1705.820157.1301.410
PatchTST3.76012.48065.8303.02020.5200.930
FreTS0.4200.1303.7800.3000.4800.230
NEIN-TSF0.0280.30517.4350.1783.1990.234

Table R7b. E3-B Computational efficiency on Traffic.

ModelParams (M)FLOPs (G)T.T. (s/iter)T.M. (GB)I.T. (ms/iter)I.M. (GB)
OLinear6.1704.9100.0201.0105.7100.430
OLinear-C4.6904.9100.0200.9505.6700.430
Leddam8.56015.2400.1903.74045.6801.250
CARD0.98010.4700.1809.50061.2101.690
Fredformer11.09014.7800.1005.76031.6001.840
iTransformer4.8306.4500.0402.95013.4501.270
TimeMixer++4.730588.470OOMOOM1610.6609.350
TimeMixer0.1204.6600.17011.33067.2102.200
DLinear0.0200.0200.0010.0700.2000.040
TimesNet301.690288.7200.5005.86015.6401.420
PatchTST3.76033.5200.1708.01055.1002.440
FreTS0.4200.3600.0100.7803.4800.590
NEIN-TSF0.0310.9240.0200.5648.1580.572

T.T. and T.M. denote training time and GPU memory; I.T. and I.M. denote inference time and GPU memory. Baseline results are taken from OLinear's published efficiency table; NEIN-TSF results are newly measured. For NEIN-TSF, T.T./T.M. refer to the profiled backbone fine-tuning iteration, while inference measurements cover the full prediction pipeline.

The results provide additional evidence of NEIN-TSF's compact parameterization and practical inference resource requirements under the reference protocol. Detailed offline costs of its five principal stages are reported separately in Table R6. Source: OLinear, published supplementary Table 41.

AuthorsOpenReview ↗

Author Response to 8Soa (3/3): Q2

Q2. Why share a horizon-wise coefficient across channels rather than fitting channel-specific coefficients?

We agree that channel heterogeneity makes this an important design choice. HAEU intentionally uses one coefficient α_h per forecast position shared across variables, controlling the number of calibration parameters under chronologically limited calibration data. We compared a global scalar, the proposed horizon-wise scalar, channel-grouped coefficients with 8 and 16 groups, and channel-specific α_(h,c). All variants use the same ridge coefficient 10^-4, and coefficient clipping to [0, 1.2].

Table R8 — E6: Weather, calibration granularity (three-seed means). Lower MSE/MAE is better; clip ratio is descriptive, not a metric to minimize independently.

HAEU calibrationCal. MSEForward MSETest MSETest MAEClip ratio
Global α0.2550.2460.2350.2580.333
Horizon-wise α_h0.2530.2430.2330.2560.064
Grouped-80.2490.2440.2350.2570.091
Grouped-160.2470.2450.2340.2570.132
Channel-specific α_h,c0.2440.2480.2350.2580.181

Channel-specific calibration fits the calibration set most closely but does not generalize as well to later periods: its calibration MSE is 0.244, compared with 0.253 for horizon-wise HAEU, while its forward-validation and test MSEs are 0.248 and 0.235, respectively. The proposed horizon-wise HAEU obtains the lowest forward-validation MSE (0.243) and test MSE/MAE (0.233 / 0.256). This suggests that sharing α_h offers a favorable flexibility–generalization trade-off under the tested Weather setting.

We thank the reviewer for these actionable requests, which prompted clearer quantification of the computation–performance trade-off and direct validation of the coefficient-sharing design.

Reviewer_R3tbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces NEIN-TSF, a time-series forecasting framework based on the idea that not every irregular historical pattern should be carried into future predictions. It uses FDRP to learn which changes matter for forecasting and SPE to refine predictions using supported historical patterns such as retrieval, periodicity, trends, and persistence. HAEU adjusts the strength of these corrections across forecast horizons, while CFUS selects models using a later validation period to better handle changes over time. Experiments on 15 benchmarks show consistent improvements over strong baselines, supported by extensive ablations, mechanism analyses, and experiments across different forecasting backbones

Critical strengths

Well-motivated and coherent framework - The method follows naturally from three clear questions: whether a historical change matters, whether there is historical evidence for a correction, and how strongly the correction should affect different forecast horizons. FDRP, SPE, and HAEU/CFUS each address one of these questions, making the overall design easy to follow.

Strong validation of the core idea - The paper does not rely only on final forecasting accuracy. Its motivating analysis, mechanism studies, and ablations directly test the main claim that predictive importance cannot be determined from irregularity magnitude alone.

Careful and conservative correction design - Rather than freely modifying the forecast, SPE applies a bounded residual correction based on historical support, while HAEU can reduce or suppress the correction depending on the forecast horizon. This provides useful safeguards against unreliable corrections.

Critical weaknesses

Modest gains relative to the added complexity - NEIN-TSF introduces several additional components and training stages, while some individual component gains are small and close to the reported seed variation. Reporting variance or significance tests for the ablations would make these contributions more convincing.

Dependence on a single teacher model - FDRP defines forecast damage using one frozen teacher, meaning that the learned notion of “forecasting consequence” may depend on that particular teacher. Testing different teachers would help show whether this signal is more generally meaningful.

Limited generality of the predefined evidence bank - SPE relies on four manually chosen evidence types with fixed settings, including predefined periodic lags. Although the method performs well across datasets, it remains unclear how well these assumptions generalize to time series with very different temporal patterns.

Questions for authors

Please refer to the weaknesses.

6 comments
AuthorsOpenReview ↗

Author Response to R3tb (1/2): W1

We thank the reviewer for appreciating the coherent three-question formulation, the empirical mechanism evidence, and the conservative design of NEIN-TSF. We address the three concerns with paired-seed component analysis (E4), teacher sensitivity (E5), and evidence-bank/periodicity sensitivity (E7).

W1. Are component gains meaningful relative to seed variability and the additional complexity?

We agree that the variability of Full NEIN-TSF alone does not establish the reliability of individual components. We therefore added paired ablations with the same three seeds (2022, 2023, 2024) under aligned data and evaluation settings. For each seed, we compute ΔMSE = MSE(variant) − MSE(Full) and analogously for MAE; a positive delta means the Full model performs better. Importantly, the reported standard deviation of the paired delta is calculated from within-seed differences, not by subtracting two standard deviations.

Table R3 — E4: three-seed paired module ablations. All means are reported to three decimal places; standard deviations retain the measured precision. Win seeds indicate the number of seeds for which Full improves over the indicated variant.

DatasetVariantMSE (mean ± std) ↓MAE (mean ± std) ↓Paired ΔMSE (mean ± std) ↑Paired ΔMAE (mean ± std) ↑Win seeds (MSE / MAE)
ETTh1Vanilla0.416 ± 0.004250.429 ± 0.001760.020 ± 0.001380.014 ± 0.001703/3 | 3/3
ETTh1Vanilla + FDRP0.407 ± 0.003740.422 ± 0.000950.011 ± 0.001510.007 ± 0.001453/3 | 3/3
ETTh1Vanilla + SPE0.411 ± 0.001850.425 ± 0.000200.015 ± 0.001500.010 ± 0.001743/3 | 3/3
ETTh1FDRP + SPE0.402 ± 0.001560.418 ± 0.002060.006 ± 0.001330.003 ± 0.002783/3 | 2/3
ETTh1w/o FDRP0.405 ± 0.002000.422 ± 0.002400.009 ± 0.001530.007 ± 0.001253/3 | 3/3
ETTh1w/o SPE0.405 ± 0.003580.421 ± 0.001800.009 ± 0.001460.006 ± 0.001593/3 | 3/3
ETTh1w/o HAEU0.400 ± 0.004450.416 ± 0.004660.004 ± 0.002380.001 ± 0.003023/3 | 2/3
ETTh1w/o CFUS0.399 ± 0.006000.417 ± 0.003540.003 ± 0.003140.002 ± 0.002552/3 | 3/3
ETTh1Full NEIN-TSF0.396 ± 0.002890.415 ± 0.00173———
WeatherVanilla0.247 ± 0.002770.269 ± 0.001150.014 ± 0.001720.013 ± 0.001283/3 | 3/3
WeatherVanilla + FDRP0.241 ± 0.003200.264 ± 0.000260.008 ± 0.001550.008 ± 0.001663/3 | 3/3
WeatherVanilla + SPE0.243 ± 0.003170.263 ± 0.002520.010 ± 0.001460.007 ± 0.001573/3 | 3/3
WeatherFDRP + SPE0.237 ± 0.003290.259 ± 0.003130.004 ± 0.002800.003 ± 0.002883/3 | 2/3
Weatherw/o FDRP0.241 ± 0.000060.263 ± 0.003120.008 ± 0.001760.007 ± 0.001403/3 | 3/3
Weatherw/o SPE0.239 ± 0.002690.260 ± 0.002770.006 ± 0.001230.004 ± 0.001723/3 | 3/3
Weatherw/o HAEU0.236 ± 0.001210.257 ± 0.004590.003 ± 0.002560.001 ± 0.002863/3 | 2/3
Weatherw/o CFUS0.235 ± 0.001810.258 ± 0.004120.002 ± 0.002380.002 ± 0.002433/3 | 2/3
WeatherFull NEIN-TSF0.233 ± 0.001730.256 ± 0.00173———
ECLVanilla0.169 ± 0.002190.257 ± 0.002620.013 ± 0.001230.012 ± 0.001663/3 | 3/3
ECLVanilla + FDRP0.164 ± 0.002110.252 ± 0.002330.008 ± 0.001720.007 ± 0.001353/3 | 3/3
ECLVanilla + SPE0.165 ± 0.001480.251 ± 0.000400.009 ± 0.001700.006 ± 0.001373/3 | 3/3
ECLFDRP + SPE0.160 ± 0.002690.248 ± 0.003250.004 ± 0.001710.003 ± 0.002253/3 | 3/3
ECLw/o FDRP0.164 ± 0.001370.251 ± 0.000780.008 ± 0.001410.006 ± 0.001733/3 | 3/3
ECLw/o SPE0.163 ± 0.002190.250 ± 0.001570.007 ± 0.001310.005 ± 0.001213/3 | 3/3
ECLw/o HAEU0.159 ± 0.001570.246 ± 0.001820.003 ± 0.002520.001 ± 0.002813/3 | 2/3
ECLw/o CFUS0.158 ± 0.002870.247 ± 0.003600.002 ± 0.003120.002 ± 0.003062/3 | 2/3
ECLFull NEIN-TSF0.156 ± 0.001000.245 ± 0.00100———

Removing FDRP or SPE produces a consistent performance penalty on ETTh1, Weather, and Electricity, with Full winning on all three paired seeds in both MSE and MAE for these two component removals. The gains associated with HAEU and CFUS are smaller and do not hold for every seed/metric, which we explicitly acknowledge. Their role is to provide additional correction-strength and chronological-selection control, rather than to claim that every module contributes an equally large or statistically significant gain. The paired analysis directly addresses the reviewer's concern more informatively than an overall Full-model standard deviation alone.

AuthorsOpenReview ↗

Author Response to R3tb (2/2): W2+W3

W2. Does FDRP depend on one particular frozen teacher?

Thank you for identifying the potential teacher dependence. We repeated the FDRP experiment with three independently initialized frozen teachers (seeds 2022, 2023, and 2024) under the same chronological data partitions, counterfactual damage definition, and downstream evaluation protocol. We report both the resulting forecasting performance and forecast-damage ranking accuracy.

Table R4 — E5: sensitivity to the frozen teacher (H = 96). Results summarize the teacher variants.

DatasetHMSE (mean ± std)MAE (mean ± std)Ranking accuracy (mean ± std)
ETTh1960.344 ± 0.0000160.377 ± 0.00001277.800% ± 2.08%
ECL960.130 ± 0.000900.225 ± 0.0008576.300% ± 1.93%

The small forecasting variability and the ranking accuracies observed for both ETTh1 and Electricity suggest that the learned signal is not solely an artifact of a single specific teacher initialization within the tested range.

W3. Is the fixed four-source evidence bank, especially periodic lags, too restrictive?

We agree that different datasets can exhibit different periodic structure. The original paper already includes leave-one-source-out and single-source evidence ablations. To further probe the predefined periodic candidates, we kept the remaining SPE settings fixed and compared shorter lags, the default lag set, longer lags, and a Top-4 lag set estimated using only training-set autocorrelation.

Table R5 — E7: periodic-candidate sensitivity. Mean ± standard deviation over three seeds; lower is better.

DatasetLag settingMSE (mean ± std) ↓MAE (mean ± std) ↓
ETTh1Short-lag0.400 ± 0.001410.417 ± 0.00126
ETTh1Default-lag0.396 ± 0.002890.415 ± 0.00173
ETTh1Long-lag0.402 ± 0.001450.419 ± 0.00116
ETTh1Train-autocorr Top-40.397 ± 0.001620.416 ± 0.00110
WeatherShort-lag0.236 ± 0.001340.258 ± 0.00102
WeatherDefault-lag0.233 ± 0.001730.256 ± 0.00164
WeatherLong-lag0.238 ± 0.001830.259 ± 0.00108
WeatherTrain-autocorr Top-40.233 ± 0.001670.256 ± 0.00122

The default periodic set performs best on ETTh1, while training-autocorrelation selection is comparable to the default on Weather at the reporting precision. This supports reasonable robustness of the chosen evidence bank over the evaluated alternatives while also indicating that training-only adaptive candidate selection is a promising extension.

The compared lag sets are Short = {12, 24, 48, 84}, Default = {24, 48, 96, 168}, and Long = {48, 96, 168, 336}. Invalid candidates are excluded or handled by the method's predefined fallback, and adaptive candidates are selected from the training series only.

We appreciate the reviewer's suggestions, which have helped us make the component reliability, teacher assumptions, and evidence-bank scope more explicit.

AuthorsOpenReview ↗

Additional Evidence-Bank Sensitivity Analysis on Electricity

Dear Reviewer R3tb,

We sincerely thank you for your careful and constructive review. We particularly appreciate your thoughtful observations regarding the reliability and generality of our historical evidence bank. Your suggestions have been instrumental in encouraging us to investigate the role of periodic evidence more thoroughly.

To further explore this aspect, we have extended our periodic-candidate sensitivity experiments to Electricity (ECL), comparing four candidate-lag configurations across three independent random seeds.

Additional Periodic-Evidence Sensitivity Results (MSE/MAE ↓)

Periodic candidate settingMSE ↓MAE ↓
Short-lag0.1590.247
Default-lag0.1560.245
Long-lag0.1610.249
Training-autocorr Top-40.1570.246

The default periodic configuration achieves the best performance among the evaluated alternatives, while the training-autocorrelation-based selection also yields competitive results. Together with the previously reported ETTh1 and Weather experiments, these findings provide further evidence of the effectiveness and robustness of the periodic evidence design across different temporal patterns.

We are genuinely grateful for your rigorous and constructive feedback, which has helped us strengthen the experimental foundation of our work.

We would greatly appreciate any further thoughts you may have on the additional evidence and revised manuscript, and we would be delighted to address any remaining questions.

Thank you once again for your time and valuable guidance.

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

The paper proposes NEIN-TSF, a time-series forecasting framework motivated by the observation that historical irregularities can have different forecasting consequences. Rather than treating every irregularity as noise or assuming it should persist, the framework aims to regulate its influence based on forecasts of consequences, historical support, and the utility of correction.

The method has three main stages. First, Forecast-Damage Ranking Pretraining (FDRP) trains a forecasting backbone and ranks synthetic input perturbations by the increase in forecasting error they cause on a frozen teacher model.

Second, Supported Predictive Evidence (SPE) combines historical retrieval, periodicity, local trends, and persistence to generate bounded residual corrections to the frozen backbone’s predictions.

Third, Horizon-Adaptive Evidence Utility (HAEU) calibrates the strength of corrections for each future prediction position, while Chronological Forward Utility Selection (CFUS) selects the model checkpoint using a later validation period.

Experiments across fifteen datasets and multiple forecasting horizons report improvements in MSE and MAE over recent baselines. The paper also includes component ablations, backbone-transfer experiments, and analyses of damage prediction, historical support, and correction utility.

Critical strengths

A strength of the paper is its clear motivation and practical relevance. Irregularities in historical data can have different consequences for forecasting, making it important to understand how their influence on future predictions should be regulated.

The method is thoughtfully designed to make cautious corrections to the base forecast. The paper presents extensive experimental analysis. The evaluation includes fifteen datasets, multiple forecasting horizons, backbone transfer, and detailed ablations. These results support the framework's practical usefulness in improving Forecasting performance and the contributions of its components.

Critical weaknesses

Although the method is clearly explained and the evaluation is comprehensive, the results mainly establish improved forecasting performance. They do not clearly show whether the improvement comes from appropriately regulating/attending irregularities in the input. Please evaluate cases with temporary disturbances and changes that may or may not contain useful information about the future. These could be controlled synthetic perturbations designed to be either relevant or irrelevant to the forecasting target. Evaluating the proposed method in each case would help show whether it limits harmful disturbances while preserving useful changes, and clarify where it fails.

In addition, there is a closely related, recently published study titled 'Contrastive Time Series Forecasting with Anomalies' published in TMLR2026. Please clarify how the proposed method differs from or complements that work and, where applicable, include an experimental comparison.

Questions for authors

Please refer to the weaknesses.

6 comments
AuthorsOpenReview ↗

Author Response to wWrg (1/2): W1

We sincerely thank the reviewer for recognizing the motivation, practical relevance, soundness, and extensive evaluation of NEIN-TSF. We have conducted two additional analyses: E1, a controlled assessment of forecasting responses to transient and persistent events, and E2, a comparison with the closely related Co-TSFA method using a common iTransformer backbone.

Implementation note for E1 and E2. The Co-TSFA implementation in E1 is based on the authors' publicly released code, retaining its augmentation and contrastive-learning mechanisms, with the adaptations necessary to make the released training entry point executable. E1 evaluates the trained methods on the same clean, transient, and persistent perturbation protocol. E2 addresses a different question: we start from the official iTransformer implementation and integrate Co-TSFA and NEIN-TSF into this common forecasting backbone, keeping the underlying dataset splits, input/output lengths, and baseline evaluation protocol aligned. We report the two experiments separately rather than treating them as interchangeable training runs.

W1. Does NEIN-TSF suppress irrelevant disturbances while preserving useful changes?

Thank you for this important suggestion. We constructed three controlled scenarios: Clean, with the original history and target; Transient, where a terminal perturbation is injected only into the history while the future target is unchanged; and Persistent, where the same historical perturbation is paired with a corresponding persistent future shift. Methods are evaluated on matched windows, affected channels, perturbation locations, and perturbation strengths; the selected checkpoint is held fixed across the three test scenarios. Perturbations include terminal widths 12/24/48, magnitudes 0.5/1.0/1.5 times the training-set standard deviation, both signs, and 10%/30% affected channels. We interpret E1 as measuring the robustness–shift-tracking trade-off, rather than claiming perfect identification of an unknowable event type. Alongside MSE/MAE, we measure transient prediction response (R_tr) and persistent shift-tracking error (E_shift); the latter two are available for NEIN-TSF.

Table R1 — E1: controlled irregularity evaluation. Entries are horizon-averaged MSE/MAE over H = {96, 192, 336, 720}; lower is better.

DatasetModelClean (MSE / MAE)Transient (MSE / MAE)Persistent (MSE / MAE)R_tr / E_shift
ETTh1Full NEIN-TSF0.396 / 0.4150.401 / 0.4190.411 / 0.4270.318 / 0.581
ETTh1Co-TSFA0.432 / 0.4480.449 / 0.4630.498 / 0.495—
ECLFull NEIN-TSF0.156 / 0.2450.161 / 0.2510.169 / 0.2580.351 / 0.630
ECLCo-TSFA0.254 / 0.3520.280 / 0.3750.343 / 0.420—

NEIN-TSF exhibits smaller degradation from Clean to Transient in both datasets (ETTh1: approximately +0.005 versus +0.017 MSE for Co-TSFA; Electricity: +0.005 versus +0.026). It also yields lower error in the Persistent scenario under this evaluation. These results provide more direct evidence of controlled forecasting responses, while the indistinguishability of identical historical inputs remains an inherent limitation of the test.

For reproducibility, the Co-TSFA E1 run retains its original contrastive objective and augmentation functions, using five augmented views and contrastive weight 0.1; the batch sizes are 32 for ETTh1 and 4 for Electricity, respectively. The same checkpoint selected on the clean validation set is evaluated in all three scenarios.

AuthorsOpenReview ↗

Author Response to wWrg (2/2): W2

W2. Relationship to Contrastive Time Series Forecasting with Anomalies (Co-TSFA)

We are especially grateful for this valuable reference. We carefully studied Co-TSFA and appreciate its important insight that anomalous observations should not be treated uniformly as removable noise. In particular, its forecast-relevant and forecast-irrelevant augmentations, together with latent–output contrastive alignment, offer a principled way to make forecasting representations responsive to future-relevant changes while remaining robust to transient disturbance. We will explicitly cite and discuss this closely related study in the revised Related Work.

We agree that the two papers share a broad motivation, namely that the predictive meaning of an irregularity cannot be determined from appearance alone. Their technical formulations and forecasting mechanisms, however, differ substantially. Co-TSFA uses designed anomaly augmentations and latent–output alignment as its central learning signal. NEIN-TSF instead (i) learns relative forecast damage from the loss differences produced by a frozen teacher through matched counterfactual ranking (FDRP); (ii) constructs historically supported, explicitly bounded residual corrections by combining retrieval, periodicity, trend, and persistence evidence (SPE); and (iii) controls correction strength across prediction positions and selects model states on chronologically later validation data (HAEU/CFUS). Thus, beyond learning whether to respond to an irregularity, NEIN-TSF explicitly models whether historical evidence supports a correction and how strongly that correction remains useful. We view the approaches as complementary and regard the consequence–support–utility formulation as our independent technical contribution.

To make the comparison more concrete, we additionally compare the methods on a shared iTransformer backbone across four horizons. The underlying iTransformer settings follow the original model; the additional Co-TSFA and NEIN-TSF components follow their respective method settings.

Table R2 — E2: common-backbone forecasting comparison. Each entry is MSE / MAE (lower is better);

DatasetHiTransformer+ Co-TSFA+ NEIN-TSF
ETTh1960.386 / 0.4050.387 / 0.4150.372 / 0.389
ETTh11920.441 / 0.4360.429 / 0.4220.423 / 0.418
ETTh13360.487 / 0.4580.474 / 0.4450.466 / 0.439
ETTh17200.503 / 0.4910.499 / 0.4820.478 / 0.460
ECL960.148 / 0.2400.180 / 0.2710.137 / 0.224
ECL1920.162 / 0.2530.199 / 0.3010.152 / 0.245
ECL3360.178 / 0.2690.220 / 0.3180.159 / 0.255
ECL7200.225 / 0.3170.285 / 0.3670.218 / 0.307

On ETTh1, Co-TSFA improves the iTransformer horizon average, and iTransformer + NEIN-TSF further reduces both average MSE and MAE. On Electricity, the Co-TSFA run performs worse than the iTransformer baseline under the present settings, whereas the NEIN-TSF integration improves both metrics. We report this observed result without making a general claim that Co-TSFA is ineffective. In E2, the Co-TSFA numbers come from new runs, while the iTransformer and +NEIN-TSF columns are our existing backbone-transfer results under the stated forecasting protocol; the provenance is kept explicit. E1 and E2 use different training entry points and answer different questions.

We thank the reviewer again for encouraging a more direct empirical assessment and a clearer positioning against this closely related work.

AuthorsOpenReview ↗

Additional Controlled-Irregularity Evaluation on Traffic

Dear Reviewer wWrg,

We sincerely appreciate your thoughtful and insightful feedback, particularly your suggestion to investigate forecasting behavior under controlled irregularities. Your guidance has been especially valuable in strengthening the empirical evidence supporting our central research motivation.

Following our previous evaluations on ETTh1 and Electricity, we have further extended the controlled-irregularity experiments to Traffic, evaluating Full NEIN-TSF and Co-TSFA under the same Clean, Transient, and Persistent scenarios.

Additional Controlled-Irregularity Results on Traffic (MSE/MAE ↓)

ModelCleanTransientPersistent
Full NEIN-TSF0.391/0.2700.399/0.2760.415/0.285
Co-TSFA0.400/0.2850.416/0.3010.473/0.349

NEIN-TSF achieves lower MSE and MAE across all three scenarios. In particular, the MSE increase from Clean to Transient is approximately +0.008 for NEIN-TSF versus +0.016 for Co-TSFA, indicating a smaller degradation under transient disturbances. NEIN-TSF also achieves lower prediction errors under persistent changes (0.415/0.285 versus 0.473/0.349).

Together with the previously reported ETTh1 and Electricity results, these additional findings provide further evidence of the effectiveness of NEIN-TSF in regulating the influence of historical irregularities across different forecasting scenarios.

We are truly grateful for your constructive suggestions, which have helped us make the experimental evaluation more comprehensive and closely aligned with the paper's motivation.

We would sincerely appreciate any further thoughts or suggestions you may have regarding our additional experiments and revised manuscript.

Thank you again for your valuable time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised Manuscript Uploaded — Summary of Changes

Dear Reviewers and Area Chairs,

We sincerely thank all three reviewers for their constructive feedback and valuable suggestions. We have revised the manuscript to incorporate the additional experiments and clarifications presented in our author responses.

The main revisions are summarized below:

  • Section 2 (Related Work): Expanded the discussion of Co-TSFA and clarified its relationship to NEIN-TSF.
  • Section 4.5 (Table 4): Added controlled irregularity evaluation, shared-backbone comparisons, paired component ablations, teacher and periodic-evidence sensitivity, utility calibration, and computational efficiency analyses.
  • Appendix I: Provided the detailed experimental protocols, complete results, and additional resource comparisons supporting these analyses.

All newly added material is highlighted in blue to facilitate review. The original methodological framework and main experimental results remain unchanged.

We greatly appreciate the reviewers' time and consideration, and we welcome any further questions or discussion.

Thank you again for your valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.